applied

sciences
Review
A Comprehensive Survey of Recommender Systems Based on
Deep Learning
Hongde Zhou 1 , Fei Xiong 1 and Hongshu Chen 2, *

1 Key Laboratory of Communication and Information Systems, Beijing Municipal Commission of Education,
Beijing Jiaotong University, Beijing 100044, China; [email protected] (H.Z.); [email protected] (F.X.)
2 School of Management and Economics, Beijing Institute of Technology, Beijing 100081, China
* Correspondence: [email protected]

Abstract: With the increasing abundance of information resources and the development of deep
learning techniques, recommender systems (RSs) based on deep learning have gradually become
a research focus. Although RSs have evolved in recent years, a systematic review of existing RS
approaches is still warranted. The main focus of this paper is on recommendation models that
incorporate deep learning techniques. The objective is to guide novice researchers interested in this
field through the investigation and application of the proposed recommendation models. Specifically,
we first categorize existing RS approaches into four types: content-based recommendations, sequence
recommendations, cross-domain recommendations, and social recommendation methods. We then
introduce the definitions and address the challenges associated with these RS methodologies. Sub-
sequently, we propose a comprehensive categorization framework and novel taxonomies for these
methodologies, providing a thorough account of their research advancements. Finally, we discuss
future developments regarding this topic.

Keywords: recommender systems; deep learning; social networks; sequence recommendation;

cross-domain recommendation

1. Introduction
Citation: Zhou, H.; Xiong, F.; Chen, With the rapid progression of technology and the development of the Internet, we
H. A Comprehensive Survey of have transitioned from an era of information scarcity to the age of big data. The growth in
Recommender Systems Based on the amount of information available has led to the challenge of “information overload” [1].
Deep Learning. Appl. Sci. 2023, 13, In the context of the digital age, recommender systems (RSs) have arisen. RSs analyze
11378. https://doi.org/10.3390/ data to discern user preferences for items and assist users in efficiently sifting through
app132011378 information [2], directing them towards the content that is most relevant and valuable
Academic Editor: Kiril Tenekedjiev for their interests. Currently, RSs are widely adopted and have provided many economic
benefits. Essentially, RSs are founded on the premise that when users exhibit similar item
Received: 19 September 2023
ratings or behaviors, they are likely to demonstrate similar ratings or actions on other
Revised: 5 October 2023
items [3].
Accepted: 13 October 2023
Lately, due to the continuous improvement in computational capabilities, artificial
Published: 17 October 2023
neural networks (ANNs) have started to garner widespread attention [4]. This has pro-
pelled deep learning to emerge as a burgeoning field in computer science. An ANN is
composed of layers of nodes. Neural networks acquire preferences by leveraging training
Copyright: © 2023 by the authors. data and improving their accuracy over time. Deep learning builds upon the foundation of
Licensee MDPI, Basel, Switzerland. machine learning, incorporating the concept of neural networks [5]. Using deep learning
This article is an open access article techniques, we can effectively process complex data and uncover deeply hidden features
distributed under the terms and and the relationships between these features, thereby greatly enhancing data representa-
conditions of the Creative Commons tion. In 2016, after years of continuous research in the video recommendation domain,
Attribution (CC BY) license (https:// YouTube integrated deep neural networks into its recommender system and applied them
creativecommons.org/licenses/by/
4.0/).

Appl. Sci. 2023, 13, 11378. https://doi.org/10.3390/app132011378 https://www.mdpi.com/journal/applsci

Appl. Sci. 2023, 13, 11378 2 of 31

to video recommendations with outstanding results [6]. This achievement served as inspi-
ration for an increasing number of recommendation models to incorporate deep learning
techniques [7].
In an RS, ‘item’ denotes a product or service recommended by the system to its users.
Recommendation of products to a user requires the analysis of the past preferences of
similar users and leveraging item information. At the same time, users’ preferences are
dynamically evolving [8], so researchers employ time-based/sequential methods to model
users’ dynamic preferences and sequential patterns that change over time [9].
The integration of recommendation techniques with neural networks has paved the
way for broader application prospects [10]. Users frequently interact with multiple plat-
forms using various services. By utilizing cross-domain recommendations, insights from
different domains can be leveraged to provide more comprehensive and relevant sugges-
tions [11]. This approach effectively tackles data sparsity, mitigates the cold start problem,
and captures a more holistic view of a user’s multifaceted preferences. Some studies [12–14]
have indicated that social relationships between users can be effectively used to alleviate
issues such as the cold start problem. People have proposed integrating information-rich
social networks with neural networks and incorporating the social–relational attributes
of users or items (such as friendships and tag categorizations) into traditional recommen-
dation frameworks, thereby overcoming the limitations of conventional recommendation
algorithms. However, many recommendation algorithms still fall in terms of handling data
sparsity and processing large-scale data, which to some extent compromises the precision
of recommendations produced [15]. RS algorithms must have the ability to deal with highly
sparse data and need to scale with increasing numbers of users and items. At the same
time, they must provide satisfactory results and address other issues, such as data noise
and privacy protection.
In the evolution of recommender systems, various challenges have arisen, such as
information overload, cold start problems, long tail effects, and so on. In recent years, deep
learning technology has emerged as a pivotal force in the field of recommender systems.
Deep learning models, known for their exceptional performance, have become a hot spot
in both academia and industry, offering solutions to many of the core challenges currently
confronting recommender systems.
What makes deep learning compelling is its effectiveness in addressing the numerous
challenges encountered by recommender systems. First of all, deep learning models
excel at learning complex user and project features, making them adept at capturing
subtle differences in user behavior. Second, deep learning models can handle large-scale
data and are therefore able to cope with information overload. In addition, personalized
recommendations that take advantage of potential representations of users and items
can solve the cold-start problem to some extent. In addition, deep learning can improve
the robustness of recommender systems and improve user experience. The reason why
deep learning is so captivating lies in its effectiveness in addressing numerous challenges
encountered by recommender systems. First, deep learning models can learn intricate
features of users and items, allowing for a more precise capture of subtle variations in
user behavior. Second, deep learning models can handle large-scale data, thus mitigating
the issue of information overload. Furthermore, personalized recommendations based on
the latent representations of users and items can, to some extent, alleviate the cold-start
problem. Additionally, deep learning can enhance the robustness of recommender systems,
improving the user experience.
In this study, we provide a comprehensive overview of recommender systems with
a particular focus on their integration with deep learning. We introduce various types of
recommendation methods, technological trends, and application domains that leverage
deep learning. We offer detailed descriptions from four perspectives: content-based recom-
mendation, sequence recommendation, cross-domain recommendation, and social recom-
mendation. Additionally, we delve into the application of deep learning in recommender
systems, including model principles, performance enhancements, and application cases,
Appl. Sci. 2023, 13, 11378 3 of 31

with the aim of providing readers with a deep understanding of this field. Furthermore,
we discuss potential future trends in deep learning-based recommender systems, including
model interpretability, multi-modal recommendation, privacy, and fairness considerations.
The primary contributions of this paper can be summarized as follows.
• We present a comprehensive examination of recommender systems, with a specific
emphasis on their integration with deep learning. We categorize them in terms of their
developmental perspective, providing a comprehensive view of the evolution of the
recommender systems field.
• We conduct a review of the research progress of recommender systems integrated
with deep learning, focusing on methods for applying deep learning to collaborative
filtering. Specifically, we perform a comprehensive analysis of four recommendation
approaches that incorporate deep learning: content-based recommendation, sequence
recommendation, cross-domain recommendation, and social recommendation.
• We identify future research directions in the field of deep learning-based recommender
systems, contributing to the advancement of the research community.
Throughout this review study, in comparison to existing literature, we have pinpointed
articles that excessively concentrate on content-based recommendation and collaborative
filtering, with relatively less attention given to reviews concerning the direction of deep
learning. Consequently, the following gaps have been identified: Despite the exceptional
performance of deep learning models, their interpretability remains a challenge. This paper
aims to explore ways to improve the interpretability of models to meet the requirements of
both users and regulatory authorities. Future research could delve deeper into methods
for enhancing model interpretability to bolster user trust. With the proliferation of multi-
modal data, recommender systems need to better integrate diverse types of information.
We believe that future research can explore methods for the more profound fusion of multi-
modal data. Recommender systems should also make strides in addressing user privacy
concerns and guaranteeing fairness in recommendation results. Future research should
focus on how to protect user privacy and ensure fairness while providing personalized
recommendations. In summary, this review aims to highlight these research gaps and offer
valuable guidance for future research efforts in this field.

