A Survey of Graph Neural Networks for Recommender

Systems: Challenges, Methods, and Directions

CHEN GAO, YU ZHENG, NIAN LI, YINFENG LI, YINGRONG QIN, JINGHUA PIAO,
YUHAN QUAN, JIANXIN CHANG, and DEPENG JIN, Beijing National Research Center for
Information Science and Technology (BNRist), Department of Electronic Engineering, Tsinghua University,
China
XIANGNAN HE, School of Information Science and Technology, University of Science and Technology of
China, China
YONG LI, Beijing National Research Center for Information Science and Technology (BNRist), Department
of Electronic Engineering, Tsinghua University, China 3
Recommender system is one of the most important information services on today’s Internet. Recently, graph
neural networks have become the new state-of-the-art approach to recommender systems. In this survey, we
conduct a comprehensive review of the literature on graph neural network-based recommender systems. We
first introduce the background and the history of the development of both recommender systems and graph
neural networks. For recommender systems, in general, there are four aspects for categorizing existing works:
stage, scenario, objective, and application. For graph neural networks, the existing methods consist of two cat-
egories: spectral models and spatial ones. We then discuss the motivation of applying graph neural networks
into recommender systems, mainly consisting of the high-order connectivity, the structural property of data
and the enhanced supervision signal. We then systematically analyze the challenges in graph construction,
embedding propagation/aggregation, model optimization, and computation efficiency. Afterward and primar-
ily, we provide a comprehensive overview of a multitude of existing works of graph neural network-based
recommender systems, following the taxonomy above. Finally, we raise discussions on the open problems
and promising future directions in this area. We summarize the representative papers along with their code
repositories in https://github.com/tsinghua-fib-lab/GNN-Recommender-Systems.
CCS Concepts: • Information systems → Recommender systems; • General and reference → Surveys
and overviews;
Additional Key Words and Phrases: Recommender systems, graph neural networks, graph representation
learning, information retrieval

Chen Gao and Yu Zheng contributed equally to this article.

This work is supported by the National Key Research and Development Program of China under grant 2020AAA0106000.
This work is also supported by the National Natural Science Foundation of China (U19A2079). This work is also supported
by the fellowship of China Postdoctoral Science Foundation under grant 2021TQ0027 and 2022M710006.
Authors’ addresses: C. Gao, Y. Zheng, N. Li, Y. Li, Y. Qin, J. Piao, Y. Quan, J. Chang, D. Jin, and Y. Li (corresponding au-
thor), Beijing National Research Center for Information Science and Technology (BNRist), Department of Electronic Engi-
neering, Tsinghua University, China; emails: [email protected], {y-zheng19, linian21, liyf19, qyr16, pojh19, quanyh19,
changjx18}@mails.tsinghua.edu.cn, {jindp, liyong07}@tsinghua.edu.cn; X. He, School of Information Science and Technol-
ogy, University of Science and Technology of China, China; email: [email protected].
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee
provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and
the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be
honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists,
requires prior specific permission and/or a fee. Request permissions from [email protected].
© 2023 Copyright held by the owner/author(s). Publication rights licensed to ACM.
2770-6699/2023/02-ART3 $15.00
https://doi.org/10.1145/3568022

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
3:2 C. Gao et al.

ACM Reference format:

Chen Gao, Yu Zheng, Nian Li, Yinfeng Li, Yingrong Qin, Jinghua Piao, Yuhan Quan, Jianxin Chang, Depeng Jin,
Xiangnan He, and Yong Li. 2023. A Survey of Graph Neural Networks for Recommender Systems: Challenges,
Methods, and Directions. ACM Trans. Recomm. Syst. 1, 1, Article 3 (February 2023), 51 pages.
https://doi.org/10.1145/3568022

1 INTRODUCTION
Recommender system is a kind of filtering system in which the goal is to present personalized
information to users, which improves the user experience and promotes business profit. As one of
the typical applications of machine learning driven by the real world, it is an extremely hot topic
in both industrial and academia nowadays.
To recap the history of recommender systems, it can be generally divided into three stages,
shallow models [85, 148, 149], neural models [28, 57, 66], and GNN-based models [65, 181, 222].
The earliest recommendation models captured the collaborative filtering (CF) effect by directly
calculating the similarity of interactions. Then model-based CF methods, such as matrix factor-
ization (MF) [85] or factorization machine [148], were proposed to approach recommendation as
a representation learning problem. However, these methods are faced with critical challenges such
as complex user behaviors or data input. To address it, neural network-based models [28, 57, 66]
are proposed. For example, neural collaborative filtering (NCF) was developed to extend the in-
ner product in MF with multi-layer perceptrons (MLP) to improve its capacity. Similarly, deep
factorization machine (DeepFM) [57] combined the shallow model factorization machine
(FM) [148] with MLP. However, these methods are still highly limited, since their paradigms of
prediction and training ignore the high-order structural information in observed data. For example,
the optimization goal of NCF is to predict user-item interaction, and the training samples include
observed positive user-item interactions and unobserved negative user-item interactions. It means
that during the parameter updating for a specific user, only the items interacted by him/her are
involved.
Recently, the advances in graph neural networks provide a strong and fundamental opportu-
nity to address the above issues in recommender systems. Specifically, graph neural networks
adopt embedding propagation to aggregate neighborhood embedding iteratively. By stacking the
propagation layers, each node can access high-order neighbors’ information, rather than only the
first-order neighbors’ as the traditional methods do. With its advantages in handling the structural
data and exploring structural information, GNN-based methods have become the new state-of-the-
art approaches in recommender systems.
To well apply graph neural networks into recommender systems, there are some critical chal-
lenges required to be addressed. First, the data input of recommender system should be carefully
and properly constructed into graph, with nodes representing elements and edges representing
relations. Second, for the specific task, the components in graph neural networks should be adap-
tively designed, including how to propagate and aggregate, in which existing works have explored
various choices with different advantages and disadvantages. Third, the optimization of the GNN-
based model, including the optimization goal, loss function, data sampling, and so on [218], should
be consistent with the task requirement. Last, since recommender systems have strict limitations
on the computation cost, and also due to GNNs’ embedding propagation operations introducing
a number of computations, the efficient deployment of graph neural networks in recommender
systems is another critical challenge.
In this article, we aim to provide a systematic and comprehensive review of the research ef-
forts, especially on how they improve recommendation with graph neural networks and ad-
dress the corresponding challenges. To fulfill a clear understanding, we categorize research of
ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
A Survey of Graph Neural Networks for Recommender Systems 3:3

Fig. 1. An illustration of typical recommender systems (stages, scenarios, objectives, and applications).

recommender systems from four perspectives, stage, scenario, objectives, and applications, as il-
lustrated in Figure 1. We summarize the representative papers along with their code repositories
in https://github.com/tsinghua-fib-lab/GNN-Recommender-Systems.
It is worth mentioning that there is one existing survey [202] of graph neural network-based rec-
ommender system, compared with which our differences are as follows: First, it mainly considers
collaborative filtering, sequential recommendation, social recommendation, and knowledge-based
recommendation, which belong to different aspects based on the taxonomy and thus cannot be
listed side-by-side. For example, the sequential recommendation is only one specific recommen-
dation scenario with a special setting of input and output, as pointed out by this survey. Second,
important recommendation stages (such as ranking), recommendation scenarios (such as cross-
domain recommendation), and beyond-accuracy recommendation objectives are ignored. Espe-
cially, the beyond-accuracy recommender system is now a very important topic in both academia
and industry. Last, since this area is developing rapidly, there are many recent papers in top-tier
venues not covered by Reference [202]. It is worth mentioning that there is a recently published
book chapter [32] that introduces GNN’s basic concepts and several representative applications in
modern recommender systems.
Furthermore, in addition to Reference [202], there are some other surveys closely relevant to
this survey as follows:
• Reference [108] is a very recent survey on recommender systems based on heterogeneous
information network (HIN), which divides literature into three categories: similarity mea-
surement, matrix factorization, and graph representation learning. As for the category of
graph representation learning, the authors further divide it into two sub-categories: two-
stage training-based methods and end-to-end training-based methods (including relation-
based ones and meta-path-based ones). As for the end-to-end methods, this survey briefly
introduces some graph neural network-based methods. That is, Reference [108] only covers
a very small fraction of GNN-based recommenders.
• Reference [176] is a survey on graph-based recommender systems. It is very short (nine
pages) and only uses one page to briefly introduce very-limited typical works of GNN-based
recommenders. Reference [39] is a survey of recommender systems based on graph embed-
ding techniques. It mainly discusses the traditional graph embedding methods and only dis-
cusses a few GNN-based methods.

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
3:4 C. Gao et al.

• Reference [196] is a very recent survey on deep learning-based recommender systems. It

mainly discusses two parts of works: deep learning-based collaborative filtering and deep
learning-based feature learning for recommendation. In the first part, the authors have in-
troduced some representative GNN-based collaborative filtering methods; in the second part,
the authors introduced a few works that use GNN to extract visual features.
• There are other surveys on specific topics of recommendation, such as Reference [175] for
session-based recommendation, References [30, 59] for knowledge-graph based recommen-
dation, Reference [26] for explainable recommendation, Reference [191] for news recommen-
dation. There are some other surveys focusing on specific techniques of recommendation,
such as automated machine learning [241] and self-supervised learning [241].
In summary, our survey is largely different compared with these surveys, as our survey provides
extensive and up-to-date introduction and discussions about GNN-based recommenders.
The structure of this survey is organized as follows: We first introduce the background of
recommender systems from four kinds of perspectives (stage, scenario, objective, application) and
the background of graph neural networks in Section 2. We then discuss the challenges of applying
graph neural networks to recommender systems from four aspects in Section 3. Then, we elaborate
on the representative methods of graph nerual network-based recommendation in Section 4 by
following the taxonomy in the above section. We discuss the most critical open problems in this
area and provide ideas of the future directions in Section 5 and conclude this survey in Section 6.

2 BACKGROUND
2.1 Recommender Systems
2.1.1 Overview. In this section, we present the background of recommender systems from four
perspectives: stages, scenarios, objectives, and applications. Specifically, in industrial applications,
due to the real-world requirements of system engineering, the recommender systems are always
split into three stages, matching, ranking, and re-ranking, forming a standard pipeline. Each stage
has different characteristics on data input, output, model design, and so on. Besides the standard
stages, there are many specific recommendation scenarios with a special definition. For example,
in the past 20 years, social recommendation has been attracting attention, defined as improving
recommender systems based on social relations. Last, different recommender systems have differ-
ent objectives, of which accuracy is always the most important one, as it directly determines the
system’s utility. Recently, recommender systems have been assigned other requirements such as
recommending diversified items to avoid boring user experience, making sure the system treats all
users fairly, protecting user privacy from attack, and so on. As for the applications, GNN models
can be widely deployed in e-commerce recommendation, point-of-interest recommendation, news
recommendation, movie recommendation, music recommendation, and so on.
2.1.2 Stages. The item pool, i.e., all the items available for recommender systems, is usually
large and can include millions of items. Thus, common recommender systems follow a multi-stage
architecture, filtering items stage-by-stage from the large-scale item pool to the final recommen-
dations exposed to users, tens of items [34, 190]. Generally, a modern recommender system is
composed of the following three stages:
• Matching. This first stage generates hundreds of candidate items from the extremely large
item pool (million-level or even billion-level), which significantly reduces the scale. Con-
sidering the large scale of data input in this stage, and due to the strict latency restrictions
of online serving, complicated algorithms cannot be adopted, such as very deep neural
networks [34, 81]. In other words, models in this stage are usually concise. That is, the core
task of this stage is to retrieve potentially relevant items with high efficiency and attain
ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
A Survey of Graph Neural Networks for Recommender Systems 3:5

Fig. 2. The typical pipeline of recommender systems.

coarse-grained modeling of user interests. It is worth noting that a recommender system

in the real world usually contains multiple matching channels with multiple models, such
as embedding matching, geographical matching, popularity matching, social matching,
and so on.
• Ranking. After the matching stage, multiple sources of candidate items from different
channels are merged into one list and then scored by a single ranking model. Specifically,
the ranking model ranks these items according to the scores, and the top dozens of
items are selected. Since the amount of input items in this stage is relatively small, the
system can afford much more complicated algorithms to achieve higher recommendation
accuracy [80, 101, 156]. For example, rich features including user profiles and item attributes
can be taken into consideration, and advanced techniques such as self-attention [80] can
be utilized. Since many features are involved, the key challenge in this stage is to design
appropriate models for capturing complicated feature interactions.
• Re-ranking. Although the obtained item list after the ranking stage is optimized with
respect to relevance, it may not meet other important requirements, such as freshness,
diversity, fairness and so on [143]. Therefore, a re-ranking stage is necessary, which usually
removes certain items or changes the order of the list to fulfill additional criteria and
satisfy business needs. The main concern in this stage is to consider multiple relationships
among the top-scored items [5, 255]. For example, similar or substitutable items can lead to
information redundancy when they are displayed closely in the recommendations.
Figure 2 illustrates the typical pipeline of a recommender system, as well as the comparisons of
the three stages above.
2.1.3 Scenarios. In the following, we will elaborate on the different scenarios of recommender
systems, including social recommendation, sequential recommendation, session recommendation,
bundle recommendation, cross-domain recommendation, and multi-behavior recommendation.
• Social Recommendation. In the past few years, social platforms have dramatically
changed users’ daily lives. With the ability to interact with other users, individual behaviors
are driven by both personal and social factors. Specifically, users’ behavior may be influenced
by what their friends might do or think, which is known as social influence [33]. For exam-
ple, users in WeChat∗ Video platform may like some videos only because of their WeChat
friends’ like behaviors. At the same time, social homophily is another popular phenomenon

∗ http://wechat.com/.

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
3:6 C. Gao et al.

Fig. 3. An illustration of social recommendation. User interactions are affected by both their own preferences
and the social factor.

Fig. 4. An illustration of sequential recommendation. Given a user’s historical sequence, recommender sys-
tem aims to predict the next item.

on many social platforms, i.e., people tend to build social relations with others who have
similar preferences with them [127]. Taking social e-commerce as an example, users from a
common family possibly share similar product preferences, such as food, clothes, daily ne-
cessities, and so on. Hence, social relations are often integrated into recommender systems
to enhance the final performance, which is called social recommendation. Figure 3 illustrates
the data input of social recommendation, of which user interactions are determined by both
preferences and social factors (social influence and social homophily).
• Sequential Recommendation. In recommender systems, users will produce a large num-
ber of interaction behaviors over time. The sequential recommendation method extracts in-
formation from these behavioral sequences and predicts the user’s next interaction item, as
shown in Figure 4. To symbolize this problem, for the sequence of items {x 1 , x 2 , . . . , x n } that
the user has interacted with in order, the goal of the system is to predict the next item x n+1
that the user will interact with. In recommender systems, the user’s historical behaviors play
an important role in modeling the user’s interest. Many commonly used recommendation
methods like collaborative filtering [66] train the model by taking each user behavior as a
sample. They directly model the user’s preference on a single item, but sequential recom-
mendation uses the user’s historical behavior sequence to learn timestamp-aware sequen-
tial patterns to recommend the next item that the user may be interested in. In sequential
recommendation, there are two main challenges. First, for each sample, i.e., each sequence,
the user’s interest needs to be extracted from the sequence to predict the next item. Espe-
cially when the sequence length increases, it is very challenging to simultaneously model
the short-term, long-term, and dynamic interests of users. Second, in addition to modeling
within a sequence, since items may occur in multiple sequences or users have multiple se-
quences, collaborative signals between different sequences require to be captured for better
representation learning.
• Session-based Recommendation. In many real-world scenarios, such as some small retail-
ers and mobile stream media (e.g., YouTube and TikTok), it is impossible or not necessary to

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
A Survey of Graph Neural Networks for Recommender Systems 3:7

Fig. 5. An illustration of session-based recommendation. Given a anonymous short session, recommender

system aims to predict the next item.

Fig. 6. An illustration of bundle recommendation.

Fig. 7. An illustration of cross-domain recommendation.

track the user-ID’s behaviors over a long period due to limited storage resources. In other
words, user profiles and long-term historical interactions are unavailable, and only the short
session data from anonymous users are provided. Hence, conventional recommendation
methods (e.g., collaborative filtering) may perform poorly in this scenarios. This motivates
the problem of session-based recommendation (SBR) [31], which aims at predicting the
next item with a given anonymous behavioral session data, as shown in Figure 5. Distinct
from the sequential recommendation, the subsequent sessions of the same user are handled
independently in SBR, since the behavior of users in each session only shows session-based
traits [67].
• Bundle Recommendation. The existing recommender systems mainly focus on recom-
mending independent items to users. The bundle is a collection of items, which is an impor-
tant marketing strategy for product promotion. Bundle recommendation aims to recommend
a combination of items for users to consume [12, 23, 114, 141], as illustrated in Figure 6. Bun-
dle recommendation is very common nowadays on online platforms, e.g., the music playlists
on Spotify, the pinboards on Pinterest, the computer sets on Amazon, and the furniture suites
on IKEA. It is worth mentioning that bundle recommendations are also used to solve inter-
esting and meaningful problems such as fashion outfits [95] and drug packages [246].
• Cross-domain Recommendation. As an increasing number of users interact with multi-
modal information across multiple domains, cross-domain recommendation (CDR) has
been demonstrated to be a promising method to alleviate cold start and data sparsity
problems [50, 52, 70, 79, 124, 132, 238], as illustrated in Figure 7. CDR methods can be

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
3:8 C. Gao et al.

Fig. 8. An illustration of multi-behavior recommendation.

roughly divided into two categories, single-target CDR (STCDR) and dual-target CDR
(DTCDR) [35]. CDR methods transfer the information from the source domain to the tar-
get domain in one direction; DTCDR emphasizes the mutual utilization of information from
both the source domain and target domain, which can be extended to multiple-target CDR
(MTCDR). Since utilizing information from multiple domains can improve performance,
cross-domain recommendation has become an important scenario in recommender systems.
• Multi-behavior Recommendation. Users interact with recommender systems under mul-
tiple types of behaviors instead of only one type of behavior. For example, when a user clicks
on the video, he/she may also perform behaviors such as collecting or commenting. In an
e-commerce website, users often click, add into shopping carts, share, or collect a product
before purchasing it, as shown in Figure 8. Although the ultimate goal of the recommender
system is to recommend products that users will purchase, the purchase behavior is very
sparse compared to the user’s click, sharing, and other behaviors. To symbolize this problem,
For each user u and item v, suppose there are K different types of behaviors {y1 , y2 , . . . , yK }.
For the ith behavior, if the user has an observed behavior, then yi = 1, otherwise yi = 0. The
goal of the recommender system is to improve the prediction accuracy of a certain type of
target behavior yt .
For multi-behavior recommendation, generally speaking, there are two main challenges.
First, different behaviors have different influences on the target behavior. Some behaviors
may be strong signals, and some may be weak signals [207, 208]. At the same time, this in-
fluence is different for each user. It is challenging to model the influence of these different
behaviors on the target behavior accurately. Second, it is challenging to learn comprehensive
representations from different types of behaviors for items. Different behaviors reflect users’
different preferences for items; in other words, different behaviors have different meanings.
To obtain a better representation, the meaning of different behaviors needs to be integrated
into the representation learning.

2.1.4 Objectives. The most important objective of recommender systems, of course, is accuracy.
In the following, we elaborate on the three other important beyond-accuracy objectives, including
diversity, explainability, and fairness.
• Diversity. Two types of diversity are usually considered in recommender systems, namely,
individual-level diversity and system-level diversity. Specifically, individual-level diversity
is an important objective that measures the dissimilarity of the recommended items for each
user, since repeated similar items make users reluctant to explore the system. In other words,

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
A Survey of Graph Neural Networks for Recommender Systems 3:9

Fig. 9. Illustration of individual-level diversity and system-level diversity.

individual-level diversity reflects how many topics the recommendation list covers and how
balanced is the recommended items distributed on different topics. Here, topics depend on
the recommendation task, e.g., topics can be different product categories for e-commerce
recommendation and different genres for music recommendation [242]. Regarding system-
level diversity, it compares the recommendation results of different users and expects them
to be dis-similar to each other. In other words, low system-level diversity means always rec-
ommending popular items to all the users while ignoring long-tail items. Therefore, system-
level diversity is sometimes called long-tail recommendation. Figure 9 briefly illustrates the
two types of diversity and their differences. For both individual-level and system-level diver-
sity, there are two main challenges. First, the signal strength of different items varies greatly.
For each user, there exist dominant topics and disadvantaged topics, e.g., a user’s interaction
records with electronics might be much more frequent than with clothes. Similarly, the signal
strength of long-tail items is also far weaker than popular items. Therefore, it is challeng-
ing to recommend relevant content with such weak supervision from either disadvantaged
topics or long-tail items for individual-level and system-level diversity, respectively. Second,
diversity may sometimes contradict recommendation accuracy, resulting in the accuracy-
diversity dilemma; thus, it is challenging to balance the two aspects.
• Explainability. As current recommender systems mostly adopt a deep learning paradigm,
they can arouse urgent needs on the explainability of recommendation [236]. The focus of
explainable recommender systems is not only to produce accurate recommendation results
but to generate persuasive explanations for how and why the item is recommended to a
specific user [118, 184, 236]. Increasing the explainability of recommender systems can en-
hance users’ perceived transparency [154], persuasiveness [166], and trustworthiness [86],
and facilitate practitioners to debug and refine the system [236]. This survey mainly focuses
on improving explainability with machine learning techniques. Specifically, past research
adopts two different approaches [236]: one makes efforts to design intrinsic explainable

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
3:10 C. Gao et al.

models, ensuring the explainability of recommendation results by designing models with

transparent logic (rather than merely “black box”), e.g., Explicit Factor Models [237], Hidden
Factor and Topic Model [126], and TriRank [64]. The others compromise slightly: They de-
sign post hoc separate models to explain the results generated by “black box” recommender
systems, e.g., Explanation Mining [142]. Currently, there are two challenges. First, represent-
ing explainable information requires graph-structural item attributes, which are difficult to
model without the power of GNN. Second, reasoning recommendations depend on external
knowledge in the knowledge graph, which also poses challenges to the task.
• Fairness. As a typical data-driven system, recommender systems could be biased by data
and the algorithm, arousing increasing concerns on the fairness [4, 41, 98, 128]. Specifically,
according to the involved stakeholders, fairness in recommender systems can be divided
into two categories [4, 98, 128]: user fairness, which attempts to ensure no algorithmic bias
among specific users or demographic groups [9, 90, 96], and item fairness, which indicates
fair exposures of different items, or no popularity bias among different items [2, 3, 98, 128].
Here, we focus on the user fairness and leave item fairness in the section of diversity for their
close connection in terms of interpretations and solutions [2, 125]. Specifically, researchers
adopt two methods to enhance fairness: one is directly debiasing recommendation results
in the training process [9, 254], while the other endeavors to rank items to alleviate the
unfairness in a post-processing method [55, 153]. Indeed, increasing evidence shows that
the utilization of graph data, e.g., user-user, could intensify concerns on fairness [36, 147,
195]. Thus, it is challenging to debias unfairness in recommendation in the context of rich
graph data. Furthermore, it is even harder to boost user fairness in recommendation from
the perspective of graph.
2.1.5 Applications. Recommender systems widely exist in today’s information services, with
various kinds of applications, of which the representative ones are as follows:
Product recommendation, also known as E-commerce recommendation, is one of the most fa-
mous applications of recommender systems. For recommendation models in the e-commerce sce-
nario, the business value is highly concerned. Therefore, it is crucial to handle multiple types of
behaviors closely relevant to the platform profit, including adding-to-cart or purchasing. Some
works [122] propose to optimize the click-through rate and conversion rate at the same time. In
addition, in e-commerce platform, products may have rich attributes, such as price [244], cate-
gory [248], and so on, based on which heterogeneous graphs can be constructed [117]. Represen-
tative benchmark datasets for product recommendation include Amazon,† Tmall,‡ and so on.
POI (Point-of-Interest) recommendation is also a popular application that aims to recom-
mend new locations/point-of-interests for users’ next visitation. In the point-of-interest recom-
mendation, there are two important factors: spatial factor and temporal factor. The spatial factor
refers to naturally existed geographical attributes of POIs, i.e., the geographic location. In addi-
tion, the users’ visitations are also largely limited by their geographical activity areas, since the
user cannot visit POIs as easily as browsing/purchasing products on e-commerce websites. Be-
sides, the temporal factor is also of great importance, since users’ visitation/check-in behaviors
always form a sequence. This motivates the problem of next-POI or successive POI recommen-
dation [46, 102, 189, 213]. Representative benchmark datasets for POI recommendation includes
Yelp,§ Gowalla, and so on.

† http://jmcauley.ucsd.edu/data/amazon.
‡ https://tianchi.aliyun.com/dataset/dataDetail?dataId=649.
§ https://www.yelp.com/dataset.
 https://snap.stanford.edu/data/loc-gowalla.html.

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
A Survey of Graph Neural Networks for Recommender Systems 3:11

Fig. 10. The overall procedure of implementing a GNN model: construct graphs from data (e.g., table or text),
design tailored GNN for generating representations, map representations to prediction results, and further
define loss function with labels for optimization.

News recommendation helps users find preferred news, which is also another typical ap-
plication. Different from other recommendation applications, news recommendation requires
proper modeling of the texts of news. Therefore, methods of natural language processing can be
combined with recommendation models for extracting news features better [134]. Besides, users
are always interested in up-to-date news and may refuse out-of-date ones. Thus, it is required
to accurately filter news from the fast-changing candidate pools, which is quite challenging. As
for news recommendation, MIND dataset [192] is the recently released representative benchmark
dataset.
Movie recommendation is one of the earliest recommender systems. For example, Netflix’s
movie recommendation competition [7] motivated many pioneer recommendation researches [7,
84]. The earlier setting of movie recommendation is to estimate the users’ rating scores on movies,
from one to five, which is named explicit feedback. Recently, the binary implicit feedback has be-
come more popular setting [66, 149].
There are also other types of recommendation applications, such as video recommenda-
tion [11, 37, 110], music recommendation [87, 167, 186], job recommendation [139], food recom-
mendation [131], and so on.

2.2 Graph Neural Networks

With the rapid emergence of vast volumes of graph data, such as social networks, molecular struc-
tures, and knowledge graphs, a wave of graph neural network (GNN) studies has sprung up in
recent years [8, 47, 49, 83, 168, 229]. The rise of GNN mainly originates from the advancement of
convolutional neural network (CNN) and graph representation learning (GRL) [205, 250].
When applied to regular Euclidean data such as images or texts, CNN is extremely effective in
extracting localized features. However, for non-Euclidean data like graphs, CNN requires general-
ization to handle the situations where operation objects (e.g., pixels in images or nodes on graphs)
are non-fixed in size. In terms of GRL, it aims to generate low-dimensional vectors for graph nodes,
edges, or subgraphs, which represent complex connection structures of graphs. For example, a pi-
oneering work, DeepWalk [144], learns node representations by using SkipGram [130] on a gener-
ated path with random walks on graphs. Combining CNN and GRL, various GNNs are developed
to distill structural information and learn high-level representations. In later parts, we will intro-
duce several general and primary stages for designing a GNN model to accomplish tasks on graphs,
illustrated in Figure 10. Specifically, Sections 2.2.1, 2.2.2, and 2.2.3 elaborate on how to construct
graphs, design specialized and effective graph neural networks, and optimize models, respectively.
2.2.1 Graph Construction. We first utilize the unified formula to define a graph, i.e., G = (V , E),
where V and E denote the set of graph nodes and edges, respectively, and each edge in E joins any
number of nodes in V . In recent years, GNN-based models focus on designing specialized networks
for the following three categories of graphs:

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
3:12 C. Gao et al.

• Homogeneous graph, of which each edge connects only two nodes, and there is only one
type of node and edge.
• Heterogeneous graph, of which each edge connects only two nodes, and there are multiple
types of nodes or edges.
• Hypergraph, of which each edge joins more than two nodes.
In many information services nowadays, relational data is naturally represented in the form
of graphs. For example, implicit social media relationships can be considered as a unified graph,
with nodes representing individuals and edges connecting people who follow each other. However,
since non-structured data such as images and texts do not explicitly contain graphs, it is necessary
to define nodes and edges manually for building graphs. Taking text data used in Natural Lan-
guage Processsing (NLP) as an example, words/documents are described as nodes, and edges
among them are constructed according to Term Frequency-Inverse Document Frequency (IF-
ITF) [220]. An emerging research direction in representation learning on graphs is knowledge
graph (KG), a representative instance of the heterogeneous graph. KG integrates multiple data
attributes and relationships, in which nodes and edges are redefined as entities and relations, re-
spectively. Specifically, the entities in KG can cover a wide range of elements, including persons,
movies, books, and so on. The relations are utilized to describe how entities associate with each
other. For example, a movie can relate to persons (e.g., actors or directors), countries, languages,
and so on. Except for regular graphs, the hypergraph is also explored recently to handle more com-
plex data (e.g., beyond pairwise relations and multiple modals) flexibly [47], in which each edge
can connect more than two nodes.
In summary, constructing graphs necessitates either pre-existing graph data or abstracting the
concept of graph nodes and edges from non-structured data.

2.2.2 Network Design. Generally speaking, GNN models can be categorized into spectral and
spatial models. Spectral models consider graphs as signals and process them with graph convolu-
tion in the spectral domain. Specifically, graph signals are first transformed into spectral domain
by Fourier transform defined on graphs, then a filter is applied, at last the processed signals are
transformed back to spatial domain [152]. The formulation of processing the graph signal x with
the filter g is
g  x = F −1 (F (g)  F (x)) , (1)

where F denotes the graph Fourier transform.

In contrast, spatial models conduct the convolution on graph structures directly to extract local-
ized features via weighted aggregation like CNNs. Despite the fact that these two types of models
start from different places, they fall into the same principle of collecting neighborhood information
iteratively to capture high-order correlations among graph nodes and edges. Here, “information”
is represented as embeddings, i.e., low-dimensional vectors. To this end, the primary and pivotal
operation of GNN is to propagate embeddings on graphs following structural connections, includ-
ing aggregating neighborhood embeddings and fusing them with the target (a node or an edge)
embedding to update the graph embeddings layer-by-layer.
In the following, we will introduce several groundbreaking GNN models to elaborate on how
neural networks are implemented on graphs. The frequently used notations are explained in
Table 1.
• GCN [83]. This is a typical spectral model that combines graph convolution and neural
networks to achieve the graph task of semi-supervised classification. In detail, GCN approx-
imates the filter in convolution by the first order following Reference [63]. Then the node

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
A Survey of Graph Neural Networks for Recommender Systems 3:13

Table 1. Frequently Used Notations

V /E The set of graph nodes/edges

U/I The set of users/items
Ni The neighborhood set of graph node i
hli The embedding of graph node i in the lth propagation layer
A The adjacency matrix of graph
Wl Learnable transformation matrix in the lth propagation layer
δ (·) Nonlinear activation function
 Concatenation operation
 Hadamard product

embeddings are updated as follows:

Hl +1 = δ (D̃− 2 ÃD̃− 2 Hl Wl ),
1 1
(2)
of which the derivation can refer to Reference [205]. Hl ∈ R |V |×D is the embedding matrix of
graph nodes in the lth layer of convolution, where D is the embedding dimension. Besides,
à ∈ R |V |× |V | is the adjacency matrix of the graph with self-loop, of which each entry Ãi j = 1

if the node i connects with j or i = j; otherwise Ãi j = 0, and D̃ii = j Ãi j .
• GraphSAGE [62]. This is a pioneered spatial GNN model that samples neighbors of the
target node, aggregates their embeddings, and merges with the target embedding to update.
 
hlNi = AGGREGATEl {hlj , ∀j ∈ Ni } ,
   (3)
hli +1 = δ Wl hli hlNi ,
where Ni denotes the sampled neighbors of the target node i. The function AGGREGATE
has various options, such as MEAN, LSTM [68], and so on.
• GAT [168]. This is a spatial GNN model that addresses several key challenges of spectral
models, such as poor ability of generalization from a specific graph structure to another
sophisticated computation of matrix inverse. GAT utilizes attention mechanisms to aggre-
gate neighborhood features (embeddings) by specifying different weights to different nodes.
Specifically, the propagation is formulated as follows:

hli +1 = δ  α i j Wl hlj  ,
j ∈Ni     (4)
exp LeakyReLU aT Wl hli Wl hlj
αi j =      ,
T Wl hl Wl hl
k ∈Ni exp LeakyReLU a i k

where α i j is the propagation weight from node j to node i, and Ni is the neighborhood set
of node i, including i itself. As shown in the second equation, the attention mechanism is
implemented via a fully connected layer parameterized by a learnable vector a, followed by
the softmax function.
• HetGNN [229]. This is a spatial GNN tailored for heterogeneous graphs. Considering the het-
erogeneous graph consists of multiple types of nodes and edges, HetGNN first divides neigh-
bors into subsets according to their types. Hereafter, an aggregator function is conducted for
each type of neighbor to gather localized information, combining LSTM and MEAN opera-
tions. Furthermore, different types of neighborhood information are aggregated based on

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
3:14 C. Gao et al.

Fig. 11. Comparison of several typical GNN models. For graph type, nodes and edges type are represented
by colors and line-styles, respectively. For aggregation, line-width indicates neighborhood weight.

the attention mechanism. Detailed formulas are omitted, since the implementation follows
the above works.
• HGNN [47]. This is a spectral model implementing GNN on the hypergraph. The convolu-
tion is defined as follows:
−1 −1
Hl +1 = Dv 2 EDe−1 ET Dv 2 Hl Wl , (5)
where each entry Eui of E ∈ R |V |× |E | denotes whether the hyperedge u contains the node i,
each diagonal entry of Dv ∈ R |V |× |V | denotes how many hyperedges the node is included in,
and each diagonal entry of De ∈ R |E |× |E | denotes how many nodes the hyperedge includes.
Generally, this convolution operation can be considered as two stages of propagating neigh-
borhood embeddings: (1) propagation from nodes to the hyperedge connecting them and
(2) propagation from hyperedges to the node they meet.
The commonality and difference among typical GNN models above are illustrated in Figure 11.
To further capture high-order structural information on the graph, the convolution or embed-
ding propagation mentioned above will be performed for L times. In most cases, L ≤ 4, due to
the over-smoothing problem. Specifically, when the number of propagation layers is too large, the
propagated embeddings will be too similar, losing distinguishing ability. Relevant studies focusing
on developing deep and effective GNN models will be introduced in Section 5.1.1.
2.2.3 Model Optimization. After the processing of the designed network in Section 2.2.2, overall
embeddings of nodes or edges encoding feature semantics as well as graph structures are produced.
To perform the downstream graph learning tasks, these embeddings will be further transformed to
targets (e.g., the probability that a node belongs to a class) by general neural networks (e.g., MLP).
There are mainly classification, prediction, and regression tasks on graphs, including three lev-
els: node, edge, and subgraph. Despite the disparity of various tasks, there is a standard procedure
for model optimization. Specifically, relevant embeddings will be mapped and come with labels to

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
A Survey of Graph Neural Networks for Recommender Systems 3:15

formulate the loss function, and then existing optimizers are utilized for model learning. Follow-
ing this process, there are several types of mapping functions (e.g., MLP, inner product) and loss
functions (e.g., pair-wise, point-wise) to choose for specific tasks. For pair-wise loss function, the
discrimination between positive and negative samples is encouraged, and a typical formulation
BPR [149] is as follows:

L= −lnσ (s (p) − s (n)), (6)
p,n

where σ (·) is the sigmoid function. p and n denote positive and negative samples, respectively, and
s (·) is for measuring the samples. For point-wise loss function, it includes mean square error
(MSE) loss, cross-entropy loss, and so on.
For a better understanding, we take link prediction and node classification tasks as examples to
elaborate on how the GNN model is optimized. For link prediction, the likelihood that whether an
edge exists between two nodes i, j requires definition. Technically, it is usually calculated based
on similarity with node embeddings in each layer of propagation:
s (i, j) = f ({hli }, {hlj }), (7)
where f (·) denotes the mapping function. Furthermore, we can construct the training data as O =
{(i, j, k )}, consisting of observed positive and randomly selected negative samples, (i, j) and (i, k ),
respectively. Specifically, the node i connects with j on the graph but not with k. In recommender
system, the samples will indicate that user i has interacted with item j but not interacted with
item k, where k is sampled from all the other items u has not interacted before. Hereafter, if the
pair-wise loss BPR is chosen, then the optimization object will be

L= −lnσ (s (i, j) − s (i, k )). (8)
(i, j,k ) ∈ O

In terms of node classification, node embedding will be transformed to a probability distribution

representing which class it belongs to, shown as follows:
pi = f ({hli }), (9)
where pi ∈ RC×1 is the distribution and C is the number of classes. Similarly, the training data
will be O = {(i, yi )}, where yi ∈ RC×1 and i belongs to the class c means yic = 1; otherwise,
yic = 0. In general, the point-wise loss function is usually chosen for the classification task, such
as cross-entropy loss, which is formulated as

L=− yTi log pi . (10)
(i,yi ) ∈ O

In short, optimization in GNN-based models treats representations generalized by GNNs as in-

put and graph structures (e.g., edges, node classes) as labels, and loss functions are defined for
training.

2.3 Why Are GNNs Required for Recommender Systems?

Over the past decade, recommender systems have evolved rapidly from traditional factorization ap-
proaches to advanced deep neural networks based models. Particularly, GNN-based recommenders
have achieved the state-of-the-art in many aspects, including different recommendaiton stages, sce-
narios, objectives, and applications. The success of GNN-based recommenders can be explained
from the following three perspectives: (1) structural data; (2) high-order connectivity; (3) supervi-
sion signal.

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
3:16 C. Gao et al.

Structural data. Data collected from online platforms comes in many forms, including user-item
interaction (rating, click, purchase, etc.), user profile (gender, age, income, etc.), item attribute
(brand, category, price, etc.), and so on. Traditional recommender systems are not capable
of leveraging those multiple forms of data, and they usually focus on one or a few specific
data sources, which leads to sub-optimal performance, since much information is ignored. By
expressing all the data as nodes and edges on a graph, GNN provides a unified way to utilize
available data. Meanwhile, GNN shows strong power in learning representations, and thus,
high-quality embeddings for users, items, and other features can be obtained, which is critical to
the recommendation performance.
High-order connectivity. Recommendation accuracy relies on capturing similarity between
users and items, and such similarity is supposed to be reflected in the learned embedding space.
Specifically, the learned embedding for a user is similar to embeddings of items that are interacted
with by the user. Furthermore, those items that are interacted with by other users with similar pref-
erences are also relevant to the user, which is known as the collaborative filtering effect, and it is of
great importance for recommendation accuracy. In traditional approaches, the collaborative filter-
ing effect is only implicitly captured, since the training data is mainly interaction records that only
contain directly connected items. In other words, only first-order connectivity is taken into consid-
eration. The absence of high-order connectivity can largely damage recommendation performance.
In contrast, GNN-based models can effectively capture high-order connectivity. Specifically, the
collaborative filtering effect can be naturally expressed as multi-hop neighbors on the graph, and it
is incorporated into the learned representations through embedding propagation and aggregation.
Supervision signal. Supervision signals are usually sparse in the collected data, while GNN-based
model can leverage semi-supervised signals in the representation learning process to alleviate this
problem. Take the E-Commerce platform as an example: The target behavior, purchase, is pretty
sparse compared to other behaviors. Therefore, recommender systems that only use the target
behavior may have poor performance. GNN-based models can effectively incorporate multiple
non-target behaviors, such as search and add to cart, by encoding semi-supervised signals over
the graph, which can significantly improve recommendation performance [78]. Meanwhile, self-
supervised signals can also be utilized by designing auxiliary tasks on the graph, which further
improves recommendation performance.

3 CHALLENGES OF APPLYING GNNS TO RECOMMENDER SYSTEMS

Although it is well motivated to apply graph neural networks in recommender systems, there exist
four parts of critical challenges.
• How to construct appropriate graphs for specific tasks?
• How to design the mechanism of information propagation and aggregation?
• How to optimize the model?
• How to ensure the efficiency of model training and inference?
In the following, we will elaborate on the four challenges one-by-one.

3.1 Graph Construction

Obviously, the first step of applying graph neural networks is to construct graphs. This is in two
folds: constructing the data input as graph-structured data; reorganizing the recommendation
goal as a task on the graph. Taking the task of standard collaborative filtering as an example, the
data input is the observed user-item interaction data, and the output is predictions of the missing
user-item interactions. Therefore, a bipartite graph with users/items as nodes and interactions as
edges can be constructed. Besides, the CF task turns to the user-item link prediction on the graph.

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
A Survey of Graph Neural Networks for Recommender Systems 3:17

However, it is challenging to construct graphs that can well handle the task properly. It should
be carefully implemented with the consideration of the following aspects:
• Nodes. One of the main goals of learning with graph neural networks is to assign nodes the
representations. This results in that the definition of nodes largely determines the scale of the
GNN models, of which the majority of parameters are occupied by the layer-0 node embed-
dings. Note that edge embeddings are usually either not considered or computed based on
node embeddings. However, it is also a challenging problem to determine whether to distin-
guish different types of nodes. For example, in the collaborative filtering task, user and item
nodes can be modeled differently or considered the same kind of nodes. Another challeng-
ing point is handling concrete input such as numerical features like item prices, which are
always continuous numbers. To represent these features in the graph, one possible solution
is to discretize them into categorical ones, which can be then represented as nodes [244].
• Edges. The definition of edges highly affects the graph’s quality in further propagation and
aggregation, along with the model optimization. In some trivial tasks, the recommender sys-
tem’s data input can be considered a kind of relational data, such as user-item interactions or
user-user social relations. In some complex tasks, other relations can also be represented as
edges. For example, in bundle recommendation, a bundle consists of several items. Then the
edge connecting the bundle and item can reflect the relation of affiliation. When construct-
ing a graph, good designs of edges should fully consider the graph density. A too-dense
graph means there are nodes with extremely high degrees. This will make the embedding
propagation conducted by a huge number of neighbors. It will further make the propagated
embedding non-distinguished and useless. To handle too-dense edges, sampling, filtering, or
pruning on graphs are promising solutions. A too-sparse graph will also result in the poor
utility of embedding propagation, since the propagation will be conducted on only a tiny
fraction of nodes.