2. Related Work
Recommender systems are derived from advances in cognitive science, approximation
theory, information extraction, and prediction theory. They have undergone rapid evolution
since their birth and have become an important catalyst for the rise of the modern business
economy [16]. These systems filter out redundant information from vast amounts of data,
selecting items that can fulfill latent user needs. Traditional recommender systems typically
focus on whether users have shown interest or rated an item [17]. By analyzing histori-
cal interaction data, these systems uncover the underlying demands, thereby achieving
efficient recommendations.
Recommender systems are fundamentally based on exploiting binary relationships
between users and items. By utilizing historical behaviors or similarity relations, these
systems help identify items that might pique users’ interests [18]. A utility function, denoted
as “s”, is employed to calculate the recommendation score for item “i” with respect to
user “u”. In this context, both users and items are characterized by a collection of distinct
attribute features [19]. By calculating recommendation scores, the aim is to find the most
interesting i0 ∈ I for every u ∈ U, as shown in Equation (1).

∀u ∈ U, iu0 = argmaxs(u, i ) (1)

In 2008, Ma et al. from the University of Hong Kong [20] took an innovative step
by integrating user social interaction information with historical item rating data. This
was a pioneering effort that introduced the SoRec recommendation algorithm, which was
built on the foundation of probabilistic matrix factorization (PMF) [21]. Recognizing the
Appl. Sci. 2023, 13, 11378 4 of 31

increasing significance of social interactions in recommendations, research in this area

witnessed exponential growth. In 2010, Jamali [22] delved deeper into the sphere of
trust propagation among users and their social friends. They adeptly combined matrix
factorization techniques with interactive trust, leading to the conception of the Social
MF model. A year later, in 2011, building on the foundation laid by Social MF, Ma et al.
utilized social regularization for recommendation [23]. This model was predicated on the
assumption that the greater the influence of neighboring friends on a user, the more the
user’s latent preferences would converge towards that friend. When it comes to neural
networks, Sedhain et al. of Australia’s NICTA research center ingeniously applied an
autoencoder model to recommender systems in 2015, giving birth to the single-hidden-layer
neural network recommendation model known as AutoRec [24]. Using these operations,
AutoRec has generalization and expression capabilities and it is the basic application of
deep learning in this field. Pushing the boundaries further, Devooght and Bersini [25]
achieved enhanced short-term prediction accuracy in 2017 by redefining collaborative
filtering (CF) as a sequence prediction problem using neural networks. In summary, the
evolution of recommender systems, as seen over the years, has seamlessly transitioned
from leveraging social interactions to harnessing the prowess of deep neural networks.
This evolution trend underscores the dynamic nature of research in this domain and the
search for more refined and accurate recommendation techniques.
Graph neural networks (GNNs) form a collective term for models that apply neural
networks on graphs, which have powerful graph-structured data learning capabilities.
Graphs consist of nodes and edges connecting nodes and are typical non-Euclidean spatial
data structures. Social networks are quintessential examples of graph data. Using GNNs,
social network graph models can be seamlessly integrated into recommender systems.
Hence, research on social recommendation has progressively become a focal point. By
harnessing deep learning techniques to incorporate information from social networks, one
can deeply understand the relevant features of users and items, achieving better accuracy
of the recommendation while simultaneously improving user experience [26,27]. Based on
this concept, GraphSAGE [28] is an inductive learning framework that efficiently generates
embeddings for unknown vertices using their attribute information. This alleviates the
issue of models being unable to directly generalize to vertices that have not appeared
during the training process. Based on GraphSAGE, Ying et al. introduced an efficient GCN
algorithm [29], they combine random walks with graph convolutions to produce node
embeddings. GraphSAGE introduced a novel paradigm: inductive learning. Its strength
lies in its ability to generalize from specific instances to broader contexts, discerning
unknown data on unknown nodes. This advantage enables it to handle various scenarios
in the industrial sector, where graph structures undergo dynamic changes.
The primary objective of this article is to provide scholars interested in this topic
with an understanding of the main effects of employing deep learning methods in recom-
mender systems. This work particularly delves into the motivations and developments
associated with the application of various methods in such systems. Furthermore, it
aims to offer solution-oriented perspectives that address the current challenges faced by
recommender systems.

3. Overview of the Recommender Systems

Early recommender systems were primarily based on collaborative filtering and
content-based recommendation. User-based collaborative filtering methods were intro-
duced in 1992, while item-based collaborative filtering methods emerged in 1999. Content-
based recommendation methods also began to appear during this period, utilizing item
attributes and user preferences for recommendations. In the 21st century, with the rise of
social networks and the development of deep learning technology, social recommendation
methods gradually gained attention. Similarly, cross-domain recommendation methods
also started to emerge, allowing the application of a user’s historical behavior in one
domain to recommendations in other domains. In recent years, with the proliferation of
 Appl. Sci. 2023, 13, x FOR PEER REVIEW 5 of 32

Appl. Sci. 2023, 13, 11378 5 of 31

domain to recommendations in other domains. In recent years, with the proliferation of

mobile devices and online shopping, sequence recommendation methods have become
increasingly
mobile important.
devices and onlineThese methods
shopping, considerrecommendation
sequence user behavior sequences,
methods such
haveas pur-
become
chase historyimportant.
increasingly or browsing history,
These to provide
methods moreuser
consider personalized
behaviorrecommendations.
sequences, such as Deep
pur-
learning
chase methods
history can handle
or browsing large-scale
history, to providedatamore
and personalized
multi-modal recommendations.
information, enhancingDeep
the effectiveness
learning methods canof recommendations.
handle large-scale Reinforcement learninginformation,
data and multi-modal methods have also been
enhancing the
applied to recommender
effectiveness systems,Reinforcement
of recommendations. optimizing recommendation
learning methodsstrategies viabeen
have also userapplied
inter-
toactions. The development
recommender history ofrecommendation
systems, optimizing recommender systems can via
strategies be user
divided into several
interactions. The
key stages, each producing different types of recommendation methods. A
development history of recommender systems can be divided into several key stages, each summary of
some important recommender system methods is provided in the following
producing different types of recommendation methods. A summary of some important Figure 1.
recommender system methods is provided in the following Figure 1.

Figure 1. The development and classification of recommender system methods [28,30–46].

Figure 1. The development and classification of recommender system methods [28,30–46].
New technologies, compared to old ones, are not a complete revolution. New technolo-
gies are Newbuilt on the foundation
technologies, compared laidtobyoldtheir
ones, predecessors,
are not a completeincorporating
revolution. improvements
New tech-
rather than being entirely different. The field of recommender
nologies are built on the foundation laid by their predecessors, incorporating systems is continuously
improve-
evolving,
ments rather withthannewbeingmethods anddifferent.
entirely technologies emerging
The field to adapt to
of recommender evolving
systems user de-
is continu-
mands and data environments. Currently, efforts are underway
ously evolving, with new methods and technologies emerging to adapt to evolving user to refine the integration
ofdemands
deep learning
and data forenvironments.
improved recommendation
Currently, efforts results. Therefore,
are underway the journey
to refine of recom-
the integration
mender systems is a dynamic and evolving process. We will
of deep learning for improved recommendation results. Therefore, the journey of recom- begin by introducing the
most fundamental approaches, combining deep learning with
mender systems is a dynamic and evolving process. We will begin by introducing the most collaborative filtering and
content-based recommendation.
fundamental approaches, combining deep learning with collaborative filtering and con-
Collaborative
tent-based filtering is one of the first and fundamental recommender systems.
recommendation.
In 1992, GlodBerg et
Collaborative al. [47]isintroduced
filtering one of the firstthe and
concept of collaborative
fundamental recommender filtering. The key
systems. In
idea
1992, GlodBerg et al. [47] introduced the concept of collaborative filtering. The key items,
lies in finding the nearest neighbors using the interaction matrix of users and idea
and
liesthen generating
in finding the nearestvarious forms ofusing
neighbors recommendations
the interaction basedmatrixon the information
of users and items, and from
these
then generating various forms of recommendations based on the information from theseof
neighbors, which is essentially a process of matrix completion. The core concept
collaborative
neighbors, which filtering algorithms
is essentially is to provide
a process of matrixproduct recommendations
completion. The core concept or predictions
of collab-
based on similar users or similar items, encompassing two dimensions:
orative filtering algorithms is to provide product recommendations or predictions based neighborhood-
based and model-based
on similar users or similar approaches. The key to collaborative
items, encompassing filteringneighborhood-based
two dimensions: recommendations is
the
andcomputation
model-based ofapproaches.
similarity between
The keydifferent users and
to collaborative different
filtering items.
recommendations is the
Over the past
computation three decades,
of similarity betweencollaborative
different users filtering has been
and different widely used by people.
items.
However, Overasthe time
pasthas progressed,
three decades,its scalability limitations
collaborative filtering hashave beenbecome
widelyapparent, making
used by people.
itHowever,
less suitable for scenarios with large datasets. People have turned
as time has progressed, its scalability limitations have become apparent, mak- to machine learning
and
ingdata
it lessmining
suitable algorithms
for scenarios to learn and recognize
with large complex
datasets. People patterns
have turnedusing training
to machine data.
learn-
Currently, one of the most common paradigms for building
ing and data mining algorithms to learn and recognize complex patterns using training deep learning recommen-
dation algorithmsone
data. Currently, is the
of theusemost
of multi-layer
common paradigmsperceptions, for which
building can introduce
deep learningnonlinear
recom-
transformations
mendation algorithms into existing
is the collaborative
use of multi-layer filtering methods.which can introduce nonlin-
perceptions,
ear Since then, the research
transformations of recommender
into existing collaborativesystems
filtering has shifted towards deep learning-
methods.
based recommendation [48]. To address the complexities of user behavior, neural network-
based models have become increasingly favored by researchers [49]. Among them, neural
collaborative filtering (NCF) stands out as a notable attempt [30]; this method enhances the
expressiveness of the developed model by expanding the inner product within a multilayer
Appl. Sci. 2023, 13, 11378 6 of 31

perceptron (MLP). The success of this approach has further motivated researchers to explore
the integration of deep learning techniques using traditional recommendation algorithms.
This trend is effectively illustrated in the deep factorization machine (DeepFM) [31],
which cleverly integrates a shallow factorization machine (FM) [50] with an MLP. This
approach effectively combines the FM and deep components to achieve a unified represen-
tation of both low-order and high-order feature interactions, significantly enhancing the
model’s expressive power. This approach is similar to traditional machine learning, but its
training is divided into two phases. Therefore, it can converge better and faster.
Wu, Y et al. proposed FedDeepFM [32], a model grounded in federated learning,
which offers a mechanism that delivers high-quality recommendations while enhancing
user privacy. This approach is particularly suitable for scenarios that prioritize privacy
preservation or data transmission reduction. It generates training data using real data
combined with interaction-based synthetic data. Built upon DeepFM, FedDeepFM offers a
mechanism for delivering high-quality recommendations while enhancing user privacy.
The “multi-layer neural network + output layer” allows for a more extensive inter-
action of user/item vectors, introducing additional non-linear features and enhancing
the learning capability for sparse features. By incorporating attention mechanisms and
introducing attention layers between the embedding layer and the multi-layer perceptron,
it has evolved into deep interest networks (DIN) [51]. Furthermore, there are variations
such as deep interest evolution networks (DIEN) [52], which incorporate sequence models
to simulate changes in user preferences, and multi-interest networks with dynamic routing
(MIND) [53], which utilize capsule networks to extract diverse user interests and introduce
tag-based attention mechanisms for dynamic path selection.
As one of the most widely used recommendation algorithms at present, collaborative
filtering exhibits strong generality and significant effectiveness. However, collaborative
filtering always faces serious issues of data sparsity and cold start. The integration of deep
learning with collaborative filtering recommender systems not only simplifies engineering
implementation but also alleviates these problems. Content-based recommendation, which
coexisted with collaborative filtering, has also evolved in a similar manner.

3.1. Content-Based Recommendation

Content-based recommendation is the earliest and most basic type of recommendation
algorithm, and it has played a pivotal role in the entire history of recommender systems [54].
Although its effect may not be as good as that of new recommendation algorithms, it is still
very valuable and even essential.
Content-based recommendation algorithms form a relatively intuitive category of
algorithms that primarily offer recommendations by comparing item features with users’
historical preferences. In the early days, it relied primarily on specific content, such as the
characteristics of text, images, or audio. Their approach entails suggesting items to users
that exhibit content similarity to previously preferred items, determined by the historical
preferences of these users [55,56]. A content-based recommendation model is constructed
based on the target item, user information, and interactions between the user and items.
Importantly, it notably neglects the actions of other users. A schematic process is shown in
Figure 2 below.
To achieve content-based recommendation, the general steps include (1) constructing
an item feature representation based on the content features of each item; (2) building a
user feature representation based on user characteristics and behaviors; and (3) generating
a recommendation list according to the level of alignment between the features of the
target item and those of the user [57]. When integrated with neural networks, the entire
recommendation model can be divided into an input layer, a hidden layer, an output layer,
and a prediction layer. A content-based recommendation model is shown in Figure 3.
Appl.
Appl. Sci.
Sci.2023,
2023,13,
13,x11378
FOR PEER REVIEW 7 7ofof32
31

Figure 2. Diagram of a content-based recommendation algorithm.

To achieve content-based recommendation, the general steps include (1) constructing

an item feature representation based on the content features of each item; (2) building a
user feature representation based on user characteristics and behaviors; and (3) generating
a recommendation list according to the level of alignment between the features of the tar-
get item and those of the user [57]. When integrated with neural networks, the entire rec-
ommendation model can be divided into an input layer, a hidden layer, an output layer,
and a prediction layer. A content-based recommendation model is shown in Figure 3.
Figure2.
Figure Diagramof
2.Diagram ofaacontent-based
content-basedrecommendation
recommendation algorithm.
algorithm.

To achieve content-based recommendation, the general steps include (1) constructing

an item feature representation based on the content features of each item; (2) building a
user feature representation based on user characteristics and behaviors; and (3) generating
a recommendation list according to the level of alignment between the features of the tar-
get item and those of the user [57]. When integrated with neural networks, the entire rec-
ommendation model can be divided into an input layer, a hidden layer, an output layer,
and a prediction layer. A content-based recommendation model is shown in Figure 3.

Figure 3. Content-based recommendation procedure.

Figure 3. Content-based recommendation procedure.
The hidden layer is a fully connected layer consisting of 1–2 sublayers, and the score
ofThe
the hidden layer is[58].
item is output a fully
Thisconnected layer consisting
model highlights of 1–2 sublayers,
the embedding and the score
and transformation of item
of the item is
features. Theoutput [58].
transfer This model
process highlights
from the the to
input layer embedding
the next isand transformation
represented of in
as shown
item features.(2).
Equation The transfer process from the input layer to the next is represented as shown
in Equation (2). h = f ( W ih x + b h ) (2)
The transfer process between hidden layers is depicted as shown in Equation (3)
(if multiple hidden layers are present).
Figure 3. Content-based recommendation
h =procedure.
f (W h + bj ) (3)
j hj j −1

The
Thehidden
transferlayer is a to
process fully
theconnected layer consisting
output is shown in Equation of 1–2
(4). sublayers, and the score
of the item is output [58]. This model highlights the embedding and transformation of
y =the
item features. The transfer process from f (W ho h +
input bo ) to the next is represented as shown
layer (4)
in Equation (2).
where x represents the input and y signifies the output. w and b are parameters, while f is
a nonlinear activation function.
Appl. Sci. 2023, 13, 11378 8 of 31

The process of representing items and users via tagging is becoming increasingly
popular. Tags can serve as content information and be used in content-based recommenda-
tion methods [59], or they can be treated as a third entity, utilizing collaborative filtering
approaches for recommendation. However, as data representations become more diverse,
the preprocessing tasks for items become increasingly challenging. A one-dimensional
vector representation can be overly simplistic, thereby neglecting the intrinsic relation-
ships between various attributes of the given content. A tensor is a multidimensional
data structure that is capable of representing the co-occurrence relationships of multiple
types of attributes. Tensor decomposition is fundamentally a feature extraction method,
and many algorithms that leverage tensor decomposition for model construction have
been developed in the past. Peng et al. [60] constructed a tensor for user–item–tag triples
and employed Tucker decomposition to obtain abstract representations of users. The core
tensor obtained from the decomposition procedure encapsulates the associations between
different attributes, thereby taking the cooccurrence relationships between items and tags
into consideration.
Unlike traditional methods such as prod2vec, which typically only considers user
interactions to create embeddings, Meta-prod2vec not only utilizes tags but also incorpo-
rates additional information, such as product descriptions and reviews, to generate richer
product embeddings [61]. It utilizes product-side information to produce product embed-
dings. By integrating meta-information with interaction data, new embedding vectors are
generated for products. The future of content-based recommender systems is inextricably
linked with advancements in deep learning. Furthermore, feedback loops derived from
users can enhance the adaptability of the utilizing model, ensuring continuous user profile
learning and updating.
When dealing with large and complex user-item interaction matrices, traditional ma-
trix factorization models tend to underperform [62]. The paper titled “Deep learning” [63],
published in Nature, serves as a fundamental contribution to the field of deep learning. It
presents the concept of deep learning by demonstrating how computational models with
multiple processing layers can effectively represent data with varying levels of abstraction.
These methods achieve significantly improved performance in various tasks. The paper
delineates the fundamental equations of neural networks, backpropagation, CNNs, and
RNNs, offering readers a comprehensive overview of the domain’s evolution, key concepts,
and challenges.
Currently, the content-based recommendation is extensively employed in industrial-
scale recommender systems, as it possesses the following advantages.
1. Personalized recommendations: These recommendations are based on the user’s histori-
cal interests, ensuring that the recommended content aligns with the user’s preferences.
2. Simple principle with strong interpretability: Content-based recommendations can
be made based on label dimensions or by embedding items into a vector space using
similarity, making this strategy easy to implement. It is also readily accepted and
validated by users.
3. Addresses the cold-start problem to some extent: As long as sufficient content at-
tributes are available, new items can be effectively handled without relying on other
users’ behaviors.
However, some of the drawbacks of content-based recommendation led to limitations
in its effectiveness and scope of application. It has a narrow recommendation scope, and its
novelty is not pronounced. The results obtained from content-based recommender systems
tend to converge on categories of items that the user has previously shown interest in. New
users, without sufficient interaction histories, might not receive effective recommendations.
Moreover, the comprehensiveness, integrity, and accuracy of content understanding can
impact the efficacy of the recommendation process. Additionally, based on practical
experience, the recommendation accuracy of this approach is not particularly high.
Overall, content-based recommendation is a fundamental technique within recom-
mender systems. It suggests items by analyzing item feature vectors and user interest
 Appl. Sci. 2023, 13, 11378 9 of 31