3.2 Network Design

The propagation layer makes GNN different from the traditional graph learning methods. As for
the propagation, how to choose the path is critical for modeling the high-order similarity in recom-
mender systems. Besides, the propagation can also be parametric, which assigns different weights
to different nodes. For example, propagating item embeddings to a user node in the user-item in-
teraction graph captures the item-based CF effect. The weights refer to the different importance
of historically interacted items.
In the propagation, there are also various choices of aggregation functions, including mean
pooling, LSTM, max, min, and so on. Since there is no single choice that can perform the best
among all recommendation tasks or different datasets, it is vital to design a specific and proper one.
Besides, the different choices of propagation/aggregation highly affect the computation efficiency.
For example, mean pooling is widely used in GNN-based recommendation models, since it can be
computed efficiently, especially for the graph containing high-degree nodes, such as very popular
items (which can connect many users). Also, the propagation/aggregation layers can be stacked to
help nodes access higher-hops neighbors. Too-shallow layers cannot model the high-order graph
structure cannot be well modeled, and too-deep ones make the node embedding over-smoothed.
Either one of the two cases will lead to poor recommendation performance.

3.3 Model Optimization

To optimize the graph neural network-based recommendation models, the traditional loss func-
tions in recommender system always turn to graph learning losses. For example, the logloss in the

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
3:18 C. Gao et al.

optimization can be regarded as the point-wise link prediction loss. Similarly, BPR loss [149] is usu-
ally adopted in the link prediction task on graphs. Another aspect is data sampling. In GNN-based
recommendation, to sample positive or negative items, the sampling manner can highly depend
on the graph structure. For example, in social recommendation, performing random walk on the
graph can generate weak positive items (such as items interacted by friends).
In addition, sometimes, GNN-based recommendation may involve multiple tasks, such as the
link prediction tasks on different types of edges. Then, in such a case, how to balance each task
and make them enhance each other is challenging.

3.4 Computation Efficiency

In the real world, recommender systems should be trained/inferred efficiently. Therefore, to ensure
the application value of GNN-based recommendation models, their computation efficiency should
be seriously considered. Compared with traditional non-GNN recommendation methods such as
NCF or FM, GNN models’ computation cost is far higher. Complex matrix operations are involved
in each GCN layer, especially for the spectral GNN models such as GCN. With multi-layer stacking
of GCN layers, the computation cost further increases. Therefore, spatial GNN models such as
PinSage can be easier to be implemented in large-scale industrial applications. With sampling
among neighbors or pruned graph structure, efficiency can always be kept as long as we can bear
the drop in recommendation performance.

4 EXISTING METHODS
4.1 Taxonomy
In recent years, GNN has been applied to a wide range of recommendation tasks. Here, we define
the taxonomy in terms of recommendation stages, scenarios, objectives, and applications, respec-
tively. To be more specific, recommendation stages indicate the overall procedure that a recom-
mender system is implemented in the real-world platform. The procedure includes matching for
item candidate selection, ranking for capturing user preferences, and re-ranking for other criteria
beyond accuracy. The recommendation scenarios include social recommendation, sequential rec-
ommendation, cross-domain recommendation, and so on. The recommendation objectives incorpo-
rate accuracy, diversity, explainability, fairness, and so on, in which accuracy is of the most concern.
The recommendation application refers to specific industrial applications. Tables 2, 3, and 4 show
representative researches of GNN-based recommendation published in top-tier venues for different
recommendation stages, recommendation scenarios, and recommendation objectives, respectively.

4.2 GNN in Different Recommendation Stages

4.2.1 GNN in Matching. In the matching stage, efficiency is an essential problem because of the
high computation complexity for candidate selection. Specifically, only hundreds of items will be
selected from the item pool of million magnitudes for the following ranking stage, based on coarse-
grained user preferences. Therefore, proposed models in this stage barely leverage user-item inter-
actions as data input for modeling user preferences without introducing additional features such
as user ages, item price, browsing time on the application, and so on. For matching, besides the
basic challenges of graph construction and network design, there is also the challenge of CF signal
extraction, as shown in Table 5.
GNN-based models in the matching stage can be regarded as embedding matching, usually
designing specialized GNN architecture on the user-item bipartite graph [8, 160, 181, 183, 194].
Berg et al. [8] proposed to pass neighborhood messages by summing and assign weight-sharing
transformation channels for different relational edges (i.e., user-item ratings). Wang et al. [181]

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
A Survey of Graph Neural Networks for Recommender Systems 3:19

Table 2. A Summary of GNN-based Models in

Different Recommendation Stages in Top-tier Venues

Stage Model Venue Year

GCMC [8] KDD 2018
PinSage [222] KDD 2018
NGCF [181] SIGIR 2019
LightGCN [65] SIGIR 2020
NIA-GCN [160] SIGIR 2020
Matching
DGCF [183] SIGIR 2020
IMP-GCN [104] WWW 2021
SGL [194] SIGIR 2021
HS-GCN [107] TKDE 2022
LGCN [228] AAAI 2022
Fi-GNN [99] CIKM 2019
PUP [244] ICDE 2020
A2-GCN [105] TKDE 2020
Ranking
L 0 -SIGN [158] AAAI 2021
DG-ENN [60] KDD 2021
TGIN [77] WSDM 2022
Re-ranking IRGPR [111] CIKM 2020

proposed a spatial GNN in recommendation and obtain superior performance compared with
conventional CF methods such as MF [85] or NCF [66]. Sun et al. [160] argued that simple
aggregation mechanisms such as sum, mean, or max cannot model relational information among
neighbors and proposed a neighbor interaction-aware convolution to address the issue. Wang
et al. [183] developed disentangled GNN to capture independent user intentions, which extends
the set of candidate items in matching and guarantees the accuracy simultaneously. Wu et al. [194]
leverage the stability of graph structure to incorporate a contrastive learning framework to assist
representation learning. These GNN-based models can capture high-order similarity among users
and items as well as structural connectivity. In this way, the semantics that users with similar
interactions will have similar preferences are extended through multiple times of information
propagation. However, the training complexity of GNN-based models was demonstrated [181, 183]
as acceptable and comparable with non-graph models, especially when the transformation matrix
is removed [65]. Besides, Reference [222] showed that the GNN-based model could be applied
to web-scale recommender systems in real-world platforms efficiently and effectively, which
combines random walk and GraphSAGE [62] for embedding learning on a large-scale item-item
graph. Table 5 shows the commonality and difference among the GNN models in matching [104]
proposed first to filter promising neighbors and then adopt embedding propagation.
We would like to briefly discuss the advantages and disadvantages of different GNN-based
matching methods. These methods with sampling techniques such as Reference [222] can be used
for large graphs, and they are more suitable for denser edges, such as item-item relations. Recent
advances [65] remove non-linear operations to improve the performance further, but on some
datasets, the performance of methods with non-linear operations [181] are still comparable. In a
nutshell, GNN can be applied to recommendation tasks effectively, which balances the accuracy
and efficiency of generating candidates from the item pool.
4.2.2 GNN in Ranking. In the ranking stage, with a much smaller amount of candidate items,
more accurate models can be utilized, and more features can be included. Existing ranking models

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
3:20 C. Gao et al.

Table 3. A Summary of GNN-based Models in Different

Recommendation Scenarios in Top-tier Venues
Scenario Model Venue Year
DiffNet [198] SIGIR 2019
GraphRec [43] WWW 2019
DANSER [201] WWW 2019
DGRec [157] WSDM 2019
HGP [82] RecSys 2019
Social
DiffNet++ [197] TKDE 2020
MHCN [227] WWW 2021
SEPT [225] KDD 2021
GBGCN [230] ICDE 2021
KCGN [73] AAAI 2021
DiffNetLG [155] SIGIR 2021
ISSR [106] AAAI 2020
MA-GNN [119] AAAI 2020
STP-UDGAT [102] CIKM 2020
GPR [13] CIKM 2020
Sequential GES-SASRec [253] TKDE 2021
RetaGNN [69] WWW 2021
TGSRec [44] CIKM 2021
SGRec [97] IJCAI 2021
SURGE [15] SIGIR 2021
SR-GNN [203] AAAI 2019
GC-SAN [215] IJCAI 2019
TA-GNN [224] SIGIR 2020
MGNN-SPred [177] WWW 2020
LESSR [24] KDD 2020
MKM-SR [129] SIGIR 2020
GAG [146] SIGIR 2020
Session GCE-GNN [187] SIGIR 2020
SGNN-HN [137] CIKM 2020
DHCN [210] AAAI 2021
SHARE [173] SDM 2021
SERec [25] WSDM 2021
COTREC [209] CIKM 2021
DAT-MID [16] SIGIR 2021
TASRec [249] SIGIR 2021
G3 SR [40] TNNLS 2022
HG-GNN [138] WSDM 2022
CGL [136] TOIS 2022
BGCN [14] SIGIR 2020
HFGN [95] SIGIR 2020
BundleNet [38] CIKM 2020
Bundle
DPR [246] WWW 2021
DPG [247] TOIS 2022
MIDGN [240] AAAI 2022
PPGN [238] CIKM 2019
BiTGCF [109] CIKM 2020
Cross Domain DAN [170] CIKM 2020
HeroGRAPH [35] Recsys 2020
DAGCN [58] IJCAI 2021

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
A Survey of Graph Neural Networks for Recommender Systems 3:21

Table 4. A Summary of GNN-based Models for Different

Recommendation Objectives in Top-tier Venues

Objective Model Venue Year

MBGCN [78] SIGIR 2020
MGNN-SPred [177] WWW 2020
MGNN [235] CIKM 2020
LP-MRGNN [178] TKDE 2021
GNMR [206] ICDE 2021
Multi-behavior
MB-GMN [208] SIGIR 2021
KHGT [207] AAAI 2021
GHCF [17] AAAI 2021
DMBGN [212] KDD 2021
V2HT [94] CIKM 2019
Diversity BGCF [159] KDD 2020
DGCN [242] WWW 2021
RippleNet [171] CIKM 2018
EIUM [75] MM 2019
KPRN [185] AAAI 2019
RuleRec [121] WWW 2019
Explainability
PGPR [211] SIGIR 2019
KGAT [180] KDD 2019
TMER [21] WSDM 2021
FairGo [195] WWW 2021
Fairness
FairGNN [36] WSDM 2021

Table 5. Details of GNN Models in Matching Stage

Model Graph Construction Network Design Specific Design for Collaborative Filtering
GCMC [8] user-item GCN weight sharing among relations
NGCF [181] user-item GCN enhance propagated information
DGCF [183] user-item GAT disentangled representations
LightGCN [65] user-item LightGCN remove transformation and nonlinearity
SGL [194] user-item GCN self-supervision on graphs
NIA-GCN [160] user-item NIA-GCN neighbor interaction (NI)
PinSage [222] item-item GCN neighbor sampling
IMP-GCN [104] user-item GCN neighbor sampling
HS-GCN [107] user-item GCN re-define distance metrics between nodes
LGCN [228] user-item low-pass GCN trainable kernels for spectral graph convolution

usually first convert sparse features into one-hot encodings and then transform them into dense
embedding vectors. These feature embeddings are directly concatenated and fed into DNN [28, 57]
or specifically designed models [66, 148, 156] in an unstructured way to estimate the ranking score.
The main challenge of utilizing GNN for ranking is designing proper structures to capture feature
interactions, as shown in Table 6. Specifically, GNN-based ranking models usually consist of two
components, encoder and predictor, which address feature interaction from different directions. On
the one hand, special graph structures can be designed to capture the desired feature interactions in
the encoder. On the other hand, feature interaction can be taken into consideration in the predictor,
where the ranking score is estimated by integrating different feature embeddings from the GNN
encoder.

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
3:22 C. Gao et al.

Table 6. Details of GNN Models in Ranking Stage

Model Graph Construction Network Design Feature Interaction

Fi-GNN [99] fully connected GAT+GRU attention
PUP [244] pre-defined GCN two-branch FM
L 0 -SIGN [158] learned SIGN graph classification
DG-ENN [60] pre-defined LightGCN DNN
SHCF [91] pre-defined HGAT inner product
GCM [193] pre-defined context GCN FM
A2-GCN [105] pre-defined attentive GCN inner product
TGIN [77] pre-defined triangle-based GCN DNN

Li et al. [99] propose Feature Interaction Graph Neural Networks (Fi-GNN), which
constructs a weighted fully connected graph of all the input features. The encoder in Fi-GNN is
composed of a GAT and a GRU, and the predictor is achieved with attention networks. Zheng
et al. [243, 244] investigate the influence of price feature in ranking and propose a model called
Price-aware User Preference modeling (PUP). They design an encoder with GCN on a pre-
defined heterogeneous graph to capture price awareness, and a two-branch factorization machine
is utilized as the predictor. Liu [105] also approaches the recommendation problem when some
item attributes are available and similarly builds item-attributes edges. Then the item embeddings
can reflect both user behaviors and attribute information. Since not all feature interactions are
useful, L 0 -SIGN [158] automatically detects beneficial feature interactions and only reserves
those edges, resulting in a learned graph that is further fed into a graph classification model
to estimate the ranking score. In addition, Guo et al. [60] propose DG-ENN with a dual graph
of an attribute graph and a collaborative graph, which integrates the information of different
fields to refine the embedding for ranking. Furthermore, SHCF [91] and GCM [193] utilize extra
nodes and edge attributes to represent item attributes and context information, respectively.
Classical interaction predictors are adopted, such as inner product and FM. Table 6 illustrates the
differences between the above GNN-based ranking models in terms of the designs of the encoder
and predictor. Overall, these GNN-based ranking methods can be regarded as a combination
of the traditional feature-interaction methods and GNN-based representation learning. As a
result, the performance is still constrained by the feature-interaction component. A model that
well couples the graph structure, such as high-order paths, into high-order feature extraction
is desired.
4.2.3 GNN in Re-ranking. After obtaining the scores of recommended items, top items are fur-
ther re-ranked with pre-defined rules or functions to improve the recommendation quality. Specif-
ically, two key factors need to be considered in re-ranking. First, different items can have mutual
influence by certain relationships such as substitutability and complementarity. Second, different
users tend to have distinct preferences, and thus re-ranking can also be personalized. GNN pro-
vides a unified way to encode both item relationships and user preferences, and thus the major
challenge is how to fuse multiple re-ranking goals. Liu et al. [111] propose a model called IRGPR to
accomplish personalized re-ranking with the help of GNN. They propose a heterogeneous graph
to fuse the two information sources, one item relation graph to capture multiple item relationships
and one user-item scoring graph to include the initial ranking scores. User and item embeddings
are obtained after multiple message propagation layers, including global item relation propaga-
tion and personalized intent propagation. The final order of re-ranked items is generated with a
feed-forward network. In short, the current works in GNN-based re-ranking only considered a few
re-ranking goals, and others remain less explored.

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
A Survey of Graph Neural Networks for Recommender Systems 3:23

4.3 GNN in Different Recommendation Scenarios

4.3.1 GNN in Social Recommendation. In social recommendation, we have social networks that
contain the social relations of each user, and the goal is to utilize the local neighbors’ preferences
for each user in social networks to enhance the user modeling [18, 161, 199, 200]. From the per-
spective of representation learning with GNN, there are two key considerations in social recom-
mendation: (1) how to capture the social factor; (2) how to combine the social factor from friends
and user preference from his/her interaction behaviors. Here, we first summarize how the existing
works capture the social factors from two perspectives, i.e., graph construction and information
propagation.
• Graph construction. In social-aware recommender systems, a user’s final behavior is de-
cided by both the social impacts from friends and his/her preferences. One of the main
challenges in social recommendation is constructing a social graph to capture the social
influences of friends. Generally speaking, a certain user in social networks is not only influ-
enced by his/her friends (the first-order neighbors) but also influenced by friends’ friends
(the high-order neighbors). Given that normal graph can only model the pairwise relations,
normal graph-based methods [43, 61, 82, 116, 133, 155, 197, 198, 201, 217, 225, 226] stacked
multiple GNN layers to capture multi-hop high-order social relations. However, stacked
GNN layers may suffer from the over-smoothing [19] problem, which may lead to signif-
icant performance degradation. Hypergraph-based methods, such as MHCN [227], propose
to model the high-order social relations with hyperedge [47], which can connect more
than two nodes and model the high-order relations in a natural way. HOSR [113] recur-
sively propagates embeddings on the social network to reflect the influence of high-order
neighbors in the user representations. To further improve the recommendation performance,
some works [6, 73, 157, 216, 230] introduce side information when constructing the graph.
RecoGCN [216] unifies users, items, and selling agents into a heterogeneous graph to cap-
ture the complex relations in social E-commerce. GBGCN [230] constructs a graph for or-
ganizing user behaviors of two views in the group-buying recommendation, of which the
initiator view contains those initiator-item interactions, and the participant view contains
those participant-item interactions. DGRec [157] and TGRec [6] introduce temporal infor-
mation of user behaviors into social recommendation. KCGN [73] proposes to capture both
user-user and item-item relations with the developed knowledge-aware coupled graph.
• Information propagation. As for the propagation on the constructed graph for social rec-
ommendation, there are two main propagation mechanisms, i.e., GCN and GAT. Some meth-
ods [6, 61, 73, 82, 155, 198, 217, 225, 227, 230] conduct GCN on social graph and treat the
social influence of friends equally. RecoGCN [216] conducts meta-path-based GCN prop-
agation on the constructed graph to capture both the social impact and user preference.
HOSR [113] aggregates the information from neighbors with GCN to capture the high-
order relations in the social graph. MHCN [227] performs propagation with GCN on con-
structed hypergraph to obtain the high-order social relations. Some methods with GAT
mechanism [43, 116, 133, 157, 197, 201, 226], such as GraphRec [43] and DiffNet++ [197],
assume that the social influences from different neighbors on the social graph are different
and assign different weights to the social influences from different friends.
In social recommendation, user representations are learned from two distinct perspectives, i.e.,
social influence and user interactions. To combine the user representations from the above two per-
spectives, there are two strategies: (1) separately learn user representations from the social graph
and user-item bipartite graph and (2) jointly learn user representations from the unified graph that
consists of social graph and user-item bipartite graph. The methods with the first strategy, such as

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
3:24 C. Gao et al.

Table 7. Details of GNN Models for Social Recommendation

Model Graph Construction Network Design Social Signal Extraction
DiffNet [198] social graph GCN sum-pooling
GraphRec [43] social graph + user-item graph GAT concatenation
DANSER [201] social graph + user-item graph GAT & GCN –
DiffNet++ [197] heterogeneous graph GAT multi-level attention network
MHCN [227] multi-channel hypergraph + user-item graph HyperGCN sum-pooling
SEPT [225] triangle-graphs + user-item graph GCN –
RecoGCN [216] heterogeneous graph Meta-path + GCN concatenation
ESRF [226] motif-induced graph GAT sum-pooling
GNN-SoR [61] heterogeneous graph GCN concatenation
ASR [116] heterogeneous graph GAT concatenation
GBGCN [230] heterogeneous graph GCN –
DGRec [157] social graph GAT –
SR-HGNN [217] social graph + user-item graph GCN concatenation
KCGN [73] social graph + item-item graph GCN concatenation
HGP [82] group-user graph + user-item graph GCN attention mechanism
GAT-NSR [133] social graph + user-item graph GAT MLP
HOSR [113] social graph + user-item graph GCN attention mechanism
DiffNetLG [155] heterogeneous graph GCN concatenation
TGRec [6] heterogeneous graph GCN attention mechanism

Fig. 12. Illustration of GNN models for social recommendation.

DiffNet [198], GraphRec [43], and MHCN [227], first separately learn user representations from so-
cial graph and user-item graph and then combine the representations with sum-pooling [198, 227],
concatenation [43], MLP [133], or attention mechanism [6, 82, 113]. DiffNet++ [197], a typical
method with the second strategy, first aggregates the information in the user-item sub-graph and
social sub-graph with the GAT mechanism and then combines the representations with the de-
signed multi-level attention network at each layer. Table 7 shows the differences among the above
approaches for social recommendation.
To sum up, the development of social recommendation with GNN can be summarized in
Figure 12. Early efforts in social recommendation only model the social network with GNN,
such as DiffNet [198]. Then, the methods [43, 61, 82, 116, 133, 155, 197, 198, 201, 217, 225, 226]
that model both social network and user interactions with GNNs become the mainstream in

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
A Survey of Graph Neural Networks for Recommender Systems 3:25

Fig. 13. Illustration of GNN models for sequential recommendation.

GNN-based social recommendation. Moreover, some studies, such as MHCN [227] and HOSR [113],
attempt to enhance the recommendation by modeling the high-order relations in social networks
more sufficiently. Also, there exist some works [6, 73, 157, 216, 230] that introduce additional
information to further enhance the social recommendation. In short, the recent advances that can
distinguish social relations with different strengths can always achieve better performance than
early efforts roughly treating all social relations as the same.
4.3.2 GNN in Sequential Recommendation. For sequential recommendation, to improve the rec-
ommendation performance, it is necessary to extract as much effective information as possible
from the sequence and to learn the user’s interest in the sequence, including short-term interest,
long-term interest, dynamic interest, and so on, to accurately predict the next item that the user
may be interested in. Some tools for sequence modeling have been used, such as Markov chains [27]
or recurrent neural networks [80]. Graph neural networks can be well leveraged for short-term,
dynamic interest modeling or representation learning by converting the data to a graph. A general
pattern for sequential modeling with GNN is shown in Figure 13.
SURGE [15] transforms the sequence of each user into an item-item graph and adaptively learns
the weights of edges through metric learning, with only the stronger edges retained by dynamic
graph pooling. The retained graph is converted to a sequence by position flattening and finally used
to predict the next item. Ma et al. [119] consider the short-term interest modeling in the sequence
to build an item-item graph. For each item, this work only builds edges with other items close
to it in the sequence. This enables it to learn short-term user interests in the sequence while still
learning long-term user interests through other networks. The learned multiple representations
are fused together and used for final recommendation.
Besides the challenges of graph construction and network design, the methods of GNN-based
sequential recommendation should carefully design how to extract the sequential patterns in users’
behaviors, as shown in Table 8. Since GNN has the ability of high-order relationship modeling by
aggregating information from neighbor nodes, after fusing multiple sequences into one graph, it
can learn representations of both users and items in different sequences, which ca not be accom-
plished by the Markov model or recurrent neural network. Wang et al. [169] propose a simple
method that directly converts the sequence information into directed edges on the graph and then
uses GNN to learn representations. Liu et al. [106] construct a user-item bipartite graph and an
item-item graph at the same time, where the edges of the item-item graph indicate co-occurrence
in a sequence, with edge weights assigned according to the number of occurrences. The represen-
tations learned by GNN are used in the final recommendation through the recurrent neural net-
work. Different from directly converting the temporal sequence into directed edges in the graph,

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
3:26 C. Gao et al.

Table 8. Details of GNN Models for Sequential Recommendation

Model Graph Construction Network Design Sequential Modeling

SURGE [15] item-item graph GAT RNN
[169] item-item graph GCN Attention
ISSR [106] item-item and user-item graph GCN RNN
MA-GNN [119] item-item graph GCN Memory network
DGSR [234] user-item graph GAT RNN
GES-SASRec [253] item-item graph GCN RNN
RetaGNN [69] temporal heterogeneous graph GAT Self Attention
TGSRec [44] temporal user-item graph GAT GAT
SGRec [97] item-item graph GAT GAT
GME [213] item-item and user-item graph GAT GAT
STP-UDGAT [102] item-item and user-item graph GAT GAT
GPR [13] item-item and user-item graph GCN GCN

Fig. 14. Illustration of GNN models for session-based recommendation (SBR).

DGSR [234] and TGSRec [44] consider the timestamps in the sequence in the process of graph
construction. In the graph, each edge represents the interaction between the user and the item,
along with the corresponding time attribute. Then perform convolution operations on the tempo-
ral graph to learn the representations of users and items. GES-SASRec [253] and SGRec [97] focus
on the learning of item representations. For an item in a sequence, GES-SASRec [253] considers
the next item of this item in other sequences, and SGRec [97] not only considers the next item
but also considers the previous one. By aggregating the items before and after the target item in
different sequences, the representation of the item is enhanced. GPR [13] and GME [213] construct
edges between items by considering the frequency of consecutive occurrences or occurrences in
the same sequence to enhance the representation. Some works are more complicated. For exam-
ple, RetaGNN [69] considers the attributes of the items when constructing the graph, while STP-
UDGAT [102] considers the geographic location, timestamp, and frequency in the POI recommen-
dation. Table 8 summarizes the above works. Current networks of graph neural network-based
recommendation are still highly relying on an additional sequence model such as RNN, Trans-
former, and so on, for sequential modeling. The ability of graphs in sequential modeling, such as
directed edges, dynamic nodes, requires more research efforts.
4.3.3 GNN in Session-based Recommendation. In session-based recommendation, the session
data may contain both user interests and noisy signals. Suppose a session for a certain user,

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
A Survey of Graph Neural Networks for Recommender Systems 3:27

iPhone → iPad → milk → AirPods. Obviously, milk is likely clicked by mistake and then becomes
a noise, as the session reflects the user’s preference for electronic products. Hence, the two main
considerations in the session-based recommendation are (1) how to model the item transition pat-
tern in session data and (2) how to activate the user’s core interests from noisy data. From the
perspective of graph learning, the item transitions can be modeled as graph, and the information
propagation on the graph can activate the user’s actual interests.
• Graph construction. In session-based recommendation, most existing works [145, 187, 203,
215, 224] model the session data with a directed graph to capture the item transition pattern.
Distinct from sequential recommendation, the session sequence in session-based recommen-
dation is short and the user behaviors are limited, i.e., the average length of sequences in
Tmall∥ is only 6.69 [187, 210]. Hence, a session graph constructed from a single session may
only contain limited nodes and edges. To address the above challenge and sufficiently capture
the possible relations among items, there are two strategies: (1) straightforwardly capturing
relations from other sessions and (2) adding the additional edges of the session graph. For
the first strategy, A-PGNN [233], DGTN [245], and GAG [146] propose to enhance relations
of the current session graph with related sessions, and GCE-GNN [187] leverages the global
context by constructing another global graph to assist the transition patterns in the current
session. DHCN [210] regards each session as a hyperedge and represents all sessions in a hy-
pergraph to model the high-order item relations. SERec [25] enhances the global information
for each session with a knowledge graph. CAGE [151] learns the representations of semantic-
level entities by leveraging the open knowledge graph to improve the session-based news
recommendation. MKM-SR [129] enhances the information in the given session by incorpo-
rating user micro-behaviors and item knowledge graph. COTREC [209] unifies all sessions
into a global item graph from the item view and captures the relations among sessions by
line graph from the session view. DAT-MID [16] follows GCE-GNN [187] to construct both
session graph and global graph and then learns item embeddings from different domains.
TASRec [249] constructs a graph for each day to model the relations among items and en-
hance the information in each session. As for the second strategy, SGNN-HN [137] con-
structs a star graph with a “star” node to gain extra knowledge in session data. SHARE [173]
expands the hyperedge connections by sliding the contextual window on the session se-
quence. LESSR [24] proposes first to construct an edge-order preserving multigraph and
then construct a shortcut graph for each session for enriching edge links. A3 SR [40] pro-
poses to construct one global graph to represent all sessions and distinguish session-level
prediction signals based on the later embedding propagation. HG-GNN [138] proposed to
construct a heterogeneous graph in which the item transitions are captured by two types of
item-item edge: in-type and out-type. CGL [136] also proposed to construct a global-level
graph and a session-level graph, but it considers the learning on two graphs as two tasks,
based on which multi-task learning is adopted.
• Information propagation. As for the information propagation on the constructed graph,
there are four propagation mechanisms that are used in session-based recommendation,
e.g., gated GNN, GCN, GAT, and GraphSAGE. SR-GNN [203] and its related works [25, 129,
137, 215, 224, 233] combine the gated recurrent units in the propagation (gated GNN) on
the session graph. GAG [146], DCTN [245] conduct graph convolution on the constructed
directed graph. DHCN [210] proposes to perform graph convolution on both hypergraph
and line graph to obtain session representations from two different perspectives. Similar to

∥ https://www.tmall.com.

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
3:28 C. Gao et al.

Table 9. Details of GNN Models for Session-based Recommendation

Model Graph Construction Network Design Enrich Graph Structure
SR-GNN [203] directed graph gated GNN –
GC-SAN [215] directed graph gated GNN –
TA-GNN [224] directed graph gated GNN –
FGNN [145] directed graph GAT –
A-PGNN [233] directed graph gated GNN cross sessions
MGNN-SPred [177] directed & multi-relational item graph GraphSAGE cross sessions
GAG [146] directed graph GCN cross sessions
SGNN-HN [137] star graph gated GNN additional edges
GCE-GNN [187] directed graph + global graph GAT cross sessions
DGTN [245] directed graph GCN cross sessions
DHCN [210] hypergraph + line graph HyperGCN cross sessions
SHARE [173] session hypergraph HyperGAT additional edges
SERec [25] KG + directed graph GAT + gated GNN cross sessions
LESSR [24] directed graph GAT additional edges
CAGE [151] KG + article-level graph GCN cross sessions
MKM-SR [129] KG + directed graph gated GNN cross sessions
COTREC [209] item graph + line graph GCN cross sessions
DAT-MDI [16] directed graph + global graph GAT cross sessions
TASRec [249] dynamic graph GCN cross sessions
HG-GNN [138] directed graph+global graph gated GNN cross sessions
CGL [136] directed graph+global graph gated GNN cross sessions

DHCN [210], COTREC [209] performs GCN on item graph and line graph to obtain informa-
tion from item and session views, respectively. CAGE [151] conducts GCN on article-level
graph, and TASRec [249] performs graph convolution on the dynamic graph to capture the
item relations. FGNN [145] conducts GAT on a directed session graph to assign different
weights to different items. SHARE [173] performs GAT on session hypergraphs to capture
the high-order contextual relations among items. GCE-GNN [187] and DAT-MID [16] per-
form GAT on both the session graph and global graph to capture the local and global in-
formation, respectively. MGNN-SPred [177] adopts GraphSAGE on a multi-relational item
graph to capture the information from different types of neighbors.
Table 9 shows the differences among the above approaches for session-based recommendation.
To sum up, the development of session-based recommendation (SBR) with GNN can be
summarized in Figure 14. Early efforts in SBR only model each session sequence with a directed
graph, such as SR-GNN [203], GC-SAN [215], TA-GNN [224], and FGNN [145]. Then, some meth-
ods attempt to enrich the relation and information in the session graph with other sessions or
additional links. The methods in References [16, 25, 129, 146, 151, 177, 187, 209, 210, 233, 245, 249]
combine the information from other sessions to capture more information from similar sessions or
all sessions. Moreover, the methods with additional links, such as SGNN-HN [137], SHARE [173],
and LESSR [24], attempt to introduce additional edges to the given session to capture the complex
relations and information in session data. Furthermore, some studies, such as DHCN [210] and
SHARE [173], attempt to enhance the recommendation by modeling the high-order relations
in session data more sufficiently. As we can observe, the current works of GNN models for
session-based recommendation have been developed from simple graphs to complex graphs, with
even hyper-edges. Unlike sequential recommendation, session-based recommendation, in which
users have repeated interactions with specific items, is more suitable for GNN-based models.
Thus, the existing works of session-based recommendation can build more powerful and complex
graph structures. A possible improvement of this area is to explore larger-scale datasets and
further evaluate GNN’s ability in these datasets.

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
A Survey of Graph Neural Networks for Recommender Systems 3:29

Fig. 15. Illustration of GNN models for bundle recommendation.

Table 10. Details of GNN Models for Bundle Recommendation
Model Graph construction for affiliation Refining bundle Network design for High-order Relation
BGCN [14] tripartite graph meta-path propagation multi-layer GCN
HFGN [95] hierarchical graph direct aggregation multi-layer GCN
BundleNet [38] tripartite graph transformation parameters multi-layer GCN
DPR [246] tripartite graph graph induction multi-layer GCN
DPG [247] tripartite graph graph induction multi-layer GCN
MIDGN [240] tripartite graph graph induction multi-layer GCN & disentangling

4.3.4 GNN in Bundle Recommendation. The three challenges of bundle recommendation are (1)
users’ decisions towards bundles are determined by the items the bundles contain (the affiliation
relation), (2) learning bundle representations with the sparse user-bundle interactions, and (3) high-
order relations. The existing works of GNN-based bundle recommendation are shown in Figure 15
by learning from user-item interaction and user-bundle interaction together, with parameter
sharing or joint loss function [23, 141]. For the first time, Chang et al. [14] propose a GNN-based
model that unifies both two parts of interactions and the bundle-item affiliation-relations into one
graph. Then the item can serve as the bridge for embedding propagation between user-bundle and
bundle-bundle. Besides, a specially designed sampling manner for finding hard-negative samples
is further proposed for training. Deng et al. [38] construct a similar tripartite graph with the trans-
formation parameters to well extract bundle representations from included items’ representations.
Zheng et al. [246] consider the bundle recommendation in the drug package dataset and propose
to initialize a graph with auxiliary data and represent the interaction as scalar weights and vectors.
GNN layers are proposed for obtaining the drug package embeddings. This method is further
improved by Reference [247] by modeling the interaction data with the graph consisting of all
bundles to obtain bundle embeddings and introducing a goal of new-bundle generation based on
reinforcement learning. Li et al. [95] consider the problem of personalized outfit recommendation,
which can also be regarded as a kind of bundle recommendation. The authors construct a hier-
archical graph in which the users, items, and outfits are contained. GNN layers are deployed for
obtaining the representations of users and outfits. The learning of the GNN model follows a multi-
task manner. Zhao et al. [240] proposed a MIDGN model that combines the disentangle embedding
space and multi-layer embedding propagation, obtaining finer-grained user preference towards
items and bundles. The details of GNN models for bundle recommendation are shown in Table 10.
In summary, the data input of bundle recommendation can be well represented as graph-
structural data, especially for the bundles, which have been represented as a kind of new nodes.
4.3.5 GNN in Cross-domain Recommendation. Benefitting from the powerful capabilities, the
GNN-based recommendation model has gradually emerged in the cross-domain recommendation,
ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
3:30 C. Gao et al.

Fig. 16. Illustration of GNN models for cross-domain recommendation.

Table 11. Details of GNN Models of Cross-domain Recommendation

Model Graph Construction Network Design for Information Transferring

PPGN [238] cross-domain graph cross-domain propagation
BiTGCF [109] domain-specific graphs common user attributes
DAN [170] domain-specific graphs two-branched decoder
HeroGRAPH [35] cross-domain graph cross-domain propagation
DAGCN [58] cross-domain graph cross-domain propagation

as shown in Figure 16. The challenges include how to construct graph and design network archi-
tecture to transfer information across domains, as shown in Table 11. Zhao et al. [238] construct
the cross-domain graph with shared users and items from multiple domains. The proposed PPGN’s
embedding propagation layers can well learn the users’ preferences on multiple domains’ items
under a multi-task learning framework. Liu et al. [109] propose the bidirectional knowledge trans-
fer by regarding the shared users as the bridge. GNN layers are adopted to leverage the high-order
connectivity in the user-item interaction graph for better preference learning, and then it fuses
common features with domain-specific features. Guo et al. [58] construct the graph for each do-
main and deploy domain-specific GCN layers for learning user-specific embeddings. The authors
combine it with the attention mechanisms for adaptively choosing important neighbors during
the embedding propagation. Wang et al. [170] propose an encoder-decoder framework where the
encoder is implemented by graph convolutional networks. Specifically, the GCNs are deployed
on the user-item interaction graphs. Cui et al. [35] proposes to construct a heterogeneous graph,
where users and items in multiple domains can be well included. The GNN-based embedding prop-
agation is deployed on multiple domains, where a user/item can directly absorb the information
of different domains, where recurrent attention networks are used to distinguish important neigh-
bors. To sum up, the cross-domain recommendation is a general definition, which covers various
specific settings. For example, one setting may assume that the user can be overlapped, while
another setting may be on the contrary.
4.3.6 GNN in Multi-behavior Recommendation. Multiple types of behaviors can provide a
large amount of information to the recommender system, which helps the recommender system
to learn the user’s intentions better, thereby improving the recommendation performance. For

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
A Survey of Graph Neural Networks for Recommender Systems 3:31

Fig. 17. Illustration of GNN models for multi-behavior recommendation.

recommender systems based on graph neural networks, based on standard user-item bipartite
graphs, the multiple types of behaviors between users and items can naturally be modeled as
different types of edges between nodes. Therefore, most of the multi-behavior recommendation
methods based on graph neural networks are based on heterogeneous graphs. However, the focus
of multi-behavior recommendation is (1) how to model the relationship between multiple behav-
iors and target behavior and (2) how to model the semantics of the item through behavior, which
is shown in Figure 17.
To model the effect of auxiliary behaviors on target behaviors, the simplest method is to di-
rectly model all types of behaviors without considering the differences between behaviors. Zhang
et al. [235] construct all user behaviors in one graph and perform graph convolution operations.
Wang [177, 178] extracts each behavior from the graph to construct a subgraph, then learns from
the subgraph, and finally aggregates through the gating mechanism. Chen et al. [17] proposed an
intuitive method that assigns a representation to users, items, and behaviors. In the propagation
process, the representations of edges and neighbor nodes need to be composited first to obtain
a new node representation, and then it can be applied to the GNN method. Through the compo-
sition operation, the representation of the node is fused with different types of behaviors. Xia et
al. [206] also redesign the aggregation mechanism on the graph convolutional network to explic-
itly model the impact of different types of behavior. Jin et al. [78] assign different learnable weights
to different edges to model the importance of the behaviors. In addition, to capture complex multi-
behavior relationships, some works rely on knowledge. For example, Xia et al. [207] learn the
representations in different behavior spaces and then inject temporal contextual information into
the representations to model the user’s behavior dynamics, and finally, through the attention mech-
anism, discriminate the most important relationships and behaviors for the predicted target. Xia
et al. [208] use meta-graph networks to learn meta knowledge for different behaviors and then
transfer the learned meta-knowledge between different types of behaviors.
In addition to modeling the influence of different types of behaviors, different behaviors may
contain different meanings or semantics. For example, for items added to a shopping cart, users
may have similar preferences for them, or these items have a complementary relationship and
generally need to be purchased at the same time. If these items are connected through a graph, then
the representation of the item can be enhanced. To get better item representation, Yu et al. [223]
not only connect related items in the graph but also construct a new graph of the category that the
items belong to, which is used to enhance the representation of the item. It needs to construct the
item graph separately, and there are also some works that do not construct the graph separately
and directly in the user-item heterogeneous graph by using meta-path or second-order neighbors,

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
3:32 C. Gao et al.

Table 12. Details of GNN Models for Multi-behavior Recommendation

Model Graph Construction Network Design Multi-behavior Modeling
MB-GMN [208] user-item graph GCN / GAT Behavior representation
KHGT [207] user-item/item-item graph GAT Weight
MBGCN [78] user-item graph GAT Weight
MGNN-SPred [177] item-item graph GCN Weight
MGNN [235] user-item graph GCN Node representation
LP-MRGNN [178] item-item graph GCN Weight
GNNH [223] item-item/category-category graph GCN Item representation
GNMR [206] user-item graph GAT Weight
DMBGN [212] item-item graph GCN Graph representation
GHCF [17] user-item graph GCN Edge embedding

where similar items are aggregated to enhance their representation [78, 207]. These works’ details
are presented in Table 12.

4.4 GNN for Different Recommendation Objectives

4.4.1 GNN for Diversity. For individual-level diversity, retrieved items from recommender sys-
tems need to cover more topics such as different categories of products or different genres of music.
Therefore, utilizing GNN to increase diversity requires that the learned user embeddings be close
to item embeddings with various topics. However, as the embedding aggregation operation in GNN
makes user embeddings close to embeddings of items that are interacted in historical records, GNN
might discourage diversity by recommending too many similar items that belong to the dominant
topic in users’ interaction history. For example, the learned embedding from GNN for a user who
mainly interacts with electronics may be too close to embeddings of electronic items, making the
GNN only recommend electronics to this user, which leads to low diversity. Therefore, to overcome
the first challenge of weak signals from disadvantaged topics, efforts have been made to restrict
the importance of dominant topics (electronics in the above example) by constructing diversified
sub-graphs from the original user-item bipartite graph. Specifically, Sun et al. [159] propose a
model called Bayesian Graph Collaborative Filtering (BGCF), which constructs augmented
graphs with node copying [135] from high-order neighbors, such that items of diverse topics with
high similarity can be directly connected to user nodes. Zheng et al. [242] propose Diversified
Graph Convolutional Networks (DGCN) and conduct rebalanced neighbor sampling, which
down-weights dominant topics and boosts the importance of disadvantaged topics in neighbor
nodes. Figure 19 illustrates the comparison between BGCF and DGCN on graph construction for
diversity. Meanwhile, to address the second challenge of balancing between accuracy and diversity,
BGCF re-ranks the top items according to item popularity, and DGCN utilizes adversarial learning
on the item embeddings, which encourages GNN to capture user preferences that are largely inde-
pendent of item categories, which further improves recommendation diversity. The comparisons
are illustrated in Figure 18.
As for system-level diversity, the main target is to discover more relevant items from the long-
tail ones, which have much fewer training samples than those popular items. To address the weak
signals of long-tail items, Li et al. [94] propose a model called V2HT to construct an item graph
that explores item correlations with external knowledge. Specifically, four types of edges are intro-
duced, which connect frequent items and long-tail items. Then multiple GCN layers are stacked,
which propagates well-trained embeddings of frequent items to undertrained embeddings of long-
tail items. In this way, long-tail item embeddings of higher quality are obtained, since they share
the information from frequent items; thus, system-level diversity is improved with more recom-
mendation on long-tail items.

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
A Survey of Graph Neural Networks for Recommender Systems 3:33

Fig. 18. Illustration of BGCF [159] and DGCN [242] on achieving individual-level diversity.

Fig. 19. Illustration of V2HT [94], FH-HAT [214], and Tradeoff-Framework [76] on achieving system-level
diversity.