vectors. The aim is to propose new items to users that match their preferences. But people’s
preferences change over time. So, sequence recommendations came into being.

3.2. Sequential Recommendation

Traditional recommender systems focus on identifying static connections between
users and items, often neglecting dynamic shifts in user preferences over time. Sequential
recommendation approaches posit that the recommendations given to a user at a specific
moment should be determined based on the user’s prior behaviors prior to that moment.
Sequential recommendation captures the dynamic nature of user behaviors, revealing cer-
tain relationships between items over a given time span. Sequential recommender systems
can capture dynamic changes in user interests by modeling their interaction sequences.
The issue of sequential recommendation entails time-based learning to anticipate user
interactions with items. The aim of sequential recommendation is to explicitly model users’
Appl. Sci. 2023, 13, x FOR PEER REVIEW 10 of 32
sequential behaviors. We delve into the intricacies of sequence-based recommendation
models. To offer a more intuitive understanding of this concept, Figure 4 presents the
architectural diagram of a typical sequence-based recommendation model.

Figure 4. Sequential recommendation procedure.

Figure 4. Sequential recommendation procedure.
As illustrated in Figure 4, sequence-based recommendation models are capable of
As illustrated in Figure 4, sequence-based recommendation models are capable of
capturing the dynamic changes in user behaviors, thereby generating more personalized
capturing the dynamic changes in user behaviors, thereby generating more personalized
and timely recommendations. In its input layer, such a model receives a sequence of user
and timely recommendations.
behaviors, In its inputthe
which commonly represent layer, such
user’s a modelinteractions.
historical receives a sequence of user
These sequential
behaviors, which commonly represent the user s historical interactions. These
data are subsequently encoded into high-dimensional vectors. Advanced feature extraction sequential
data are performed
is then subsequently viaencoded
attentioninto high-dimensional
mechanisms or neural vectors. Advanced
networks. feature
Following thisextrac-
step, in
tion is then performed via attention mechanisms or neural networks. Following
the hidden layer, a range of more complex network architectures, such as long short-term this step,
inmemory
the hidden layer, a range of more complex network architectures, such
(LSTM) or transformer models, are employed to capture the temporal depen- as long short-
term memory
dencies (LSTM)
and other or transformer
latent models,
patterns within are employed
the sequence. to capture
Finally, the temporal
the output de-
layer typically
pendencies and other latent patterns within the sequence. Finally, the output
comprises a fully connected layer and an activation function, employed to compute scores layer typi-
cally comprises afor
or probabilities fully connected
each potentiallayer and an activation
recommended item. function, employed to compute
scores Inor summary,
probabilities
thefor each potential
development of recommended
deep learning item.
in recommender systems has wit-
In summary,
nessed the development
an orchestrated of deep learning
blend of innovations, in recommender
each building systems has These
upon its predecessor. wit-
nessed an orchestrated blend of innovations, each building upon its
advancements, rooted in complex mechanisms and models, have progressively enhanced predecessor. These
advancements,
the precision androoted in complex
context mechanisms
awareness and models,Sequence
of recommendations. have progressively
recommender enhanced
systems
the precision and context awareness of recommendations. Sequence recommender sys-
tems (SRS) can be categorized into three groups: standard sequence recommendation,
long-short term sequence recommendation, and multi-objective sequence recommender
systems.
Appl. Sci. 2023, 13, 11378 10 of 31

(SRS) can be categorized into three groups: standard sequence recommendation, long-short
term sequence recommendation, and multi-objective sequence recommender systems.

3.2.1. Standard Sequence Recommendation

Standard sequence recommendation refers to recommendation algorithms that extract
user representations using a single behavior sequence [64]. Traditional sequence recom-
mendation algorithms incorporate various prevalent sequence modelling techniques, such
as pooling, RNNs, CNNs, memory networks, and attention mechanisms. This approach
has been employed in Google’s recommendation model. Due to its simplicity and efficacy,
it stands as one of the most prevalent techniques for leveraging sequence features. Neural
networks have also found extensive applications in sequence recommendation tasks.
In the initial stages of sequence-based recommendation models, the primary methods
employed are Markov chains [33] and feature-based matrix factorization [65,66]. Compared
to deep learning techniques, Markov chains have limitations in terms of handling complex
data. Researchers employ personalized Markov chains to model individual users, leading
to enhanced efficiency. Some people have pointed out that the Markov chain approach un-
derperforms when modelling joint-level sequence patterns and fails to capture the effects of
skip behaviors between users and items on recommendation outcomes. While factorization-
based methods might model sequences by summing up sequence item vectors, they neglect
the influence of order within the user interaction sequences on the recommendation results.
Traditional sequence recommendation methods primarily utilize Markov chains and their
derivatives for predicting a user’s next action. They operate under the assumption that a
user’s subsequent action is solely dependent on their recent few actions.
Compared to traditional recommendation models, deep learning models have demon-
strated superior capabilities with respect to capturing the evolution trends of user sequences.
The evolution of deep learning in sequential recommendation can be delineated via a series
of transformative advancements. Recurrent neural networks (RNNs) have demonstrated
their potent capabilities [67] and have been extensively utilized in sequence-based rec-
ommendation. Sequence recommendation models based on RNNs can be represented as
Equation (5).
ht = σ ( Whh ht−1 + Whx xt + bh ) (5)
where xt represents the input and ht denotes the hidden state. Finally, a different activation
function for output is selected. Some people introduced CNNs to sequence recommenda-
tion tasks [68]. They underlined the fact that the existing Markov chain models can only
model point-level sequential patterns and fall short when representing set-level patterns.
They employ convolutional layers to extract local features from the user’s historical behav-
ior sequence, capturing these features from different embedding perspectives (vertical and
horizontal). Different convolutional kernels are utilized to extract features from the input
sequence. These features are subsequently used to generate the final recommendation list.
An attention mechanism offers a notable solution to this gap.
Based on this foundation, SASRec [34] introduced a sequence recommendation method
founded on self-attention to improve its effect. Their approach employs a multihead self-
attention mechanism and positional encoding to handle the input user behavior sequence.
However, for an extensive sequence, earlier interactions might be overlooked. Conse-
quently, researchers have begun to investigate recommendation models that integrate both
long- and short-term interactions.

3.2.2. Long- and Short-Term Sequence Recommendation

The evolution of long- and short-term sequence recommendation reflects the progres-
sion of recommender systems from simple collaborative filtering methods to sophisticated
deep learning models. With continual advancements in technology, models are becom-
ing increasingly adept at distinguishing and integrating two user preferences, thereby
facilitating more accurate and personalized recommendations.
Appl. Sci. 2023, 13, 11378 11 of 31

Long-term preferences reflect their relatively stable interests, while their recent pref-
erences indicate their recent interest changes [69]. Although user preferences change
dynamically over time [70], some long-term stable preferences that influence user behav-
iors remain. The existing approaches tend to combine two preferences [71].
Taking into account that interactions between two behaviors may exert diverse effects
on the user’s present interests [72]. Compared to recent interaction sequences, long-term
interactions tend to evolve more slowly [73], thereby reducing the demand for real-time
responsiveness [74]. SHAN [75] categorizes user behaviors into two behaviors, employing
a multilayer attention network for modelling purposes. By coupling users’ two preferences,
an optimal user representation is generated, enhancing the output recommendation results.
This hybrid representation is calculated as shown in Equation (6).

∑
hybrid long
ut = β0 ut−1 + βj vj (6)
j∈ Stu

where vj represents the rating embedding of an item, which is indicative of the short-term
embedding, while ut−1 denotes the long-term embedding. β denotes varying attention
scores. This formulation captures the dynamic nature and differentiates between the
contributions of items towards predicting the next.
RNNs encounter challenges like gradient vanishing and gradient explosion when deal-
ing with lengthy sequences, diminishing their proficiency in handling extended sequences.
To tackle this problem, long short-term memory networks (LSTMs) [76] were initially
introduced, introducing the fundamental concept of gating mechanisms that effectively
addressed the issue of vanishing gradients.
Recommender systems often use LSTM or gate recurrent units (GRUs) for user mod-
elling. LSTM is an enhanced version of an RNN [35]. It has been demonstrated that LSTM
outperforms conventional RNNs in sequence recommendation tasks. A GRU, another
variant of LSTM, is computationally simpler yet equally effective. Recognizing that the two-
term impacts of preferences might differ, attention mechanisms have also been employed.
The attention mechanisms used in deep learning stem from the idea that humans are drawn
to the significant parts of a target. Their origin can be traced back to the research performed
by Bahdanau et al. [77]. They used attention mechanisms to model the importance of the
output using different parts of the input sentence. Building upon this foundation, vanilla
attention was introduced to function as an RNN decoder, and it has been extensively used
in sequence recommendation [78]. However, the self-attention mechanism (originating
from the transformer in Google’s 2017 neural machine translation work [79]) has also been
deployed in sequence recommendation. Contrary to vanilla attention, self-attention does
not involve an RNN structure, but it has achieved better performance in recommender
systems than RNN-based models [80].
Building upon this idea, a GRU [81] offered streamlined versions of the LSTM gat-
ing processes. Venturing into the realm of natural language processing, the transformer
model [82] gained widespread acclaim. It established unparalleled standards in terms of
discerning the dependencies between sequences, which paved the way for swift parallel
computations and accelerated sequence information extraction [83].
This foundation was ingeniously adopted in Alibaba’s DIN model. The self-attention
mechanism of this model was integrated within the recommendation domain, amalgamat-
ing both fundamental user details and context information. SDM [36] takes a nuanced
approach by segregating user interests. It employs LSTM, complemented by multi-head
attention mechanisms, to home in on immediate user inclinations. The culmination of this
model witnessed the amalgamation of both fleeting and lasting user interests, encapsulating
the user’s essence in a vector representation. Compared to previous models, this approach
employs gating mechanisms to act as weights, ultimately outputting a user behavior vector,
as shown in Equations (7) and (8).

Gut = sigmoid (W 1 eu + W 2 sut + W 3 pu + b) (7)

Appl. Sci. 2023, 13, 11378 12 of 31

otu = (1 − Gtu ) pu + Gtu sit (8)

Here, eu is the user embedding, while su and qu are data derived from user embed-
dings processed via different attention networks. This facilitates a more effective data
processing strategy. Unlike prior works that simply used parameters as weights, this
current approach enables short-term behaviors to be better integrated.
Finally, the Trans2D model [84] showcased a pioneering expandable attention mecha-
nism. This module adeptly discerns user predilections for specific item attributes, which
are gleaned from their behavioral sequence data, incrementally elevating the efficacy
of the recommendation process. Compared to recommendation algorithms based on
RNNs, those relying on convolutional neural networks (CNNs) present greater challenges.
Tuan et al. [85] devised a 3DCNN model for high-dimensional modelling, demonstrating
its efficacy when the given user sequences are associated with intricate element features.
Yan et al. [86] encoded interaction sequences into three-dimensional vectors, employing 2D
convolutional filters to learn local characteristics.
Tang and Wang et al. [87] proposed the Caser model, building upon the foundations of
prior convolutional techniques. They employed both horizontal and vertical convolutional
filters to capture continuous patterns in pointwise, conjoint, and skip behaviors, achieving
promising results. The vertical convolution operation is shown in Equations (9) and (10).

cik = Φc ( Ei:i+h−1 Fk ) (9)

L
ceik = ∑ Felk · El (10)
l =1

where E serves as a matrix formed by stacking embeddings to facilitate the convolution

operations. Fk slides over E from top to bottom during the horizontal convolution pro-
cess, capturing the user’s behavioral patterns. Behavioral patterns are how users interact
with different items at different points in time. The vertical convolution operates in the
embedding space, focusing on how to group or categorize different items based on their
embedding representations. By combining these two types of convolutions, the model is
capable of capturing user behavior patterns in two distinct dimensions. It also adapts the
embedding representations according to the dynamics of the user behaviors. However, this
approach separately learns embeddings for users and items, failing to capture inherent
interaction information and overlooking contextual details.
GRU4Rec [37] integrates an RNN into session-based recommender systems, treating
the interactions within a session as the historical context for sequence modelling. The
highlight of this work is its rational optimization of sequence recommendation based
on an RNN. With continued research advancements, a multitude of improved models
have emerged.
These models aim to delve deeper into item information by optimizing their model
architectures and training methodologies, thereby enhancing the produced recommenda-
tion outcomes. One such enhanced RNN model evolved from GRU4Rec employs data
augmentation techniques to bolster the stability of the model training process [88]. An-
other modified RNN model adjusts for batch generation by considering dwell time to
more precisely capture user behaviors, discovering the potential impact of dwell time on
the results [89]. However, a significant challenge is that when all candidate items rank
above popular items based on a popularity-driven sampling method, the learning speed of
the developed model is constrained, particularly when recommending long-tailed items.
Addressing this issue, an RNN model with top-k gains [90] introduced a novel sampling
strategy for GRU4Rec, blending uniform sampling with popularity-based sampling to
markedly improve performance. To enhance the modeling of item information, not only
was the item’s ID considered but other features such as text descriptions and images
were also incorporated. This was effectively realized using a parallel RNN (p-RNN) ar-
Appl. Sci. 2023, 13, 11378 13 of 31

chitecture for efficient recommendation [91]. In terms of training strategies, the authors
innovatively introduced methods such as alternating training, residual training, and inter-
laced training to optimize the performance of the p-RNN. Furthermore, Dietmar explored
multiple approaches to combine a session-based KNN classifier with GRU4Rec, such as
switching, cascading, and weighted blending, further enhancing the accuracy of the output
recommendations [92].
The RUM model [93] incorporates a user memory module to preserve such interaction
information. Utilizing a dual-layer memory mechanism, the model employs an external
memory mechanism to simulate the user’s memory process, thereby capturing both the
long-term and short-term preferences of the user. Recurrent structures are employed
to capture the user’s recent activities, whereas an external memory unit stores the user’s
historical behaviors. Attention mechanisms are applied to select the most relevant historical
information. A multilayer network architecture is finally used to capture deep interaction
information. By dynamically updating the memory unit, the model reflects user preference
changes, as indicated in the following Equation (11).

addi = tan h ( A T q i + ba ), muk ← muk + zi k · addi (11)

The above equation is utilized to update the user’s preference memory. A and ba act
as trainable additive parameters with the aim of erasing existing information prior to the
incorporation of new data. Transformers have achieved significant breakthroughs in NLP
tasks, with substantial pretrained models such as BERT leading the way. Bert4Rec [94]
adapts this architecture for recommender systems.
To achieve optimal recall results, it is essential for the employed system to account
for both the long- and short-term interests of the target user. In practical applications,
recommendation models can discern these two preferences from the user’s historical
sequence and ultimately merge them using gated unit modules.