Table 13. Details of GNN Models for Diversified Recommendation (How They Address Challenges)

Model Weak Preference Signals Accuracy-diversity Balance

BGCF [159] Node Copying Re-rank
DGCN [242] Neighbor Sampling Adversarial training
V2HT [94] Utilize item correlations -
FH-HAT [214] Heterogeneous graph Diversity loss
Tradeoff-Framework [76] NN & FN graph Joint training

In addition, a few studies [76, 214] utilize GNN to improve both individual-level and system-level
diversity. Specifically, Xie et al. [214] propose FH-GAT, which addresses the challenge of weak
signals by constructing a heterogeneous interaction graph to express diverse user preferences. A
neighbor similarity-based loss is conducted on the heterogeneous graph to balance accuracy and
diversity. Isufi et al. [76] propose two GCN on the nearest neighbor (NN) graph and the furthest
neighbor (FN) graph, where NN guarantees accuracy and FN enhances the weak signals of diverse
items. Meanwhile, the two GCN are jointly optimized with a hyper-parameter to achieve a tradeoff
between accuracy and diversity. Table 13 shows the differences among the above approaches.
4.4.2 GNN for Explainability. With the proliferation of GNN, researchers also make endeavors
to improve the explainability of recommender systems with GNN’s power of modeling logical 1000
relations, as illustrated in Figure 20. He et al. [64] construct a heterogeneous graph with three kinds
of nodes, including the user, the item, and the aspect (the specific item property extracted from
textual reviews). Therefore, they cast the recommendation task into a ternary relation ranking

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
3:34 C. Gao et al.

Fig. 20. Illustration of GNN models for explainable recommendation.

task and propose TriRank with a high degree of explainability by explicitly modeling aspects in
reviews.
Inspired by this work, the following research further explores rich information in dimensions
of users and items, generally organized in the form of knowledge graph, to enhance the explain-
ability [236]. Ai et al. [5] construct a knowledge graph with entities, i.e., users and items, and
relations, e.g., “User A purchased Item B belonging to Category C.” Moreover, they embed each
entity for recommendation and adopt the shortest relation path between a user and an item in
the knowledge graph to indicate the recommendation explanations. Different from separately uti-
lizing embedding-based and path-based methods like Ai et al. [5], Wang et al. [171] proposed an
end-to-end framework, RippleNet, that combines the two knowledge graph-aware recommenda-
tion methods together. Here, the knowledge graph contains the related knowledge of the recom-
mended items, such as the type and author of a movie. In this way, explanations can be generated
by the path between users’ history to an item with high scores.
The meta-path-based utilization of knowledge graph can also benefit other specific recom-
mendation tasks, e.g., sequential recommendation [75, 185]. Huang et al. [75] extract semantic
meta-paths between a user and an item from knowledge graph to help sequential recommendation.
Further, they encode and rank the meta-paths to generate the recommendation list, and these
meta-paths also indicate the respective explanation. Similarly, Wang et al. [185] also leverage
knowledge graph to improve the performance of the sequential recommendation task and encode
the meta-path between a user and an item with recurrent neural networks. Chen et al. [21] further
model the temporal meta-paths by capturing historical item features and path-defined context
with neural networks.
However, these meta-path-based solutions also face some challenges in terms of how to obtain
these meta-paths [121, 211]. First, since pre-defined meta-paths require extensive domain knowl-
edge, Ma et al. [121] jointly combine the discovery of inductive rules (meta-paths) from the item-
centric knowledge graph, which equips the framework with explainability and the learning of
a rule-guided recommendation model. Moreover, to overcome the computational difficulties of
enumerating all potential meta-paths, Xian et al. [211] replace the enumeration method with the
reinforcement reasoning approach to identify proper meta-paths for scalability.
Besides meta-path-based solutions, Wang et al. [180] propose a new method named Knowledge
Graph Attention Network (KGAT), where the attention mechanism can offer explainability to

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
A Survey of Graph Neural Networks for Recommender Systems 3:35

Table 14. Details of GNN Models for Explainable Recommendation

Model Network Design Source of Explainability
ECFKG [5] Meta-path based GE Meta-path over knowledge graph
RippleNet [171] Meta-path based GraphSage Meta-path over knowledge graph
EIUM [75] Meta-path based GE Meta-path over knowledge graph
KPRN [185] Meta-path based GE Meta-path over knowledge graph
TMER [21] Meta-path based GE Meta-path over knowledge graph and temporal dependency
RuleRec [121] Meta-path based GE Discovered meta-path over knowledge graph
PGPR [211] Meta-path based GE Discovered meta-path over knowledge graph
KGAT [180] GAT Attention mechanism
HAGERec [219] GAT Attention mechanism

some extent. Yang et al. [219] develop a hierarchical attention graph convolutional network to
model higher-order relations in the heterogeneous knowledge graph, where the explainability is
also dependent on the attention mechanism. Explainability in recommendation includes design-
ing explainable model and finding explanations for those models without explainability (such as a
black-box model). That is, the explainable model can naturally generate explainable recommenda-
tion results. As for GNN-based recommendation models for explainability, they can utilize paths
on heterogenous graphs or knowledge graphs to generate recommendation. Since the model ar-
chitecture is explicit and explainable, they are classified into the first category of explainability in
recommendation. We compare the technical details in Table 14. The existing works of GNN-based
recommendation for explainability mostly extract the relations between attributes or behaviors
with recommendation results, which ignores the causality, which may lead to wrong explanations.
Designing suitable causal inference methods into graph neural network-based models is important
and promising.

4.4.3 GNN for Fairness. Despite the power of graph data in recommendation, it might inherit or
even amplify discrimination and the societal bias in recommendation [36, 147, 195]. Past research
has proven that compared with models that only adopt node attributes, the user unfairness is
magnified due to the utilization of graph structures [36].
To curb the fairness issue in recommendation, some researchers propose to learn fair graph
embeddings [10, 147, 195]. Rahman et al. [147] extend the well-known graph embedding method,
node2vec [56], to a more fair version, Fairwalk, which can generate a more diverse network neigh-
borhood representation for social recommendations by sampling the next node based on its sen-
sitive attributes. Thus, all nodes’ sensitive attributes are indispensable. Bose et al. [10] propose an
adversarial framework to minimize the sensitive information in graph embeddings with a discrimi-
nator enforcing the fairness constraints. Moreover, considering the evaluations of fairness can vary,
fairness constraints are flexible according to the task. However, similar to Fairwalk [147], all nodes’
sensitive attributes are required. Wu et al. [195] learn fair embeddings for recommendation from
a graph-based perspective. They propose FairGo, which adopts a graph-based adversarial learn-
ing method to map embeddings from any recommendation models into a sensitive-information-
filtered space, therefore eliminating potential leakage of sensitive information from both original
recommendation embeddings and user-centric graph structures. Indeed, except for learning fair
embeddings, Dai et al. [36] propose FairGNN, which learns fair GNN classifiers with limited known
sensitive attributes in an adversarial learning paradigm with fairness constraints. Different from
fair graph embeddings, fair GNN classifiers are to ensure node classification task (rather than
graph embeddings) independent of sensitive data. Moreover, they develop GNN-based sensitive
data estimators to overcome the issue of missing sensitive data in the real world. We compare the
technical details in Table 15.

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
3:36 C. Gao et al.

Table 15. Details of GNN Models for Fair Recommendation

Model Network Design Achieving Fairness

Fairwalk [147] Node2vec (GE) learning fair graph embeddings by sampling
CFCGE [10] Invariant GE learning fair graph embeddings by adversarial learning
FairGo [195] HetGNN(GNN) learning fair graph embeddings by adversarial learning
FairGNN [36] GraphSage(GNN) learning fair GNN classifiers by adversarial learning

Fig. 21. Illustration of GNN models for fair recommendation.

In short, existing GNN-based recommendation models for achieving fairness, as shown in Fig-
ure 21, always lead to performance drop to some extent, requiring solutions for addressing the
tradeoff dilemma.

4.5 GNN for Specific Recommendation Applications

Graph neural networks are also widely used to handle the specific challenges in different applica-
tions of recommender systems. As for e-commerce/product recommendation, most of the existing
works have been introduced in the above sections. Li et al. [100] propose to stack multiple GNN
modules and use a deterministic clustering algorithm to help improve the efficiency of GNN in
large-scale e-commerce applications. Liu et al. [112] propose to leverage the topology of item-
relations for building graph neural networks in e-commerce recommendation. As for the Point-of-
Interest recommendation, GGLR [13] uses the sequence of POIs a user visited to construct the POI-
POI graph, where the edges between POIs denote the frequency of users consecutively visiting two
POIs. Zhang et al. [231] propose to combine social networks and user-item interactions together,
which deploys embedding aggregation on both social-connected users and visited POIs. As for
news recommendation, Hu et al. [71] propose to introduce the preference disentanglement [179]
into the user-news embedding propagation.
There are many other papers for specific applications, but they can be well categorized into and
covered by the corresponding stages, scenarios, and objectives, and thus we omit them here.

5 OPEN PROBLEMS AND FUTURE DIRECTIONS

5.1 More Complex GNNs
5.1.1 Deeper GNN. Due to the over-smoothing problem, more and more studies focus on
properly increasing GNN’s layers to capture higher-order connectivity correlations on graphs as
well as improve models’ performance [20, 88, 150, 251]. Despite these advancements, there is no

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
A Survey of Graph Neural Networks for Recommender Systems 3:37

universal solution for constructing very deep GNN like CNN, and relevant works propose different
strategies. Lai et al. [88] develop a meta-policy to adaptively select the number of propagation for
each graph node through training with reinforcement learning (RL). The experiment results
show that partial nodes require more than three propagation layers to boost model performance.
Rong et al. [150] alleviate the over-smoothing problem by randomly removing graph edges, which
acts as a message-passing reducer. Claudio et al. [51] consider GNN as a dynamic system, and
learned representations are the system’s fixed points. Following this assumption, the transforma-
tion matrix in propagation is first fixed under stability conditions. Furthermore, only embeddings
are updated in the learning procedure, leaving the matrix untrained. In this way, GNN can be
trained faster and go deeper. Li et al. [92] transfer the concepts of residual/dense connections and
dilated convolutions from CNNs to assist deeper GNNs. As for future works, the performance leap
compared with current shallow GNNs should be a fundamental problem in developing very deep
GNNs, like groundbreaking works in the area of CNN [74, 164]. At the same time, the computation
and time complexity must also be acceptable.

5.1.2 Dynamic GNN. Existing GNN-based recommendation models are almost based on the
static graph, as mentioned above, while there are plenty of dynamics in recommender systems. For
example, in the sequential recommendation or session-based recommendation, the users’ data is
collected in a dynamic manner, naturally. In addition, modeling the dynamic user preferences is one
of the most critical challenges in these recommendation scenarios. In addition, the platform may
dynamically involve new users, new products, new features, and so on, which poses challenges to
static graph neural networks. Recently, dynamic graph neural networks [93, 123] have attracted
attention, which deploys embedding propagation operations on dynamically constructed graphs.
Given the time-evolving property of recommender systems, the dynamic graph neural network-
based recommendation model will be a promising research direction with broad applications in
the real world.

5.1.3 Efficiency and Scalability in Larger GNN. Early works on GNN follow the full-batch gra-
dient descent algorithm, where the whole adjacency matrix is multiplied on the node embeddings
during each inference step, which can not handle real-world recommender systems, since the num-
ber of nodes and edges can reach a million-level scale. A recent work [239] even proposes a GNN-
based solution for both the searching and recommendation engines, which results in higher re-
quirements of computational efficiency. Hamilton et al. propose GraphSAGE [62], which performs
neighbor sampling and only updates the related sub-graph instead of the whole graph during each
inference step. The sampling strategy is also adopted in a few other works [22, 29], which reduces
the computation complexity of GNN and improves the scalability. Ying et al. [222] successfully ap-
ply GraphSAGE to web-scale recommender systems, which can efficiently compute embeddings
for billions of items. Some other works [45] design specific query-recall acceleration strategies
for an efficient recommendation. In addition, a few open-source tools have been released that
can accelerate the research and development of GNN-based recommendation, such as PyG [48],
DGL [174], and AliGraph [252]. Recently, a graph learning system with both effectiveness and scal-
ability, PlatoGL [103], has been proposed. It provides a practical solution to combine large-scale
GNN with real-time recommendation, a fundamental problem for today’s industrial recommenda-
tions. We refer to another survey [1] for details on computing and accelerating GNN.
Despite these existing approaches, achieving large-scale GNN-based recommendation is still a
challenging task, especially in the ranking stage where thousands of features are involved, which
results in a large and complicated heterogeneous graph.

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
3:38 C. Gao et al.

5.2 Advanced Machine Learning Methods

5.2.1 Self-supervised GNN. The direct supervision from interaction data is relatively sparse
compared with the scale of the graph. Therefore, it is necessary to include more supervision signals
from the graph structure itself or the recommendation task. For example, Yu et al. [227] and Wu
et al. [194] attempt to enhance GNN-based recommendation by designing auxiliary tasks from
the graph structure with self-supervision. Data augmentations such as node dropout are utilized
to generate sample pairs for contrastive training. We believe it is a promising future direction to
leverage extra self-supervised tasks to learn meaningful and robust representations for GNN-based
recommender systems.

5.2.2 AutoML-enhanced GNN. Recommendation scenarios are diverse and vastly different from
each other; thus, there exists no silver bullet GNN model that can generalize across all scenar-
ios. Recently, AutoML (Automated Machine Learning) [221] has been proposed, which can
automatically design appropriate models for specific tasks. For GNN-based recommendation, the
search space is quite large, including multiple options for the neighbor sampler, aggregator, inter-
action function, and so on. Consequently, AutoML can, to a great extent, reduce human effort in
discovering advanced model structures. A few works [54, 72] have been proposed that search to
combine GNN layers and aggregate neighbors. Therefore, designing AutoML algorithms to search
for GNN-based recommender systems [188] is a promising future direction.

5.3 Others
5.3.1 About the Taxonomy. In this survey, we have presented a taxonomy for GNN-based
recommendation, in which we discuss the existing works from four aspects: stage, scenario, ob-
jective, and application. Here, we discuss the important future directions following the presented
taxonomy.
• Stage. The traditional recommendation divides the whole recommendation process into sev-
eral stages mainly due to the too-large size of the original candidate pool (million or even
billion). That is, different stages of recommender engines have different requirements for
models. Currently, with the fast growth of computing power, the boundaries between dif-
ferent stages get more and more blurred, which can be understood in two folds. First, the
model in the ranking stage starts to take a huge-size item pool and multiple user behaviors
into consideration. Second, the multiple objectives in the re-ranking stage, such as diversity,
can also be achieved by the model of the recall stage or ranking stage, making the re-ranking
stage not so necessary.
• Scenario. With the fast development of online information services, recommendation sce-
narios are also changing rapidly. For example, the user can interact with the recommender
systems more actively by chatting with the robot, defined as a new scenario of conversa-
tional recommendation, which we will discuss in detail later. Recently, some works [239]
consider the search and recommendation simultaneously, two core concepts of information
retrieval, which try to design a unified graph neural network model to provide two kinds of
services.
• Objectives. Existing works of GNN-based recommendation models have spent much effort
in achieving various objectives, including accuracy, diversity, explainability, fairness, and so
on. However, they are still suffering from the tradeoff dilemma, in which only one goal can
be optimized simultaneously. Recently, some works of deep learning-based recommendation,
such as MMOE [120], PLE [165], and so on, studied how to deploy multiple separate mod-
els, each of which is designed towards one objective, and then optimize them together. We

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
A Survey of Graph Neural Networks for Recommender Systems 3:39

believe that these techniques in multi-objective learning can also be well used in addressing
the tradeoff in GNN-based recommendation models.
• Applications. On the one hand, a powerful recommender system helps existing non-
personalized applications attract more users and promote platform profits. On the other
hand, relying on its benefit of both user-side and platform-side, the recommender system can
motivate some new forms of applications, such as micro-video recommendation in TikTok.
In these applications where content has abundant features, deploying cross-modal graph
neural network models in both content understanding and user behavior modeling is a
promising research direction.

5.3.2 KG-enhanced Recommendation with GNN. Recent research has demonstrated the power
of KG in recommendation by enriching the user-item bipartite graph with knowledge [53]. Specifi-
cally, the utilization of knowledge graph with GNN significantly addresses some practical issues in
recommender systems [53], for example, cold start problem [42, 232] and dynamicity [163]. More-
over, knowledge graph also offers a novel solution to some scenarios in recommendations, e.g.,
sequential recommendation [75, 185], and objectives, e.g., explainability [140, 180, 219]. There are
already many works, including KGCN [172], KGAT [180], KGIN [182], and so on. Currently, the ma-
jority of research first uses GNN to learn embeddings of knowledge graphs and then incorporate
these embeddings into the recommendation model so an end-to-end model can be trained [169].
Therefore, we point out that the utilization of KG in recommendation with GNN can be further
enhanced from the perspectives of data, scenario, and model. Specifically, the incorporated KG
mainly records the rich item-item relations, e.g., Movie A belongs to Category B [53], but user-
user knowledge is lacking in formal and plausible definition and thus substantially overlooked.
Future work can consider creating user-centric KG based on abundant knowledge in sociology.
Moreover, considering its success in scenarios, such as sequential recommendation [75, 185], it is
promising that the leverage of KG could further enhance recommendation quality from the aspects
that require more external knowledge, such as diversity and fairness. Further, existing methods of
leveraging KG in recommendation cannot fully model complex relations between a user and a spe-
cific item or its attributes. Thus, designing a better framework to carve out these complex relations
is another future direction.

5.3.3 Conversational Recommendation with GNN. In existing recommender systems, there may
exist the issue of information asymmetry that the system can only estimate users’ preferences
based on their historically collected behavior data. To address it, recently, conversational (inter-
active) recommendation researches [89, 115, 162, 204] proposed a new paradigm in which users
interact with the information system through conversations, and then new data can be dynami-
cally collected. Specifically, users can chat with the system to explicitly convey their consumption
demands or offer positive/negative feedback on the recommended items. As for future work, the
advances in representation learning with graph neural networks can be combined with preference
learning in the conversational recommendation.

6 CONCLUSION
There is a rapid development of graph neural network models in the research field of recommender
systems. This article provides an extensive survey systematically presenting the challenges, meth-
ods, and future directions in this area. Not only the history of development but also the most
recent advances are well covered and introduced. We hope this survey can well help both junior
and experienced researchers in the relative areas.

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
3:40 C. Gao et al.

REFERENCES
[1] Sergi Abadal, Akshay Jain, Robert Guirado, Jorge López-Alonso, and Eduard Alarcón. 2021. Computing graph neural
networks: A survey from algorithms to accelerators. Comput. Surv. 54, 9 (2021), 1–38.
[2] Himan Abdollahpouri, Robin Burke, and Bamshad Mobasher. 2017. Controlling popularity bias in learning-to-rank
recommendation. In Proceedings of the 11th ACM Conference on Recommender Systems. 42–46.
[3] Himan Abdollahpouri, Masoud Mansoury, Robin Burke, and Bamshad Mobasher. 2019. The unfairness of popularity
bias in recommendation. In Proceedings of the RecSys Workshop on Recommendation in Multistakeholder Environments
(RMSE). 1–7.
[4] Himan Abdollahpouri, Masoud Mansoury, Robin Burke, and Bamshad Mobasher. 2020. The connection between pop-
ularity bias, calibration, and fairness in recommendation. In Proceedings of the 14th ACM Conference on Recommender
Systems. 726–731.
[5] Qingyao Ai, Vahid Azizi, Xu Chen, and Yongfeng Zhang. 2018. Learning heterogeneous knowledge base embeddings
for explainable recommendation. Algorithms 11, 9 (2018), 137.
[6] Ting Bai, Youjie Zhang, Bin Wu, and Jian-Yun Nie. 2020. Temporal graph neural networks for social recommendation.
In Proceedings of the IEEE International Conference on Big Data. 898–903.
[7] Robert M. Bell and Yehuda Koren. 2007. Lessons from the Netflix prize challenge. ACM SIGKDD Explor. Newslett. 9,
2 (2007), 75–79.
[8] Rianne van den Berg, Thomas N. Kipf, and Max Welling. 2018. Graph convolutional matrix completion. In Proceedings
of the KDD Deep Learning Day. 1–7.
[9] Alex Beutel, Jilin Chen, Tulsee Doshi, Hai Qian, Li Wei, Yi Wu, Lukasz Heldt, Zhe Zhao, Lichan Hong, Ed H. Chi, et al.
2019. Fairness in recommendation ranking through pairwise comparisons. In Proceedings of the 25th ACM SIGKDD
International Conference on Knowledge Discovery & Data Mining. 2212–2220.
[10] Avishek Bose and William Hamilton. 2019. Compositional fairness constraints for graph embeddings. In Proceedings
of the International Conference on Machine Learning. PMLR, 715–724.
[11] Desheng Cai, Shengsheng Qian, Quan Fang, Jun Hu, Wenkui Ding, and Changsheng Xu. 2022. Heterogeneous graph
contrastive learning network for personalized micro-video recommendation. IEEE Trans. Multim. (2022), 1–1.
[12] Da Cao, Liqiang Nie, Xiangnan He, Xiaochi Wei, Shunzhi Zhu, and Tat-Seng Chua. 2017. Embedding factorization
models for jointly recommending items and user generated lists. In Proceedings of the 40th International ACM SIGIR
Conference on Research and Development in Information Retrieval. 585–594.
[13] Buru Chang, Gwanghoon Jang, Seoyoon Kim, and Jaewoo Kang. 2020. Learning graph-based geographical latent
representation for point-of-interest recommendation. In Proceedings of the 29th ACM International Conference on
Information & Knowledge Management. 135–144.
[14] Jianxin Chang, Chen Gao, Xiangnan He, Depeng Jin, and Yong Li. 2020. Bundle recommendation with graph con-
volutional networks. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in
Information Retrieval. 1673–1676.
[15] Jianxin Chang, Chen Gao, Yu Zheng, Yiqun Hui, Yanan Niu, Yang Song, Depeng Jin, and Yong Li. 2021. Sequential
recommendation with graph neural networks. In Proceedings of the 44th International ACM SIGIR Conference on
Research and Development in Information Retrieval. 378–387.
[16] Chen Chen, Jie Guo, and Bin Song. 2021. Dual attention transfer in session-based recommendation with multi-
dimensional integration. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development
in Information Retrieval. 869–878.
[17] Chong Chen, Weizhi Ma, Min Zhang, Zhaowei Wang, Xiuqiang He, Chenyang Wang, Yiqun Liu, and Shaoping Ma.
2021. Graph heterogeneous multi-relational recommendation. In Proceedings of the AAAI Conference on Artificial
Intelligence, Vol. 35. 3958–3966.
[18] Chong Chen, Min Zhang, Chenyang Wang, Weizhi Ma, Minming Li, Yiqun Liu, and Shaoping Ma. 2019. An efficient
adaptive transfer neural network for social-aware recommendation. In Proceedings of the 42nd International ACM
SIGIR Conference on Research and Development in Information Retrieval. 225–234.
[19] Deli Chen, Yankai Lin, Wei Li, Peng Li, Jie Zhou, and Xu Sun. 2020. Measuring and relieving the over-smoothing
problem for graph neural networks from the topological view. In Proceedings of the AAAI Conference on Artificial
Intelligence, Vol. 34. 3438–3445.
[20] Deli Chen, Yankai Lin, Wei Li, Peng Li, Jie Zhou, and Xu Sun. 2020. Measuring and relieving the over-smoothing
problem for graph neural networks from the topological view. In Proceedings of the AAAI Conference on Artificial
Intelligence, Vol. 34. 3438–3445.
[21] Hongxu Chen, Yicong Li, Xiangguo Sun, Guandong Xu, and Hongzhi Yin. 2021. Temporal meta-path guided ex-
plainable recommendation. In Proceedings of the 14th ACM International Conference on Web Search and Data Mining.
1056–1064.