3.2.3. Multi-Objective Sequence Recommendation

Researchers have recognized that in addition to relevance, diversity and novelty
are also important objectives in recommender systems. This implies that these systems
should not only recommend items that users are likely to enjoy but also those they have
not previously encountered. In the realm of multiobjective sequence recommendation,
understanding and modelling user behaviors is one of the core tasks. Traditional methods
often overlook the personalized needs of users when addressing different learning objec-
tives [95]. As time has evolved, researchers have started to realize that user interests are
not monolithic and often span multiple domains, prompting the emergence of a series of
methods. These methods attempt to capture this diversity by encoding the target user’s
sequential behavior into different interest representation vectors [96]. In simple terms, the
input is subjected to grouped convolution operations, which are determined by coupling
coefficients obtained via an iterative dynamic routing process and subsequent weighted
embedding vectors. Through such convolutional operations, each group of outputs can be
regarded as an embedding vector corresponding to a specific interest.
Complexity increases when users exhibit a diverse array of behavioral sequences
in real-world applications, such as clicks, shares, and purchases. Researchers need to
model multiple types of behavioral sequences to uncover users’ genuine preferences more
comprehensively [97]. This issue also underscores the critical importance of explicitly
considering users’ historical interactions to enhance performance [98]. Further studies
have revealed that the actions within a user’s behavioral sequence are often heterogeneous
and polysemic [99,100]. To capture the underlying intents of these actions more effectively,
researchers have started assigning varying weights to different types of behaviors. For
instance, a purchase action is typically considered a stronger indicator of a user’s genuine
interest and preference for a product or service compared to a mere click action. Therefore,
distinguishing different types of behaviors becomes particularly critical [101]. This concept
has been concretely implemented in the CBS model. CBS categorizes the given behavioral
Appl. Sci. 2023, 13, 11378 14 of 31

sequence into target sequences and supporting sequences based on the types of behaviors
observed [102]. This enables the target sequences that are closely related to the most
predictive types of behavior (e.g., purchasing) to be highlighted. The authors proposed
a novel approach that integrates a basket representation into a recurrent layer to capture
sequential effects. The implicit recurrent representation of ht is presented as shown in
Equation (12).
h t = g ( Φ b bt + Φ h h t − 1 + Ω h ) (12)
This approach takes both temporal ordering and continuity into account. The formula
explicitly incorporates basket representation, placing particular emphasis on the influence
that user choices at specific time steps exert on the hidden state. This distinguishes it from
more traditional sequence recommendation methods.
A similar line of thinking has also been implemented in BINN [103]. BINN aims to
capture users’ current interests by utilizing all types of behaviors, such as clicks, purchases,
and favorites, while exclusively employing behaviors related to purchases (e.g., buying,
adding to cart, and favoriting) to reflect users’ long-term preferences.
In another study [104], the authors attempted to integrate the specific representations
of each type of behavior with the corresponding item embedding vectors, aiming to capture
users’ interests more comprehensively. The researchers integrated a masked beam search
and determinantal point process (DPP) selection to produce a high-quality and diversified
bundle list with an appropriate bundle size. The formula is presented in Equation (13).

exp(htT e j − mt,j )
masked_so f tmax(ht , E, mt ) j = (13)
∑N
ĵ=1
exp(htT ê j − mt, ĵ )

Specifically, the beam search retains the top k most promising candidate bundle lists
in each step and subsequently extends these lists in the next step to progressively construct
longer recommended bundle lists. This paper employed a beam search to generate more
accurate and relevant bundle recommendation lists while also ensuring the computational
efficiency of the model.
After all, with the continual advancements achieved in deep learning and natural
language processing technologies, sequential recommender systems are well-primed for
further optimization and development. More advanced model architectures, such as trans-
formers and BERT, can be employed to capture complex sequential patterns. Moreover,
integrating these techniques with other recommendation approaches, such as those based
on knowledge graphs, may enhance both the diversity and accuracy of sequential rec-
ommender systems. As the Internet of Things and smart home technologies continue to
proliferate, the utilization of sequential recommendation is expected to further broaden in
everyday life.
Recent research on sequential recommendation remains highly active, with some
researchers considering temporal information in sequential recommendation [105]. They
aggregate sequential information and collaborative signals in user behavior sequences,
thereby taking a more comprehensive approach to considering the information in sequen-
tial behaviors. In order to enhance the transition probabilities between items in sequences,
researchers employ contrastive learning for sequential recommendation [106]. They in-
troduce two informative augmentation operators leveraging item correlations to create
high-quality views for contrastive learning.
Traditionally, recommender systems have often employed static strategies to charac-
terize the interactions between users and items, relying on long-term historical behaviors to
infer user interests. However, the selection of a particular item by a user is not solely based
on their long-term stable preferences but is increasingly driven by their fluctuating short-
term interests [107]. In fact, user interests dynamically evolve over time. This is precisely
where sequential recommendation comes into play. Sequential recommendation models the
user-product interaction history as a dynamic sequence and utilizes the temporal dependen-
cies within this sequence to capture the evolving user preferences, leading to more precise
Appl. Sci. 2023, 13, 11378 15 of 31

and timely recommendations. Deep learning enables the hierarchical treatment of complex
problems, swiftly identifying the latent patterns and relationships between different layers
of data. The integration of deep learning with a sequential recommendation can also adjust
recommendations based on dynamic user preference changes, further enhancing their
effectiveness [108].
Sequential recommender systems excel at accurately predicting users’ immediate
interests and dynamically adapting to their behavior changes and shifting preferences.
This approach not only guarantees content diversity but also furnishes recommendations
that are highly pertinent to users’ recent activities, even if these suggestions may not fully
correspond with users’ earlier behavioral patterns. While sequential recommendation
undoubtedly offers unique advantages, it is important to mention that handling time-series
data can introduce additional computational demands on the model [109]. Furthermore, an
undue emphasis on users’ short-term actions carries its own set of risks, potentially leading
to a neglect of their long-term interests and preferences.

3.3. Cross-Domain Recommendation

In the preceding sections, our discussion was primarily focused on generating recom-
mendations based on users’ historical behaviors and content characteristics, usually within
a single application scenario or domain. However, to reap benefits from a variety of services,
users frequently engage with multiple social media platforms [110]. Cross-domain recom-
mendation (CDR) has been introduced to utilize richer information from multiple domains,
aiming to improve recommendation performance. Studies indicate strong correlations
between user-generated data across different domains [111]. Traditional data mining tech-
niques are primarily designed for single-domain analysis, and by overlooking data from
other domains, they often encounter issues related to data sparsity [112]. Consequently,
addressing the utilization of cross-domain data to enhance the comprehensiveness and
accuracy of recommendations has become a prominent research focus. By aggregating data
from different domains, CDR methods not only compensate for the missing information in
a single domain but also holistically harness the value embedded in social media data. The
framework is shown in Figure 5 below.
As depicted in Figure 5, the input layer is designed to accommodate user behavior and
item attributes from disparate domains. After performing feature extraction, the data acquired
from these diverse domains are aggregated into a shared hidden layer, which is tasked with
capturing cross-domain patterns and relationships. Within the hidden layer, neural networks
can be employed to integrate information from various domains. By employing this design,
the model can generate an output layer that synthesizes information from multiple domains,
thereby facilitating more comprehensive and accurate recommendations.
Cross-domain recommender systems aim to improve recommendation performance [113]
by utilizing data from various domains, particularly in scenarios with data sparsity or cold
start challenges. These systems take users’ behaviors and preferences across different domains
into account, harnessing comprehensive multidomain information to provide a more holistic
user profile and, in turn, bolster recommendation accuracy. Cross-domain recommender
systems exploit multidomain learning and knowledge transfer [114], diversifying their recom-
mended content to satisfy the needs of varied scenarios and domains. However, multidomain
data processing and the associated knowledge transfer step may also result in increased
computational costs.
Traditional recommender systems grapple with two main challenges: the cold start
problem and data sparsity [115]. CDR methods offer an effective remedy for these issues by
analyzing user interactions from other domains to capture user preferences from specific
perspectives. This can be employed to enhance data in the target domain or supplement
information during the initialization process for new users. While traditional recommender
systems build and analyze a recommendation model within the current domain, cross-
domain recommendations necessitate deciding which information should be transferred
between distinct domains and the means of performing such transfers. Cross-domain
 ommendations based on users historical behaviors and content characteristics, usually
within a single application scenario or domain. However, to reap benefits from a variety
of services, users frequently engage with multiple social media platforms [110]. Cross-
domain recommendation (CDR) has been introduced to utilize richer information from
multiple domains, aiming to improve recommendation performance. Studies indicate
Appl. Sci. 2023, 13, 11378 strong correlations between user-generated data across different domains [111]. Tradi- 16 of 31
tional data mining techniques are primarily designed for single-domain analysis, and by
overlooking data from other domains, they often encounter issues related to data sparsity
[112]. Consequently, addressing the utilization of cross-domain data to enhance the com-
recommendation approaches can harness user interaction data from other domains for
prehensiveness and accuracy of recommendations has become a prominent research fo-
auxiliary analysis purposes,
cus. By aggregating data fromcapturing certain user
different domains, CDRpreferences. This compensate
methods not only enables enhanced
for
recommendations
the missing information in a single domain but also holistically harness themultiple
to be produced in the target domain or even across domains
value embed-
and
ded provides additional
in social media information
data. The framework forisnew user
shown in initialization,
Figure 5 below.thereby addressing the
two challenges that are endemic to traditional recommender systems.

Figure 5. Cross-domain recommendation procedure.

Figure 5. Cross-domain recommendation procedure.

In recent years, researchers have focused on addressing data sparsity and cold-start
issues in the target domain for cross-domain recommendation. They have conducted
extensive studies on deep learning-based cross-domain recommendation methods. These
approaches consider deep auxiliary information, including neighboring domain descrip-
tions and rating information [116–118]. For instance, a multilayer perceptron (MLP) utilizes
the latent features of common users in the source domain as inputs and generates latent
rating features for those users in the target domain as outputs during the integration step.
Through this process, the network is trained, obtaining a nonlinear mapping function. Op-
timization is finally achieved via backpropagation, resulting in improved recommendation
outcomes. Later, Natarajan, S et al. proposed CD-SemMF [38], which uses the semantic
relevance of a project to obtain better results.
In the realm of CDR, multiple domains are encountered, each with an information
volume denoted as IN. Within this framework, Domain A possesses dense data, and
a newer Domain B is characterized by data sparsity and suboptimal recommendation
performance. The inequality INA  INB suggests that the information volume in Domain
A significantly surpasses that of Domain B. This discrepancy prompts the consideration
of whether information from Domain A can be transferred for utilization in Domain B, a
process termed cross-domain transfer. Domain A is designated as the source or auxiliary
domain, symbolized as DS , whereas Domain B is identified as the target domain, denoted
as DT . The source domain can comprise multiple distinct domains, which are aimed at
supplementing and enriching the information within the target domain DT . The respective
user sets in domains S and T are US and UT , the item sets are IS and IT , the rating sets are RS
and RT, and the comment sets are S and CT , respectively. Thus, the objective is to transfer
the knowledge embedded in the source domain to enhance the recommendation quality
of the target domain. The distinction between cross-domain recommendation and other
Appl. Sci. 2023, 13, x FOR PEER REVIEW 18 of 32

Appl. Sci. 2023, 13, 11378 quality of the target domain. The distinction between cross-domain recommendation17and of 31
other methods lies in its feature input, which consolidates multiple domains. Let V repre-
sent the user s behavioral features; user embedding for cross-domain recommendation is
illustrated in Equation (14).
methods lies in its feature input, which consolidates multiple domains. Let V represent the

in Equation (14).
ui = [w1 ⋅ f (u is) +for
user’s behavioral features; user embedding
⋅ u ti , v1h , v 2h ,...,
w2cross-domain v hn]
recommendation is illustrated
(14)
ui = [w1 · f (uis ) + w2 · uit , v1h , v2h , ..., vnh ] (14)
Cross-domain recommender systems comprise three pivotal modules: a user model-
ling module, an itemrecommender
Cross-domain modelling module,
systemsand a recommendation
comprise three pivotal algorithm
modules: module,
a user mod- as
illustrated in Figure
elling module, 6. modelling module, and a recommendation algorithm module, as
an item
illustrated in Figure 6.

Figure 6. Process of the cross-domain recommender system.

Figure 6. Process of the cross-domain recommender system.
CDR seeks to merge data from various domains by integrating supplementary in-
CDR seeks
formation fromto merge
other data from
domains various recommendations
to improve domains by integrating withinsupplementary
the target domain infor-
or
mation from other domains to improve recommendations within
even across multiple domains. However, it is not a matter of arbitrarily combining any the target domain or
even
group across multiple
of domains fordomains.
cross-domain However, it is not a matter
recommendation purposes;of arbitrarily
typically, combining
there should anybe
group of domains for cross-domain recommendation purposes;
some overlapping information among the different domains. Compared to conventional typically, there should be
some overlapping
recommender information
systems, among the different
CDR necessitates domains.
meticulously Compared
considering to conventional
which information
recommender systems,between
should be transferred CDR necessitates
domains and meticulously
the methodologyconsidering
used for which information
such transfers.
should be transferred
CDR between
tasks are closely domains
linked and the methodology
to user-related aspects, with used
twofor such transfers.
fundamental factors
CDRthis
driving tasks are closelythe
connection: linked
range to of
user-related aspects, with two fundamental
items for recommendations and the diversity factorsof
driving this connection: the range of items for recommendations
target users, giving rise to a variety of recommendation scenarios. CDR algorithms canand the diversity of tar-
get
be users, giving
classified rise to atovariety
according the target of recommendation
domain into three scenarios. CDRsingle-domain
categories: algorithms can be
CDR,
classified according
cross-domain CDR,to themulti-target
and target domain CDR. into three categories: single-domain CDR, cross-
domain CDR, and multi-target CDR.
3.3.1. Single-Target CDR
3.3.1. Single-target
Single-TargetCDR CDR is the conventional approach in the CDR domain, and most CDR
methods primarily
Single-target CDR focus
is on
thethis scenario. Itapproach
conventional involvesin recommending
the CDR domain, itemsandto users
most in the
CDR
target domain
methods by utilizing
primarily focus oninformation
this scenario. acquired fromrecommending
It involves the source domain.itemsSingle-objective
to users in the
CDR methods predominantly focus on transferring valuable knowledge.
target domain by utilizing information acquired from the source domain. Single-objective
There arepredominantly
CDR methods also various recommendation
focus on transferring methods, it initially
valuable establishes connections
knowledge.
based on common content elements such as user/item
There are also various recommendation methods, it initially establishes attributes [119], ‘like’ data, and
connections
browsing or viewing histories [120]. Subsequently, they employ
based on common content elements such as user/item attributes [119], like data, and these connections for
the cross-domain transfer of user/item data or knowledge. Embedding-based
browsing or viewing histories [120]. Subsequently, they employ these connections for the transfer
methods often
cross-domain projectofinstances
transfer user/itemfrom databoth the sourceEmbedding-based
or knowledge. and target domains into ameth-
transfer novel
feature
ods oftenspace,
project where instances
instances from from bothsource
both the domainsand display similarities.
target domains into aOther
novelmethods
feature
space, where instances from both domains display similarities. Other methods oftenmodels,
often adopt distinct techniques or philosophies, including Bayesian latent factor adopt
interest drift, ternary relations, and reinforcement learning. Rating pattern-based transfer
distinct techniques or philosophies, including Bayesian latent factor models, interest drift,
methods typically begin by learning the distinct rating patterns of users from the source
ternary relations, and reinforcement learning. Rating pattern-based transfer methods typ-
domain. Subsequently, these patterns from the source domain are harnessed to enhance
ically begin by learning the distinct rating patterns of users from the source domain.
the model’s recommendation performance in the target domain, facilitating knowledge
transfer. Shapira et al. and Tiroshi et al. [121] both focused their efforts on leveraging social
networks to enhance recommender systems. While the former harnessed Facebook’s friend
relationships to reinforce user models in the target domain, the latter delved deeper, em-
Appl. Sci. 2023, 13, 11378 18 of 31

ploying random walk algorithms to extract latent user information within social networks.
These approaches can be perceived as endeavours to uncover both explicit and implicit
user relationships within the constructed network.
Subsequently, Jiang et al. [122] implemented a semisupervised transfer learning
methodology, exploring the user interest similarities between the source and target do-
mains, thereby indicating the transferability of user interests across different domains.
Their research contradicts the traditional belief, suggesting that even with a small number
of overlapping users, valuable information can be provided for the entire system. By ex-
tracting and utilizing information from overlapping user groups, semisupervised transfer
learning approaches can be employed to deal with the issue.
However, these studies also highlighted challenges, emphasizing that the effective-
ness of the recommendation process heavily relies on the extent of user overlap across
different domains. This identified limitation provides a direction for subsequent research,
specifically on how to deliver effective recommendations even when the user intersection
is minimal. To exploit these latent and obscured relationships, studies have increasingly
pivoted towards multidomain recommendation (MDR). Compared to single-target CDR,
the focus of MDR primarily lies in handling recommendations across different domains.
Although multidomain methods can be applied within the context of CDR, their scopes in
practical applications often face certain constraints. To address data sparsity and the other
challenges inherent in multidomain recommendation scenarios, Zhang et al. introduced
MCF. Successive studies [123–125] integrated more advanced techniques such as feature
fusion aiming to further enhance the accuracy of the output recommendations.
In summary, from harnessing both explicit and implicit information within social
networks to the transfer of interests across domains and onwards to the exploration of
multidomain recommendations, researchers have continuously striven to identify more
effective and universal recommendation methods that can cater to user needs across various
domains and scenarios.