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
A Survey of Graph Neural Networks for Recommender Systems 3:41

[22] Jie Chen, Tengfei Ma, and Cao Xiao. 2018. FastGCN: Fast learning with graph convolutional networks via importance
sampling. In Proceedings of the International Conference on Learning Representations. 1–15.
[23] Liang Chen, Yang Liu, Xiangnan He, Lianli Gao, and Zibin Zheng. 2019. Matching user with item set: Collaborative
bundle recommendation with deep attention network. In Proceedings of the 28th International Joint Conference on
Artificial Intelligence. 2095–2101.
[24] Tianwen Chen and Raymond Chi-Wing Wong. 2020. Handling information loss of graph neural networks for session-
based recommendation. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery &
Data Mining. 1172–1180.
[25] Tianwen Chen and Raymond Chi-Wing Wong. 2021. An efficient and effective framework for session-based social
recommendation. In Proceedings of the 14th ACM International Conference on Web Search and Data Mining. 400–408.
[26] Xu Chen, Yongfeng Zhang, and Ji-Rong Wen. 2022. Measuring why in recommender systems: A comprehensive
survey on the evaluation of explainable recommendation. arXiv preprint arXiv:2202.06466 (2022).
[27] Chen Cheng, Haiqin Yang, Michael R. Lyu, and Irwin King. 2013. Where you like to go next: Successive point-of-
interest recommendation. In Proceedings of the 23rd International Joint Conference on Artificial Intelligence. 2605–2611.
[28] Heng-Tze Cheng, Levent Koc, Jeremiah Harmsen, Tal Shaked, Tushar Chandra, Hrishi Aradhye, Glen Anderson,
Greg Corrado, Wei Chai, Mustafa Ispir, et al. 2016. Wide & deep learning for recommender systems. In Proceedings
of the 1st Workshop on Deep Learning for Recommender Systems. 7–10.
[29] Wei-Lin Chiang, Xuanqing Liu, Si Si, Yang Li, Samy Bengio, and Cho-Jui Hsieh. 2019. Cluster-GCN: An efficient algo-
rithm for training deep and large graph convolutional networks. In Proceedings of the 25th ACM SIGKDD International
Conference on Knowledge Discovery & Data Mining. 257–266.
[30] Janneth Chicaiza and Priscila Valdiviezo-Diaz. 2021. A comprehensive survey of knowledge graph-based recom-
mender systems: Technologies, development, and contributions. Information 12, 6 (2021), 232.
[31] Minjin Choi, Jinhong Kim, Joonseok Lee, Hyunjung Shim, and Jongwuk Lee. 2022. S-walk: Accurate and scalable
session-based recommendation with random walks. In Proceedings of the 15th ACM International Conference on Web
Search and Data Mining. 150–160.
[32] Yunfei Chu, Jiangchao Yao, Chang Zhou, and Hongxia Yang. 2022. Graph neural networks in modern recommender
systems. In Graph Neural Networks: Foundations, Frontiers, and Applications. Springer, 423–445.
[33] Robert B. Cialdini and Noah J. Goldstein. 2004. Social influence: Compliance and conformity. Ann. Rev. Psychol.
55 (2004), 591–621.
[34] Paul Covington, Jay Adams, and Emre Sargin. 2016. Deep neural networks for YouTube recommendations. In Pro-
ceedings of the 10th ACM Conference on Recommender Systems. 191–198.
[35] Qiang Cui, Tao Wei, Yafeng Zhang, and Qing Zhang. 2020. HeroGRAPH: A heterogeneous graph framework for multi-
target cross-domain recommendation. In Proceedings of ORSUM@ RecSys: 3rd Workshop on Online Recommender
Systems and User Modeling.
[36] Enyan Dai and Suhang Wang. 2021. Say no to the discrimination: Learning fair graph neural networks with limited
sensitive attribute information. In Proceedings of the 14th ACM International Conference on Web Search and Data
Mining, 680–688.
[37] James Davidson, Benjamin Liebald, Junning Liu, Palash Nandy, Taylor Van Vleet, Ullas Gargi, Sujoy Gupta, Yu He,
Mike Lambert, Blake Livingston, et al. 2010. The YouTube video recommendation system. In Proceedings of the 4th
ACM Conference on Recommender Systems. 293–296.
[38] Qilin Deng, Kai Wang, Minghao Zhao, Zhene Zou, Runze Wu, Jianrong Tao, Changjie Fan, and Liang Chen. 2020.
Personalized bundle recommendation in online games. In Proceedings of the 29th ACM International Conference on
Information & Knowledge Management. 2381–2388.
[39] Yue Deng. 2021. Recommender systems based on graph embedding techniques: A comprehensive review. arXiv
preprint arXiv:2109.09587 (2021).
[40] Zhi-Hong Deng, Chang-Dong Wang, Ling Huang, Jian-Huang Lai, and Philip S. Yu. 2022. G3 SR: Global graph guided
session-based recommendation. IEEE Trans. Neural Netw. Learn. Syst. (2022), 1–14.
[41] Virginie Do, Sam Corbett-Davies, Jamal Atif, and Nicolas Usunier. 2022. Online certification of preference-based fair-
ness for personalized recommender systems. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 36.
6532–6540.
[42] Shaohua Fan, Junxiong Zhu, Xiaotian Han, Chuan Shi, Linmei Hu, Biyu Ma, and Yongliang Li. 2019. Metapath-
guided heterogeneous graph neural network for intent recommendation. In Proceedings of the 25th ACM SIGKDD
International Conference on Knowledge Discovery & Data Mining. 2478–2486.
[43] Wenqi Fan, Yao Ma, Qing Li, Yuan He, Eric Zhao, Jiliang Tang, and Dawei Yin. 2019. Graph neural networks for
social recommendation. In Proceedings of the World Wide Web Conference. 417–426.
[44] Ziwei Fan, Zhiwei Liu, Jiawei Zhang, Yun Xiong, Lei Zheng, and Philip S. Yu. 2021. Continuous-time sequential rec-
ommendation with temporal graph collaborative transformer. In Proceedings of the 30th ACM International Conference
on Information & Knowledge Management. 433–442.

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
3:42 C. Gao et al.

[45] Chao Feng, Defu Lian, Xiting Wang, Zheng Liu, Xing Xie, and Enhong Chen. 2023. Reinforcement routing on prox-
imity graph for efficient recommendation. ACM Trans. Inf. Syst. 41, 1 (2023), 1046–8188.
[46] Shanshan Feng, Xutao Li, Yifeng Zeng, Gao Cong, Yeow Meng Chee, and Quan Yuan. 2015. Personalized ranking
metric embedding for next new POI recommendation. In Proceedings of the 24th International Joint Conference on
Artificial Intelligence.
[47] Yifan Feng, Haoxuan You, Zizhao Zhang, Rongrong Ji, and Yue Gao. 2019. Hypergraph neural networks. In Proceed-
ings of the AAAI Conference on Artificial Intelligence, Vol. 33. 3558–3565.
[48] Matthias Fey and Jan Eric Lenssen. 2019. Fast graph representation learning with PyTorch geometric. In Proceedings
of the Workshop of International Conference on Learning Representations. 1–9.
[49] Alex Fout, Jonathon Byrd, Basir Shariat, and Asa Ben-Hur. 2017. Protein interface prediction using graph convolu-
tional networks. Adv. Neural Inf. Process. Syst. 30 (2017), 1–10.
[50] Wenjing Fu, Zhaohui Peng, Senzhang Wang, Yang Xu, and Jin Li. 2019. Deeply fusing reviews and contents for cold
start users in cross-domain recommendation systems. In Proceedings of the AAAI Conference on Artificial Intelligence,
Vol. 33. 94–101.
[51] Claudio Gallicchio and Alessio Micheli. 2020. Fast and deep graph neural networks. In Proceedings of the AAAI
Conference on Artificial Intelligence, Vol. 34. 3898–3905.
[52] Chen Gao, Xiangning Chen, Fuli Feng, Kai Zhao, Xiangnan He, Yong Li, and Depeng Jin. 2019. Cross-domain recom-
mendation without sharing user-relevant data. In Proceedings of the World Wide Web Conference. 491–502.
[53] Yang Gao, Yi-Fan Li, Yu Lin, Hang Gao, and Latifur Khan. 2020. Deep learning on knowledge graph for recommender
system: A survey. arXiv preprint arXiv:2004.00387 (2020).
[54] Yang Gao, Hong Yang, Peng Zhang, Chuan Zhou, and Yue Hu. 2021. Graph neural architecture search. In Proceedings
of the 29th International Conference on International Joint Conferences on Artificial Intelligence. 1403–1409.
[55] Sahin Cem Geyik, Stuart Ambler, and Krishnaram Kenthapadi. 2019. Fairness-aware ranking in search & recom-
mendation systems with application to LinkedIn talent search. In Proceedings of the 25th ACM SIGKDD International
Conference on Knowledge Discovery & Data Mining. 2221–2231.
[56] Aditya Grover and Jure Leskovec. 2016. node2vec: Scalable feature learning for networks. In Proceedings of the 22nd
ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 855–864.
[57] Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He. 2017. DeepFM: A factorization-machine
based neural network for CTR prediction. In Proceedings of the 26th International Joint Conference on Artificial Intel-
ligence. 1725–1731.
[58] Lei Guo, Li Tang, Tong Chen, Lei Zhu, Quoc Viet Hung Nguyen, and Hongzhi Yin. 2021. DA-GCN: A domain-aware
attentive graph convolution network for shared-account cross-domain sequential recommendation. In Proceedings
of the 30th International Joint Conference on Artificial Intelligence. 2483–2489.
[59] Qingyu Guo, Fuzhen Zhuang, Chuan Qin, Hengshu Zhu, Xing Xie, Hui Xiong, and Qing He. 2022. A survey on
knowledge graph-based recommender systems. IEEE Trans. Knowl. Data Eng. 34, 8 (2022), 3549–3568.
[60] Wei Guo, Rong Su, Renhao Tan, Huifeng Guo, Yingxue Zhang, Zhirong Liu, Ruiming Tang, and Xiuqiang He. 2021.
Dual graph enhanced embedding neural network for CTR prediction. In Proceedings of the 27th ACM SIGKDD Con-
ference on Knowledge Discovery & Data Mining. 496–504.
[61] Zhiwei Guo and Heng Wang. 2020. A deep graph neural network-based mechanism for social recommendations.
IEEE Trans. Industr. Inform. 17, 4 (2020), 2776–2783.
[62] William L. Hamilton, Rex Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs. In Pro-
ceedings of the 31st International Conference on Neural Information Processing Systems. 1025–1035.
[63] David K. Hammond, Pierre Vandergheynst, and Rémi Gribonval. 2011. Wavelets on graphs via spectral graph theory.
Appl. Computat. Harmon. Anal. 30, 2 (2011), 129–150.
[64] Xiangnan He, Tao Chen, Min-Yen Kan, and Xiao Chen. 2015. TriRank: Review-aware explainable recommendation
by modeling aspects. In Proceedings of the 24th ACM International on Conference on Information and Knowledge Man-
agement. 1661–1670.
[65] Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. 2020. LightGCN: Simplifying
and powering graph convolution network for recommendation. In Proceedings of the 43rd International ACM SIGIR
Conference on Research and Development in Information Retrieval. 639–648.
[66] Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural collaborative filtering.
In Proceedings of the 26th International Conference on World Wide Web. 173–182.
[67] Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk. 2016. Session-based recommendations
with recurrent neural networks. In Proceedings of the International Conference on Learning Representations. 1–10.
[68] Sepp Hochreiter and Júrgen Schmidhuber. 1997. Long short-term memory. Neural Computat. 9, 8 (1997), 1735–1780.
[69] Cheng Hsu and Cheng-Te Li. 2021. RetaGNN: Relational temporal attentive graph neural networks for holistic se-
quential recommendation. In Proceedings of the Web Conference. 2968–2979.

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
A Survey of Graph Neural Networks for Recommender Systems 3:43

[70] Guangneng Hu, Yu Zhang, and Qiang Yang. 2018. CoNet: Collaborative cross networks for cross-domain recommen-
dation. In Proceedings of the 27th ACM International Conference on Information and Knowledge Management. 667–676.
[71] Linmei Hu, Siyong Xu, Chen Li, Cheng Yang, Chuan Shi, Nan Duan, Xing Xie, and Ming Zhou. 2020. Graph neural
news recommendation with unsupervised preference disentanglement. In Proceedings of the 58th Annual Meeting of
the Association for Computational Linguistics. 4255–4264.
[72] Zhao Huan, Yao Quanming, and Tu Weiwei. 2021. Search to aggregate neighborhood for graph neural network. In
Proceedings of the IEEE 37th International Conference on Data Engineering. 552–563.
[73] Chao Huang, Huance Xu, Yong Xu, Peng Dai, Lianghao Xia, Mengyin Lu, Liefeng Bo, Hao Xing, Xiaoping Lai, and
Yanfang Ye. 2021. Knowledge-aware coupled graph neural network for social recommendation. In Proceedings of the
AAAI Conference on Artificial Intelligence.
[74] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q. Weinberger. 2017. Densely connected convolutional
networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 4700–4708.
[75] Xiaowen Huang, Quan Fang, Shengsheng Qian, Jitao Sang, Yan Li, and Changsheng Xu. 2019. Explainable interaction-
driven user modeling over knowledge graph for sequential recommendation. In Proceedings of the 27th ACM Inter-
national Conference on Multimedia. 548–556.
[76] Elvin Isufi, Matteo Pocchiari, and Alan Hanjalic. 2021. Accuracy-diversity trade-off in recommender systems via
graph convolutions. Inf. Process. Manag. 58, 2 (2021), 102459.
[77] Wensen Jiang, Yizhu Jiao, Qingqin Wang, Chuanming Liang, Lijie Guo, Yao Zhang, Zhijun Sun, Yun Xiong, and
Yangyong Zhu. 2022. Triangle graph interest network for click-through rate prediction. In Proceedings of the 15th
ACM International Conference on Web Search and Data Mining. 401–409.
[78] Bowen Jin, Chen Gao, Xiangnan He, Depeng Jin, and Yong Li. 2020. Multi-behavior recommendation with graph
convolutional networks. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development
in Information Retrieval. 659–668.
[79] SeongKu Kang, Junyoung Hwang, Dongha Lee, and Hwanjo Yu. 2019. Semi-supervised learning for cross-domain
recommendation to cold-start users. In Proceedings of the 28th ACM International Conference on Information and
Knowledge Management. 1563–1572.
[80] Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recommendation. In Proceedings of the IEEE
International Conference on Data Mining. IEEE, 197–206.
[81] Wang-Cheng Kang and Julian McAuley. 2019. Candidate generation with binary codes for large-scale Top-N rec-
ommendation. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management.
1523–1532.
[82] Kyung-Min Kim, Donghyun Kwak, Hanock Kwak, Young-Jin Park, Sangkwon Sim, Jae-Han Cho, Minkyu Kim,
Jihun Kwon, Nako Sung, and Jung-Woo Ha. 2019. Tripartite heterogeneous graph propagation for large-scale so-
cial recommendation. arXiv preprint arXiv:1908.02569 (2019).
[83] Thomas N. Kipf and Max Welling. 2017. Semi-supervised classification with graph convolutional networks. In Pro-
ceedings of the International Conference on Learning Representations. 1–14.
[84] Yehuda Koren. 2009. The BellKor solution to the Netflix grand prize. Netflix Prize Docum. 81, 2009 (2009), 1–10.
[85] Yehuda Koren, Robert Bell, and Chris Volinsky. 2009. Matrix factorization techniques for recommender systems.
Computer 42, 8 (2009), 30–37.
[86] Johannes Kunkel, Tim Donkers, Lisa Michael, Catalin-Mihai Barbu, and Jürgen Ziegler. 2019. Let me explain: Impact
of personal and impersonal explanations on trust in recommender systems. In Proceedings of the CHI Conference on
Human Factors in Computing Systems. 1–12.
[87] Valerio La Gatta, Vincenzo Moscato, Mirko Pennone, Marco Postiglione, and Giancarlo Sperlí. 2022. Music recom-
mendation via hypergraph embedding. IEEE Trans. Neural Netw. Learn. Syst. (2022), 1–13.
[88] Kwei-Herng Lai, Daochen Zha, Kaixiong Zhou, and Xia Hu. 2020. Policy-GNN: Aggregation optimization for graph
neural networks. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data
Mining. 461–471.
[89] Wenqiang Lei, Gangyi Zhang, Xiangnan He, Yisong Miao, Xiang Wang, Liang Chen, and Tat-Seng Chua. 2020. In-
teractive path reasoning on graph for conversational recommendation. In Proceedings of the 26th ACM SIGKDD
International Conference on Knowledge Discovery & Data Mining. 2073–2083.
[90] Jurek Leonhardt, Avishek Anand, and Megha Khosla. 2018. User fairness in recommender systems. In Proceedings of
the Web Conference. 101–102.
[91] Chen Li, Linmei Hu, Chuan Shi, Guojie Song, and Yuanfu Lu. 2021. Sequence-aware heterogeneous graph neural
collaborative filtering. In Proceedings of the SIAM International Conference on Data Mining. SIAM, 64–72.
[92] Guohao Li, Matthias Muller, Ali Thabet, and Bernard Ghanem. 2019. DeepGCNs: Can GCNs go as deep as CNNs? In
Proceedings of the IEEE/CVF International Conference on Computer Vision. 9267–9276.

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
3:44 C. Gao et al.

[93] Maosen Li, Siheng Chen, Yangheng Zhao, Ya Zhang, Yanfeng Wang, and Qi Tian. 2020. Dynamic multiscale graph
neural networks for 3D skeleton based human motion prediction. In Proceedings of the IEEE/CVF Conference on
Computer Vision and Pattern Recognition. 214–223.
[94] Mengmeng Li, Tian Gan, Meng Liu, Zhiyong Cheng, Jianhua Yin, and Liqiang Nie. 2019. Long-tail hashtag recommen-
dation for micro-videos with graph convolutional network. In Proceedings of the 28th ACM International Conference
on Information and Knowledge Management. 509–518.
[95] Xingchen Li, Xiang Wang, Xiangnan He, Long Chen, Jun Xiao, and Tat-Seng Chua. 2020. Hierarchical fashion graph
network for personalized outfit recommendation. In Proceedings of the 43rd International ACM SIGIR Conference on
Research and Development in Information Retrieval. 159–168.
[96] Yunqi Li, Hanxiong Chen, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. 2021. User-oriented fairness in recom-
mendation. In Proceedings of the Web Conference. 624–632.
[97] Yang Li, Tong Chen, Yadan Luo, Hongzhi Yin, and Zi Huang. 2021. Discovering collaborative signals for next POI
recommendation with iterative Seq2Graph augmentation. In Proceedings of the 30th International Joint Conference on
Artificial Intelligence. 1491–1497.
[98] Yunqi Li, Yingqiang Ge, and Yongfeng Zhang. 2021. Tutorial on fairness of machine learning in recommender systems.
In Proceedings of the International ACM SIGIR Conference on Research and Development in Information Retrieval.
[99] Zekun Li, Zeyu Cui, Shu Wu, Xiaoyu Zhang, and Liang Wang. 2019. Fi-GNN: Modeling feature interactions via
graph neural networks for CTR prediction. In Proceedings of the 28th ACM International Conference on Information
and Knowledge Management. 539–548.
[100] Zhao Li, Xin Shen, Yuhang Jiao, Xuming Pan, Pengcheng Zou, Xianling Meng, Chengwei Yao, and Jiajun Bu. 2020.
Hierarchical bipartite graph neural networks: Towards large-scale e-commerce applications. In Proceedings of the
IEEE 36th International Conference on Data Engineering. 1677–1688.
[101] Jianxun Lian, Xiaohuan Zhou, Fuzheng Zhang, Zhongxia Chen, Xing Xie, and Guangzhong Sun. 2018. xDeepFM:
Combining explicit and implicit feature interactions for recommender systems. In Proceedings of the 24th ACM
SIGKDD International Conference on Knowledge Discovery & Data Mining. 1754–1763.
[102] Nicholas Lim, Bryan Hooi, See-Kiong Ng, Xueou Wang, Yong Liang Goh, Renrong Weng, and Jagannadan Varadara-
jan. 2020. STP-UDGAT: Spatial-temporal-preference user dimensional graph attention network for next POI rec-
ommendation. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management.
845–854.
[103] Dandan Lin, Shijie Sun, Jingtao Ding, Xuehan Ke, Hao Gu, Xing Huang, Chonggang Song, Xuri Zhang, Lingling Yi,
Jie Wen, and Chuan Chen. 2022. PlatoGL: Effective and scalable deep graph learning system for graph-enhanced
real-time recommendation. In Proceedings of the Conference on Information and Knowledge Management.
[104] Fan Liu, Zhiyong Cheng, Lei Zhu, Zan Gao, and Liqiang Nie. 2021. Interest-aware message-passing gcn for recom-
mendation. In Proceedings of the Web Conference. 1296–1305.
[105] Fan Liu, Zhiyong Cheng, Lei Zhu, Chenghao Liu, and Liqiang Nie. 2022. A2-GCN: An attribute-aware attentive GCN
model for recommendation. IEEE Trans. Knowl. Data Eng. 34, 9 (2022), 4077–4088.
[106] Feng Liu, Weiwen Liu, Xutao Li, and Yunming Ye. 2020. Inter-sequence enhanced framework for personalized se-
quential recommendation. arXiv preprint arXiv:2004.12118 (2020).
[107] Han Liu, Yinwei Wei, Jianhua Yin, and Liqiang Nie. 2022. HS-GCN: Hamming spatial graph convolutional networks
for recommendation. IEEE Trans. Knowl. Data Eng. (2022), 1–1.
[108] Jiawei Liu, Chuan Shi, Cheng Yang, Zhiyuan Lu, and Philip S. Yu. 2022. A survey on heterogeneous information
network based recommender systems: Concepts, methods, applications and resources. AI Open 3 (2022), 40–57.
[109] Meng Liu, Jianjun Li, Guohui Li, and Peng Pan. 2020. Cross domain recommendation via bi-directional transfer graph
collaborative filtering networks. In Proceedings of the 29th ACM International Conference on Information & Knowledge
Management. 885–894.
[110] Qi Liu, Ruobing Xie, Lei Chen, Shukai Liu, Ke Tu, Peng Cui, Bo Zhang, and Leyu Lin. 2020. Graph neural network
for tag ranking in tag-enhanced video recommendation. In Proceedings of the 29th ACM International Conference on
Information & Knowledge Management. 2613–2620.
[111] Weiwen Liu, Qing Liu, Ruiming Tang, Junyang Chen, Xiuqiang He, and Pheng Ann Heng. 2020. Personalized re-
ranking with item relationships for e-commerce. In Proceedings of the 29th ACM International Conference on Infor-
mation & Knowledge Management. 925–934.
[112] Weiwen Liu, Yin Zhang, Jianling Wang, Yun He, James Caverlee, Patrick P. K. Chan, Daniel S. Yeung, and Pheng-
Ann Heng. 2022. Item relationship graph neural networks for e-commerce. IEEE Trans. Neural Netw. Learn. Syst. 33,
9 (2022), 4785–4799.
[113] Yang Liu, Chen Liang, Xiangnan He, Jiaying Peng, Zibin Zheng, and Jie Tang. 2022. Modelling high-order social
relations for item recommendation. IEEE Trans. Knowl. Data Eng. 34, 9 (2022), 4385–4397.

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
A Survey of Graph Neural Networks for Recommender Systems 3:45

[114] Yidan Liu, Min Xie, and Laks V. S. Lakshmanan. 2014. Recommending user generated item lists. In Proceedings of the
8th ACM Conference on Recommender Systems. 185–192.
[115] Zeming Liu, Ding Zhou, Hao Liu, Haifeng Wang, Zheng-Yu Niu, Hua Wu, Wanxiang Che, Ting Liu, and Hui Xiong.
2022. Graph-grounded goal planning for conversational recommendation. IEEE Trans. Knowl. Data Eng. (2022), 1–1.
[116] Dongsheng Luo, Yuchen Bian, Xiang Zhang, and Jun Huan. 2020. Attentive social recommendation: Towards user
and item diversities. arXiv preprint arXiv:2011.04797 (2020).
[117] Xusheng Luo, Luxin Liu, Yonghua Yang, Le Bo, Yuanpeng Cao, Jinghang Wu, Qiang Li, Keping Yang, and Kenny Q.
Zhu. 2020. AliCoCo: Alibaba e-commerce cognitive concept net. In Proceedings of the ACM SIGMOD International
Conference on Management of Data. 313–327.
[118] Ziyu Lyu, Yue Wu, Junjie Lai, Min Yang, Chengming Li, and Wei Zhou. 2022. Knowledge enhanced graph neural
networks for explainable recommendation. IEEE Trans. Knowl. Data Eng. (2022), 1–1.
[119] Chen Ma, Liheng Ma, Yingxue Zhang, Jianing Sun, Xue Liu, and Mark Coates. 2020. Memory augmented graph
neural networks for sequential recommendation. In Proceedings of the AAAI Conference on Artificial Intelligence,
Vol. 34. 5045–5052.
[120] Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H. Chi. 2018. Modeling task relationships in multi-
task learning with multi-gate mixture-of-experts. In Proceedings of the 24th ACM SIGKDD International Conference
on Knowledge Discovery & Data Mining. 1930–1939.
[121] Weizhi Ma, Min Zhang, Yue Cao, Woojeong Jin, Chenyang Wang, Yiqun Liu, Shaoping Ma, and Xiang Ren. 2019.
Jointly learning explainable rules for recommendation with knowledge graph. In Proceedings of the World Wide Web
Conference. 1210–1221.
[122] Xiao Ma, Liqin Zhao, Guan Huang, Zhi Wang, Zelin Hu, Xiaoqiang Zhu, and Kun Gai. 2018. Entire space multi-task
model: An effective approach for estimating post-click conversion rate. In Proceedings of the 41st International ACM
SIGIR Conference on Research & Development in Information Retrieval. 1137–1140.
[123] Yao Ma, Ziyi Guo, Zhaocun Ren, Jiliang Tang, and Dawei Yin. 2020. Streaming graph neural networks. In Proceedings
of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval. 719–728.
[124] Tong Man, Huawei Shen, Xiaolong Jin, and Xueqi Cheng. 2017. Cross-domain recommendation: An embedding and
mapping approach. In Proceedings of the 26th International Joint Conference on Artificial Intelligence. 2464–2470.
[125] Masoud Mansoury, Himan Abdollahpouri, Mykola Pechenizkiy, Bamshad Mobasher, and Robin Burke. 2020. Fair-
Match: A graph-based approach for improving aggregate diversity in recommender systems. In Proceedings of the
28th ACM Conference on User Modeling, Adaptation and Personalization. 154–162.
[126] Julian McAuley and Jure Leskovec. 2013. Hidden factors and hidden topics: Understanding rating dimensions with
review text. In Proceedings of the 7th ACM Conference on Recommender Systems. 165–172.
[127] Miller McPherson, Lynn Smith-Lovin, and James M. Cook. 2001. Birds of a feather: Homophily in social networks.
Ann. Rev. Sociol. 27, 1 (2001), 415–444.
[128] Rishabh Mehrotra, James McInerney, Hugues Bouchard, Mounia Lalmas, and Fernando Diaz. 2018. Towards a fair
marketplace: Counterfactual evaluation of the trade-off between relevance, fairness & satisfaction in recommenda-
tion systems. In Proceedings of the 27th ACM International Conference on Information and Knowledge Management.
2243–2251.
[129] Wenjing Meng, Deqing Yang, and Yanghua Xiao. 2020. Incorporating user micro-behaviors and item knowledge into
multi-task learning for session-based recommendation. In Proceedings of the 43rd International ACM SIGIR Conference
on Research and Development in Information Retrieval. 1091–1100.
[130] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient estimation of word representations in
vector space. In Proceedings of the International Conference on Learning Representations. 1–12.
[131] Weiqing Min, Shuqiang Jiang, and Ramesh Jain. 2019. Food recommendation: Framework, existing solutions, and
challenges. IEEE Trans. Multim. 22, 10 (2019), 2659–2671.
[132] Nima Mirbakhsh and Charles X. Ling. 2015. Improving Top-N recommendation for cold-start users via cross-domain
information. ACM Trans. Knowl. Discov. Data 9, 4 (2015), 1–19.
[133] Nan Mu, Daren Zha, Yuanye He, and Zhihao Tang. 2019. Graph attention networks for neural social recommendation.
In Proceedings of the IEEE 31st International Conference on Tools with Artificial Intelligence. 1320–1327.
[134] Shumpei Okura, Yukihiro Tagami, Shingo Ono, and Akira Tajima. 2017. Embedding-based news recommendation
for millions of users. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and
Data Mining. 1933–1942.
[135] Soumyasundar Pal, Florence Regol, and Mark Coates. 2019. Bayesian graph convolutional neural networks using
node copying. In Proceedings of the ICML Workshop on Learning and Reasoning with Graph-structured Data.
[136] Zhiqiang Pan, Fei Cai, Wanyu Chen, Chonghao Chen, and Honghui Chen. 2022. Collaborative graph learning for
session-based recommendation. ACM Trans. Inf. Syst. 40, 4 (2022), 1–26.

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
3:46 C. Gao et al.

[137] Zhiqiang Pan, Fei Cai, Wanyu Chen, Honghui Chen, and Maarten de Rijke. 2020. Star graph neural networks for
session-based recommendation. In Proceedings of the 29th ACM International Conference on Information & Knowledge
Management. 1195–1204.
[138] Yitong Pang, Lingfei Wu, Qi Shen, Yiming Zhang, Zhihua Wei, Fangli Xu, Ethan Chang, Bo Long, and Jian Pei. 2022.
Heterogeneous global graph neural networks for personalized session-based recommendation. In Proceedings of the
15th ACM International Conference on Web Search and Data Mining. 775–783.
[139] Ioannis Paparrizos, B. Barla Cambazoglu, and Aristides Gionis. 2011. Machine learned job recommendation. In Pro-
ceedings of the 5th ACM Conference on Recommender Systems. 325–328.
[140] Sung-Jun Park, Dong-Kyu Chae, Hong-Kyun Bae, Sumin Park, and Sang-Wook Kim. 2022. Reinforcement learning
over sentiment-augmented knowledge graphs towards accurate and explainable recommendation. In Proceedings of
the 15th ACM International Conference on Web Search and Data Mining. 784–793.
[141] Apurva Pathak, Kshitiz Gupta, and Julian McAuley. 2017. Generating and personalizing bundle recommendations
on steam. In Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information
Retrieval. 1073–1076.
[142] Georgina Peake and Jun Wang. 2018. Explanation mining: Post hoc interpretability of latent factor models for rec-
ommendation systems. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery &
Data Mining. 2060–2069.
[143] Changhua Pei, Yi Zhang, Yongfeng Zhang, Fei Sun, Xiao Lin, Hanxiao Sun, Jian Wu, Peng Jiang, Junfeng Ge, Wenwu
Ou, et al. 2019. Personalized re-ranking for recommendation. In Proceedings of the 13th ACM Conference on Recom-
mender Systems. 3–11.
[144] Bryan Perozzi, Rami Al-Rfou, and Steven Skiena. 2014. DeepWalk: Online learning of social representations. In Pro-
ceedings of the 20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 701–710.
[145] Ruihong Qiu, Jingjing Li, Zi Huang, and Hongzhi Yin. 2019. Rethinking the item order in session-based recommen-
dation with graph neural networks. In Proceedings of the 28th ACM International Conference on Information and
Knowledge Management. 579–588.
[146] Ruihong Qiu, Hongzhi Yin, Zi Huang, and Tong Chen. 2020. GAG: Global attributed graph neural network for stream-
ing session-based recommendation. In Proceedings of the 43rd International ACM SIGIR Conference on Research and
Development in Information Retrieval. 669–678.
[147] Tahleen Rahman, Bartlomiej Surma, Michael Backes, and Yang Zhang. 2019. Fairwalk: Towards fair graph embedding.
In Proceedings of the 28th International Joint Conference on Artificial Intelligence. 3289–3295.
[148] Steffen Rendle. 2010. Factorization machines. In Proceedings of the International Conference on Data Mining. IEEE,
995–1000.
[149] Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme. 2009. BPR: Bayesian personalized
ranking from implicit feedback. In Proceedings of the 25th Conference on Uncertainty in Artificial Intelligence. 452–461.
[150] Yu Rong, Wenbing Huang, Tingyang Xu, and Junzhou Huang. 2019. DropEdge: Towards deep graph convolutional
networks on node classification. In Proceedings of the International Conference on Learning Representations. 1–18.
[151] Heng-Shiou Sheu and Sheng Li. 2020. Context-aware graph embedding for session-based news recommendation. In
Proceedings of the 14th ACM Conference on Recommender Systems. 657–662.
[152] David I. Shuman, Sunil K. Narang, Pascal Frossard, Antonio Ortega, and Pierre Vandergheynst. 2013. The emerging
field of signal processing on graphs: Extending high-dimensional data analysis to networks and other irregular
domains. IEEE Sig. Process. Mag. 30, 3 (2013), 83–98.
[153] Ashudeep Singh and Thorsten Joachims. 2018. Fairness of exposure in rankings. In Proceedings of the 24th ACM
SIGKDD International Conference on Knowledge Discovery & Data Mining. 2219–2228.
[154] Rashmi Sinha and Kirsten Swearingen. 2002. The role of transparency in recommender systems. In CHI’02 Extended
Abstracts on Human Factors in Computing Systems, Association for Computing Machinery, 830–831.
[155] Changhao Song, Bo Wang, Qinxue Jiang, Yehua Zhang, Ruifang He, and Yuexian Hou. 2021. Social recommenda-
tion with implicit social influence. In Proceedings of the 44th International ACM SIGIR Conference on Research and
Development in Information Retrieval. 1788–1792.
[156] Weiping Song, Chence Shi, Zhiping Xiao, Zhijian Duan, Yewen Xu, Ming Zhang, and Jian Tang. 2019. AutoInt: Au-
tomatic feature interaction learning via self-attentive neural networks. In Proceedings of the 28th ACM International
Conference on Information and Knowledge Management. 1161–1170.
[157] Weiping Song, Zhiping Xiao, Yifan Wang, Laurent Charlin, Ming Zhang, and Jian Tang. 2019. Session-based social
recommendation via dynamic graph attention networks. In Proceedings of the 12th ACM International Conference on
Web Search and Data Mining. 555–563.
[158] Yixin Su, Rui Zhang, Sarah Erfani, and Zhenghua Xu. 2021. Detecting beneficial feature interactions for recommender
systems. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 35. 4357–4365.

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
A Survey of Graph Neural Networks for Recommender Systems 3:47

[159] Jianing Sun, Wei Guo, Dengcheng Zhang, Yingxue Zhang, Florence Regol, Yaochen Hu, Huifeng Guo, Ruiming Tang,
Han Yuan, Xiuqiang He, et al. 2020. A framework for recommending accurate and diverse items using Bayesian
graph convolutional neural networks. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge
Discovery & Data Mining. 2030–2039.
[160] Jianing Sun, Yingxue Zhang, Wei Guo, Huifeng Guo, Ruiming Tang, Xiuqiang He, Chen Ma, and Mark Coates. 2020.
Neighbor interaction aware graph convolution networks for recommendation. In Proceedings of the 43rd International
ACM SIGIR Conference on Research and Development in Information Retrieval. 1289–1298.
[161] Peijie Sun, Le Wu, and Meng Wang. 2018. Attentive recurrent social recommendation. In Proceedings of the 41st
International ACM SIGIR Conference on Research & Development in Information Retrieval. 185–194.
[162] Yueming Sun and Yi Zhang. 2018. Conversational recommender system. In Proceedings of the 41st International ACM
SIGIR Conference on Research & Development in Information Retrieval. 235–244.
[163] Zhu Sun, Jie Yang, Jie Zhang, Alessandro Bozzon, Long-Kai Huang, and Chi Xu. 2018. Recurrent knowledge graph
embedding for effective recommendation. In Proceedings of the 12th ACM Conference on Recommender Systems.
297–305.
[164] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent
Vanhoucke, and Andrew Rabinovich. 2015. Going deeper with convolutions. In Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition. 1–9.
[165] Hongyan Tang, Junning Liu, Ming Zhao, and Xudong Gong. 2020. Progressive layered extraction (PLE): A novel
multi-task learning (MTL) model for personalized recommendations. In Proceedings of the 14th ACM Conference on
Recommender Systems. 269–278.
[166] Nava Tintarev and Judith Masthoff. 2007. A survey of explanations in recommender systems. In Proceedings of the
IEEE 23rd International Conference on Data Engineering Workshop. 801–810.
[167] Aäron Van Den Oord, Sander Dieleman, and Benjamin Schrauwen. 2013. Deep content-based music recommendation.
In Proceedings of the Neural Information Processing Systems Conference, Vol. 26.
[168] Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. 2018. Graph
attention networks. In Proceedings of the International Conference on Learning Representations. 1–12.
[169] Baocheng Wang and Wentao Cai. 2020. Knowledge-enhanced graph neural networks for sequential recommendation.
Information 11, 8 (2020), 388.
[170] Bei Wang, Chenrui Zhang, Hao Zhang, Xiaoqing Lyu, and Zhi Tang. 2020. Dual autoencoder network with swap
reconstruction for cold-start recommendation. In Proceedings of the 29th ACM International Conference on Information
& Knowledge Management. 2249–2252.
[171] Hongwei Wang, Fuzheng Zhang, Jialin Wang, Miao Zhao, Wenjie Li, Xing Xie, and Minyi Guo. 2018. RippleNet:
Propagating user preferences on the knowledge graph for recommender systems. In Proceedings of the 27th ACM
International Conference on Information and Knowledge Management. 417–426.
[172] Hongwei Wang, Miao Zhao, Xing Xie, Wenjie Li, and Minyi Guo. 2019. Knowledge graph convolutional networks
for recommender systems. In Proceedings of the World Wide Web Conference. 3307–3313.
[173] Jianling Wang, Kaize Ding, Ziwei Zhu, and James Caverlee. 2021. Session-based recommendation with hypergraph
attention networks. In Proceedings of the SIAM International Conference on Data Mining. SIAM, 82–90.
[174] Minjie Wang, Da Zheng, Zihao Ye, Quan Gan, Mufei Li, Xiang Song, Jinjing Zhou, Chao Ma, Lingfan Yu, Yu Gai, et al.
2019. Deep graph library: A graph-centric, highly-performant package for graph neural networks. arXiv preprint
arXiv:1909.01315 (2019).
[175] Shoujin Wang, Longbing Cao, Yan Wang, Quan Z. Sheng, Mehmet A. Orgun, and Defu Lian. 2021. A survey on
session-based recommender systems. Comput. Surv. 54, 7 (2021), 1–38.
[176] Shoujin Wang, Liang Hu, Yan Wang, Xiangnan He, Quan Z. Sheng, Mehmet A. Orgun, Longbing Cao, Francesco
Ricci, and Philip S. Yu. 2021. Graph learning based recommender systems: A review. arXiv preprint arXiv:2105.06339
(2021).
[177] Wen Wang, Wei Zhang, Shukai Liu, Qi Liu, Bo Zhang, Leyu Lin, and Hongyuan Zha. 2020. Beyond clicks: Model-
ing multi-relational item graph for session-based target behavior prediction. In Proceedings of the Web Conference.
3056–3062.
[178] Wen Wang, Wei Zhang, Shukai Liu, Qi Liu, Bo Zhang, Leyu Lin, and Hongyuan Zha. 2023. Incorporating link pre-
diction into multi-relational item graph modeling for session-based recommendation. IEEE Trans. Knowl. Data Eng.
35, 3 (2023), 2683–2696. DOI:10.1109/TKDE.2021.3111436
[179] Xin Wang, Hong Chen, Yuwei Zhou, Jianxin Ma, and Wenwu Zhu. 2023. Disentangled representation learning for
recommendation. IEEE Trans. Pattern Anal. Mach. Intell. 45, 1 (2023), 408–424. DOI:10.1109/TPAMI.2022.3153112
[180] Xiang Wang, Xiangnan He, Yixin Cao, Meng Liu, and Tat-Seng Chua. 2019. KGAT: Knowledge graph attention net-
work for recommendation. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery
& Data Mining. 950–958.

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
3:48 C. Gao et al.

[181] Xiang Wang, Xiangnan He, Meng Wang, Fuli Feng, and Tat-Seng Chua. 2019. Neural graph collaborative filtering.
In Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval.
165–174.
[182] Xiang Wang, Tinglin Huang, Dingxian Wang, Yancheng Yuan, Zhenguang Liu, Xiangnan He, and Tat-Seng Chua.
2021. Learning intents behind interactions with knowledge graph for recommendation. In Proceedings of the Web
Conference. 878–887.
[183] Xiang Wang, Hongye Jin, An Zhang, Xiangnan He, Tong Xu, and Tat-Seng Chua. 2020. Disentangled graph col-
laborative filtering. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in
Information Retrieval. 1001–1010.
[184] Xiting Wang, Kunpeng Liu, Dongjie Wang, Le Wu, Yanjie Fu, and Xing Xie. 2022. Multi-level recommendation rea-
soning over knowledge graphs with reinforcement learning. In Proceedings of the ACM Web Conference. 2098–2108.
[185] Xiang Wang, Dingxian Wang, Canran Xu, Xiangnan He, Yixin Cao, and Tat-Seng Chua. 2019. Explainable reasoning
over knowledge graphs for recommendation. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 33.
5329–5336.
[186] Xinxi Wang and Ye Wang. 2014. Improving content-based and hybrid music recommendation using deep learning.
In Proceedings of the 22nd ACM International Conference on Multimedia. 627–636.
[187] Ziyang Wang, Wei Wei, Gao Cong, Xiao-Li Li, Xian-Ling Mao, and Minghui Qiu. 2020. Global context enhanced graph
neural networks for session-based recommendation. In Proceedings of the 43rd International ACM SIGIR Conference
on Research and Development in Information Retrieval. 169–178.
[188] Zhenyi Wang, Huan Zhao, and Chuan Shi. 2022. Profiling the design space for graph neural networks based collabo-
rative filtering. In Proceedings of the 15th ACM International Conference on Web Search and Data Mining. 1109–1119.
[189] Zhaobo Wang, Yanmin Zhu, Qiaomei Zhang, Haobin Liu, Chunyang Wang, and Tong Liu. 2022. Graph-enhanced
spatial-temporal network for next poi recommendation. ACM Trans. Knowl. Discov. Data 16, 6 (2022), 21.
[190] Mark Wilhelm, Ajith Ramanathan, Alexander Bonomo, Sagar Jain, Ed H. Chi, and Jennifer Gillenwater. 2018. Prac-
tical diversified recommendations on YouTube with determinantal point processes. In Proceedings of the 27th ACM
International Conference on Information and Knowledge Management. 2165–2173.
[191] Zhaobo Wang, Yanmin Zhu, Qiaomei Zhang, Haobing Liu, Chunyang Wang, and Tong Liu. 2022. Graph-enhanced
spatial-temporal network for next POI recommendation. ACM Trans. Knowl. Discov. Data 16, 6 (2022), 21.
[192] Fangzhao Wu, Ying Qiao, Jiun-Hung Chen, Chuhan Wu, Tao Qi, Jianxun Lian, Danyang Liu, Xing Xie, Jianfeng Gao,
Winnie Wu, et al. 2020. Mind: A large-scale dataset for news recommendation. In Proceedings of the 58th Annual
Meeting of the Association for Computational Linguistics. 3597–3606.
[193] Jiancan Wu, Xiangnan He, Xiang Wang, Qifan Wang, Weijian Chen, Jianxun Lian, and Xing Xie. 2022. Graph convo-
lution machine for context-aware recommender system. Front. Comput. Sci. 16, 6 (2022), 1–12.
[194] Jiancan Wu, Xiang Wang, Fuli Feng, Xiangnan He, Liang Chen, Jianxun Lian, and Xing Xie. 2021. Self-supervised
graph learning for recommendation. In Proceedings of the 44th International ACM SIGIR Conference on Research and
Development in Information Retrieval. 726–735.
[195] Le Wu, Lei Chen, Pengyang Shao, Richang Hong, Xiting Wang, and Meng Wang. 2021. Learning fair representations
for recommendation: A graph-based perspective. In Proceedings of the Web Conference. 2198–2208.
[196] Le Wu, Xiangnan He, Xiang Wang, Kun Zhang, and Meng Wang. 2022. A survey on accuracy-oriented neural recom-
mendation: From collaborative filtering to information-rich recommendation. IEEE Trans. Knowl. Data Eng. (2022),
1–1.
[197] Le Wu, Junwei Li, Peijie Sun, Richang Hong, Yong Ge, and Meng Wang. 2022. DifNet++: A neural influence
and interest diffusion network for social recommendation. IEEE Trans. Knowl. Data Eng. 34, 10 (2022), 4753–4766.
DOI:10.1109/TKDE.2020.3048414
[198] Le Wu, Peijie Sun, Yanjie Fu, Richang Hong, Xiting Wang, and Meng Wang. 2019. A neural influence diffusion
model for social recommendation. In Proceedings of the 42nd International ACM SIGIR Conference on Research and
Development in Information Retrieval. 235–244.
[199] Le Wu, Peijie Sun, Richang Hong, Yong Ge, and Meng Wang. 2021. Collaborative neural social recommendation. IEEE
Trans. Syst. Man Cyber.: Syst. 51, 1 (2021), 464–476.
[200] Peizhi Wu, Yi Tu, Xiaojie Yuan, Adam Jatowt, and Zhenglu Yang. 2018. Neural framework for joint evolution model-
ing of user feedback and social links in dynamic social networks. In Proceedings of the International Joint Conference
on Artificial Intelligence. 1632–1638.
[201] Qitian Wu, Hengrui Zhang, Xiaofeng Gao, Peng He, Paul Weng, Han Gao, and Guihai Chen. 2019. Dual graph atten-
tion networks for deep latent representation of multifaceted social effects in recommender systems. In Proceedings
of the World Wide Web Conference. 2091–2102.
[202] Shiwen Wu, Fei Sun, Wentao Zhang, and Bin Cui. 2021. Graph neural networks in recommender systems: A survey.
arXiv preprint arXiv:2011.02260 (2021).

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
A Survey of Graph Neural Networks for Recommender Systems 3:49

[203] Shu Wu, Yuyuan Tang, Yanqiao Zhu, Liang Wang, Xing Xie, and Tieniu Tan. 2019. Session-based recommendation
with graph neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 33. 346–353.
[204] Yuxia Wu, Lizi Liao, Gangyi Zhang, Wenqiang Lei, Guoshuai Zhao, Xueming Qian, and Tat-Seng Chua. 2022. State
graph reasoning for multimodal conversational recommendation. IEEE Trans. Multim. (2022), 1–1.
[205] Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and S Yu Philip. 2020. A comprehensive
survey on graph neural networks. IEEE Trans. Neural Netw. Learn. Syst. 32, 1 (2020), 4–24.
[206] Lianghao Xia, Chao Huang, Yong Xu, Peng Dai, Mengyin Lu, and Liefeng Bo. 2021. Multi-behavior enhanced recom-
mendation with cross-interaction collaborative relation modeling. In Proceedings of the 37th International Conference
on Data Engineering. IEEE, 1931–1936.
[207] Lianghao Xia, Chao Huang, Yong Xu, Peng Dai, Xiyue Zhang, Hongsheng Yang, Jian Pei, and Liefeng Bo. 2021.
Knowledge-enhanced hierarchical graph transformer network for multi-behavior recommendation. In Proceedings
of the AAAI Conference on Artificial Intelligence, Vol. 35. 4486–4493.
[208] Lianghao Xia, Yong Xu, Chao Huang, Peng Dai, and Liefeng Bo. 2021. Graph meta network for multi-behavior recom-
mendation. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information
Retrieval. 757–766.
[209] Xin Xia, Hongzhi Yin, Junliang Yu, Yingxia Shao, and Lizhen Cui. 2021. Self-supervised graph co-training for session-
based recommendation. In Proceedings of the 30th ACM International Conference on Information & Knowledge Man-
agement. 2180–2190.
[210] Xin Xia, Hongzhi Yin, Junliang Yu, Qinyong Wang, Lizhen Cui, and Xiangliang Zhang. 2021. Self-supervised hyper-
graph convolutional networks for session-based recommendation. In Proceedings of the AAAI Conference on Artificial
Intelligence, Vol. 35. 4503–4511.
[211] Yikun Xian, Zuohui Fu, S. Muthukrishnan, Gerard de Melo, and Yongfeng Zhang. 2019. Reinforcement knowledge
graph reasoning for explainable recommendation. In Proceedings of the 42nd International ACM SIGIR Conference on
Research and Development in Information Retrieval. 285–294.
[212] Fengtong Xiao, Lin Li, Weinan Xu, Jingyu Zhao, Xiaofeng Yang, Jun Lang, and Hao Wang. 2021. DMBGN: Deep
multi-behavior graph networks for voucher redemption rate prediction. In Proceedings of the 27th ACM SIGKDD
Conference on Knowledge Discovery & Data Mining. 3786–3794.
[213] Min Xie, Hongzhi Yin, Fanjiang Xu, Hao Wang, and Xiaofang Zhou. 2016. Graph-based metric embedding for
next POI recommendation. In Proceedings of the International Conference on Web Information Systems Engineering.
Springer, 207–222.
[214] Ruobing Xie, Qi Liu, Shukai Liu, Ziwei Zhang, Peng Cui, Bo Zhang, and Leyu Lin. 2022. Improving accuracy and
diversity in matching of recommendation with diversified preference network. IEEE Trans. Big Data 8, 4 (2022),
955–967.
[215] Chengfeng Xu, Pengpeng Zhao, Yanchi Liu, Victor S. Sheng, Jiajie Xu, Fuzhen Zhuang, Junhua Fang, and Xiaofang
Zhou. 2019. Graph contextualized self-attention network for session-based recommendation. In Proceedings of the
International Conference on International Joint Conferences on Artificial Intelligence, Vol. 19. 3940–3946.
[216] Fengli Xu, Jianxun Lian, Zhenyu Han, Yong Li, Yujian Xu, and Xing Xie. 2019. Relation-aware graph convolutional
networks for agent-initiated social e-commerce recommendation. In Proceedings of the 28th ACM International Con-
ference on Information and Knowledge Management. 529–538.
[217] Huance Xu, Chao Huang, Yong Xu, Lianghao Xia, Hao Xing, and Dawei Yin. 2020. Global context enhanced social
recommendation with hierarchical graph neural networks. In Proceedings of the IEEE International Conference on
Data Mining. IEEE, 701–710.
[218] Zhen Yang, Ming Ding, Xu Zou, Jie Tang, Bin Xu, Chang Zhou, and Hongxia Yang. 2022. Region or global a principle
for negative sampling in graph-based recommendation. IEEE Trans. Knowl. Data Eng. (2022), 1–1.
[219] Zuoxi Yang and Shoubin Dong. 2020. HAGERec: Hierarchical attention graph convolutional network incorporating
knowledge graph for explainable recommendation. Knowl.-based Syst. 204 (2020), 106194.
[220] Liang Yao, Chengsheng Mao, and Yuan Luo. 2019. Graph convolutional networks for text classification. In Proceedings
of the AAAI Conference on Artificial Intelligence, Vol. 33. 7370–7377.
[221] Quanming Yao, Mengshuo Wang, Yuqiang Chen, Wenyuan Dai, Yu-Feng Li, Wei-Wei Tu, Qiang Yang, and Yang
Yu. 2018. Taking human out of learning applications: A survey on automated machine learning. arXiv preprint
arXiv:1810.13306 (2018).
[222] Rex Ying, Ruining He, Kaifeng Chen, Pong Eksombatchai, William L. Hamilton, and Jure Leskovec. 2018. Graph
convolutional neural networks for web-scale recommender systems. In Proceedings of the 24th ACM SIGKDD Inter-
national Conference on Knowledge Discovery & Data Mining. 974–983.
[223] Bo Yu, Ruoqian Zhang, Wei Chen, and Junhua Fang. 2021. Graph neural network based model for multi-behavior
session-based recommendation. GeoInformatica (2021), 1–19.

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
3:50 C. Gao et al.

[224] Feng Yu, Yanqiao Zhu, Qiang Liu, Shu Wu, Liang Wang, and Tieniu Tan. 2020. TAGNN: Target attentive graph
neural networks for session-based recommendation. In Proceedings of the 43rd International ACM SIGIR Conference
on Research and Development in Information Retrieval. 1921–1924.
[225] Junliang Yu, Hongzhi Yin, Min Gao, Xin Xia, Xiangliang Zhang, and Nguyen Quoc Viet Hung. 2021. Socially-aware
self-supervised tri-training for recommendation. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge
Discovery & Data Mining. 2084–2092.
[226] Junliang Yu, Hongzhi Yin, Jundong Li, Min Gao, Zi Huang, and Lizhen Cui. 2022. Enhance social recommendation
with adversarial graph convolutional networks. IEEE Trans. Knowl. Data Eng. 34, 8 (2023), 3727–3739. DOI:10.1109/
TKDE.2020.3033673
[227] Junliang Yu, Hongzhi Yin, Jundong Li, Qinyong Wang, Nguyen Quoc Viet Hung, and Xiangliang Zhang. 2021. Self-
supervised multi-channel hypergraph convolutional network for social recommendation. In Proceedings of the Web
Conference. 413–424.
[228] Wenhui Yu, Zixin Zhang, and Zheng Qin. 2022. Low-pass graph convolutional network for recommendation. Proc.
AAAI Conf. Artif. Intell. 36, 8 (2022), 8954–8961.
[229] Chuxu Zhang, Dongjin Song, Chao Huang, Ananthram Swami, and Nitesh V. Chawla. 2019. Heterogeneous graph
neural network. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data
Mining. 793–803.
[230] Jun Zhang, Chen Gao, Depeng Jin, and Yong Li. 2021. Group-buying recommendation for social e-commerce. In
Proceedings of the IEEE 37th International Conference on Data Engineering. 1536–1547.
[231] Jiyong Zhang, Xin Liu, Xiaofei Zhou, and Xiaowen Chu. 2021. Leveraging graph neural networks for point-of-interest
recommendations. Neurocomputing 462 (2021), 1–13.
[232] Jiani Zhang, Xingjian Shi, Shenglin Zhao, and Irwin King. 2019. STAR-GCN: Stacked and reconstructed graph con-
volutional networks for recommender systems. In Proceedings of the 28th International Joint Conference on Artificial
Intelligence. 4264–4270.
[233] Mengqi Zhang, Shu Wu, Meng Gao, Xin Jiang, Ke Xu, and Liang Wang. 2022. Personalized graph neural networks
with attention mechanism for session-aware recommendation. IEEE Trans. Knowl. Data Eng. 34, 8 (2022), 3946–3957.
[234] Mengqi Zhang, Shu Wu, Xueli Yu, Qiang Liu, and Liang Wang. 2022. Dynamic graph neural networks for sequential
recommendation. IEEE Trans. Knowl. Data Eng. (2022), 1–1.
[235] Weifeng Zhang, Jingwen Mao, Yi Cao, and Congfu Xu. 2020. Multiplex graph neural networks for multi-behavior
recommendation. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management.
2313–2316.
[236] Yongfeng Zhang and Xu Chen. 2020. Explainable recommendation: A survey and new perspectives. Found. Trendső
Inf. Retr. 14, 1 (2020), 1–101.
[237] Yongfeng Zhang, Guokun Lai, Min Zhang, Yi Zhang, Yiqun Liu, and Shaoping Ma. 2014. Explicit factor models for
explainable recommendation based on phrase-level sentiment analysis. In Proceedings of the 37th International ACM
SIGIR Conference on Research & Development in Information Retrieval. 83–92.
[238] Cheng Zhao, Chenliang Li, and Cong Fu. 2019. Cross-domain recommendation via preference propagation GraphNet.
In Proceedings of the 28th ACM International Conference on Information and Knowledge Management. 2165–2168.
[239] Kai Zhao, Yukun Zheng, Tao Zhuang, Xiang Li, and Xiaoyi Zeng. 2022. Joint learning of e-commerce search and
recommendation with a unified graph neural network. In Proceedings of the 15th ACM International Conference on
Web Search and Data Mining. 1461–1469.
[240] Sen Zhao, Wei Wei, Ding Zou, and Xianling Mao. 2022. Multi-view intent disentangle graph networks for bundle
recommendation. Proc. AAAI Conf. Artif. Intell. 36, 4 (2022), 4379–4387.
[241] Ruiqi Zheng, Liang Qu, Bin Cui, Yuhui Shi, and Hongzhi Yin. 2022. AutoML for deep recommender systems: A survey.
arXiv preprint arXiv:2203.13922 (2022).
[242] Yu Zheng, Chen Gao, Liang Chen, Depeng Jin, and Yong Li. 2021. DGCN: Diversified recommendation with graph
convolutional networks. In Proceedings of the Web Conference. 401–412.
[243] Yu Zheng, Chen Gao, Xiangnan He, Depeng Jin, and Yong Li. 2023. Incorporating price into recommendation
with graph convolutional networks. IEEE Trans. Knowl. Data Eng. 34, 8 (2022), 3727–3739. DOI:10.1109/TKDE.
2020.3033673
[244] Yu Zheng, Chen Gao, Xiangnan He, Yong Li, and Depeng Jin. 2020. Price-aware recommendation with graph convo-
lutional networks. In Proceedings of the 36th International Conference on Data Engineering. IEEE, 133–144.
[245] Yujia Zheng, Siyi Liu, Zekun Li, and Shu Wu. 2020. DGTN: Dual-channel graph transition network for session-based
recommendation. In Proceedings of the International Conference on Data Mining Workshops. IEEE, 236–242.
[246] Zhi Zheng, Chao Wang, Tong Xu, Dazhong Shen, Penggang Qin, Baoxing Huai, Tongzhu Liu, and Enhong Chen.
2021. Drug package recommendation via interaction-aware graph induction. In Proceedings of the Web Conference.
1284–1295.

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.
A Survey of Graph Neural Networks for Recommender Systems 3:51

[247] Zhi Zheng, Chao Wang, Tong Xu, Dazhong Shen, Penggang Qin, Xiangyu Zhao, Baoxing Huai, Xian Wu, and Enhong
Chen. 2023. Interaction-aware drug package recommendation via policy gradient. ACM Trans. Inf. Syst. 41, 1 (2023),
1046–8188.
[248] Guorui Zhou, Xiaoqiang Zhu, Chenru Song, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai.
2018. Deep interest network for click-through rate prediction. In Proceedings of the 24th ACM SIGKDD International
Conference on Knowledge Discovery & Data Mining. 1059–1068.
[249] Huachi Zhou, Qiaoyu Tan, Xiao Huang, Kaixiong Zhou, and Xiaoling Wang. 2021. Temporal augmented graph neural
networks for session-based recommendations. In Proceedings of the 44th International ACM SIGIR Conference on
Research and Development in Information Retrieval. 1798–1802.
[250] Jie Zhou, Ganqu Cui, Shengding Hu, Zhengyan Zhang, Cheng Yang, Zhiyuan Liu, Lifeng Wang, Changcheng Li, and
Maosong Sun. 2020. Graph neural networks: A review of methods and applications. AI Open 1 (2020), 57–81.
[251] Kaixiong Zhou, Xiao Huang, Yuening Li, Daochen Zha, Rui Chen, and Xia Hu. 2020. Towards deeper graph neural
networks with differentiable group normalization. Adv. Neural Inf. Process. Syst. 33 (2020), 1–12.
[252] Rong Zhu, Kun Zhao, Hongxia Yang, Wei Lin, Chang Zhou, Baole Ai, Yong Li, and Jingren Zhou. 2019. AliGraph: A
comprehensive graph neural network platform. Proc. VLDB Endow. 12, 12 (2019), 2094–2105.
[253] Tianyu Zhu, Leilei Sun, and Guoqing Chen. 2023. Graph-based embedding smoothing for sequential recommenda-
tion. IEEE Trans. Knowl. Data Eng. 35, 1 (2023), 496–508. DOI:10.1109/TKDE.2021.3073411
[254] Ziwei Zhu, Xia Hu, and James Caverlee. 2018. Fairness-aware tensor-based recommendation. In Proceedings of the
27th ACM International Conference on Information and Knowledge Management. 1153–1162.
[255] Tao Zhuang, Wenwu Ou, and Zhirong Wang. 2018. Globally optimized mutual influence aware ranking in
e-commerce search. In Proceedings of the 27th International Joint Conference on Artificial Intelligence. 3725–3731.

Received 13 May 2022; revised 21 August 2022; accepted 24 August 2022

ACM Transactions on Recommender Systems, Vol. 1, No. 1, Article 3. Publication date: February 2023.