3.3.2. Dual-Target CDR

Single-target approaches can only utilize auxiliary information from a richer domain
to assist a sparser domain. However, the richness of various types of information varies
across different domains. If these types of information can be effectively harnessed, it
is possible to enhance the recommendation performance in all domains simultaneously,
rather than limiting improvements to a single target domain. To address this, dual-target
CDR and multi-target CDR have been recently proposed to enhance recommendation
performance across dual/multiple domains. To achieve cross-domain recommendations
for dual-target domains, researchers have begun to develop various models. These models
often incorporate advanced techniques such as graph models and attention mechanisms to
facilitate knowledge transfer between different domains.
Research on recommender systems that operate across multiple target domains has
been increasingly explored in depth. Zhu et al. [39] were pioneers in that they introduced
a dual-target-domain CDR framework called DTCDR. This framework capitalizes on
multisource information, ensuring that more detailed embeddings are produced for both
users and items. To further integrate this information, the authors employed multitask
learning techniques to merge the embeddings of overlapping users.
Building upon this foundation, Liu et al. [126] delved deeper into the embeddings
of overlapping users. They created two distinct heterogeneous graphs utilizing ratings
and content information from both domains to produce more representative user and
item embeddings. Taking hyperparameters and data sparsity into consideration, they
achieved more accurate embeddings for these users. Subsequently, the DDTCDR model
presented in [40], another dual-target domain CDR framework, took a different approach,
emphasizing the bidirectional implicit relationships between users and items. It utilized
implicit orthogonal mapping to learn user preferences, allowing for the bidirectional
transfer of user embeddings between two domains.
Appl. Sci. 2023, 13, 11378 19 of 31

Furthermore, in [127], the authors incorporated graph embedding techniques. They

not only utilized connectivity of the user-item graph within a single domain via a novel
feature propagation layer but also facilitated bidirectional knowledge transfer between
the two domains by employing common users as bridges. Moreover, distinct from prior
cross-domain collaborative filtering methods, they integrated both shared user features
and domain-specific attributes during the transfer process.
In summary, these studies were dedicated to optimizing the performance of dual-
target-domain recommender systems. Their collective efforts have contributed valuably to
enhancing the accuracy and robustness of cross-domain recommender systems.

3.3.3. Multi-Target CDR

Multitarget domain CDR is a recommendation to users in either domain for items in
both domains. However, its objectives are more intricate, aiming to provide a comprehen-
sive solution to the data sparsity problem. The fundamental concept behind multitarget
CDR is to harness additional auxiliary information from multiple domains to enhance
recommendation performance. Theoretically, if one can identify and effectively utilize
a sufficient number of relevant domains, the longstanding data sparsity issue can be
substantially mitigated. Nonetheless, beyond the challenges that are inherent in single-
and dual-target-domain CDR scenarios, negative transfer becomes inevitable in practical
multitarget-domain CDR contexts. Specifically, with the addition of domains, the recom-
mendation performance of the constructed model might deteriorate, especially in scenarios
where the domains are sparse.
The multitarget domain task presents a particularly challenging recommendation
scenario, and as of now, no perfect solution tailored to this objective is available. In [128],
the authors employed a shared heterogeneous graph to generate richer user and item
feature embeddings across domains. The MDCDR [41] utilizes auxiliary data from source
domains to improve recommendation quality across multiple domains. However, none of
these approaches effectively addresses the negative transfer issue.
As the complexity of multiplatform and multidevice environments intensifies, the
demand and significance of cross-domain recommendations are anticipated to escalate
further. Advanced models, such as deep learning and graph neural networks, may offer
more robust support for cross-domain recommendation tasks. These developments have
introduced new challenges such as feature mapping and negative transfer. These emerging
research trends inspire us to delve into the challenges within CDR and provide an overview
of the ongoing research advancements.

3.4. Social Recommendation

Users’ decision-making processes are influenced by their social circles, and numerous
researchers have integrated social relations into recommender systems. The main goal is
to integrate users’ social information as auxiliary data into conventional recommendation
frameworks, thereby improving the accuracy of the resulting recommendations. Distinct
from other recommendation strategies, social recommendation focuses on leveraging users’
social network information to generate more personalized and accurate recommendations.
The framework for social recommendation is illustrated in Figure 7.
The primary concept involves aggregating features from both a node itself and its
neighboring nodes to generate a consolidated representation. The inputs of this approach
are users, items, and the social relationship network among the users. After performing
embedding, neural networks are generally employed to handle the relationships between
users. The final output consists of item recommendations.
Appl. Sci. 2023, 13, x FOR PEER REVIEW 21 of 32
Appl. Sci. 2023, 13, 11378 20 of 31

Figure 7. Social recommendation process.

Figure 7. Social recommendation process.
Numerous studies have integrated social relational information into collaborative
filtering-based
The primaryrecommendations with the goal
concept involves aggregating of modelling
features from both the ainterrelationships
node itself and itsbe-
tween users.
neighboring The to
nodes evolution
generateofa social recommendation
consolidated has spanned
representation. three
The inputs ofphases. Initially,
this approach
researchers
are users, items,began
andtothe
recognize the potential
social relationship value among
network of trust the
relationships
users. Afterwithin a user’s
performing
social network
embedding, neuralfornetworks
recommendation
are generally purposes.
employed Leveraging
to handlerandom walks, a trust-based
the relationships between
recommendation method was proposed in
users. The final output consists of item recommendations. [129]. The trust relationships among users were
employed
Numerous as weights to adjust
studies have the recommendation
integrated outcomes. Subsequently,
social relational information withfil-
into collaborative the
application of matrix factorization techniques, models integrating
tering-based recommendations with the goal of modelling the interrelationships between matrix factorization
with The
users. social networksofgained
evolution social popularity.
recommendation Buildinghasupon trustthree
spanned relationships, these models
phases. Initially, re-
endeavored to incorporate both strong and weak social ties into a
searchers began to recognize the potential value of trust relationships within a user s so- unified framework,
allowing
cial network recommendations
for recommendation to benefit fromLeveraging
purposes. both types random
of relations [130].
walks, With the help
a trust-based rec-of
deep learning, the field has progressed into the current era of
ommendation method was proposed in [129]. The trust relationships among users were graph-based social recom-
mendations.
employed These to
as weights models
adjustusually take data from
the recommendation a user’s social
outcomes. networkwith
Subsequently, as input and
the ap-
subsequently utilize neural networks to forecast users’ interests [131].
plication of matrix factorization techniques, models integrating matrix factorization with
A graph-based social recommendation has emerged as a focal area of academic interest.
social networks gained popularity. Building upon trust relationships, these models en-
The advantages of GNNs in terms of mining graph-structured data have facilitated re-
deavored to incorporate both strong and weak social ties into a unified framework, allow-
search on social network recommender systems. GNNs possess the capability to efficiently
ing recommendations to benefit from both types of relations [130]. With the help of deep
learn and extract features from graph structures, demonstrating commendable results in
learning, the field has progressed into the current era of graph-based social recommenda-
user-item representation learning tasks. Consequently, numerous scholars have integrated
tions. These models usually take data from a user s social network as input and subse-
social networks into recommender system research and employed GNNs to construct
quently utilize neural networks to forecast users interests [131].
recommender system models grounded in graph representation learning. For instance, the
A graph-based social recommendation has emerged as a focal area of academic in-
GC-MC model [132] and the NGCF model [27] graph convolutions to identify user-item
terest. The advantages of GNNs in terms of mining graph-structured data have facilitated
interactions in the original space, yielding enhanced recommendation outcomes in practical
research on social network recommender systems. GNNs possess the capability to effi-
applications. Moreover, some researchers have found that adding nonlinear activation
ciently learn and extract features from graph structures, demonstrating commendable re-
functions to GNNs does not significantly improve the accuracy of the resulting recom-
sults in user-item
mendations. Thisrepresentation
observation led learning tasks. Consequently,
to the proposal of a streamlined numerous scholars have
GNN recommendation
integrated social networks into recommender system research
model, LightGCN [42], which outperforms NGCF in terms of recommendation and employed GNNs to
perfor-
construct recommender system models grounded in graph representation
mance. LR-GCCF [43] eliminates nonlinear transformations from graph neural networks learning. For
instance, the GC-MC
and substitutes model
them with[132]
linear and the NGCFpropagation,
embedding model [27] graph convolutions
thereby to identify
reducing unnecessary
Appl. Sci. 2023, 13, 11378 21 of 31

operations, simplifying the complexity of the model, and still ensuring that satisfactory
performance metrics are produced.
Social recommendation algorithms can be broadly classified into three main categories:
traditional collaborative filtering methods, deep social recommendation methods relying on
graph embeddings, and social recommendation methods based on graph neural networks.

3.4.1. Traditional Collaborative Filtering-Based Social Recommender Systems

In traditional social network-based recommendations, two separate graphs need to be
constructed: one representing user-item interactions, and another representing user-user
social connections. Therefore, social recommendation algorithms utilizing graph neural
networks require knowledge extraction from both graphs simultaneously for accurate
inference. They can leverage coarse-grained and sparse user trust relationships to enhance
conventional collaborative filtering methods. Essentially, this approach involves the use
of similarity metrics to identify users similar to the target user. Subsequently, different
weights are assigned to ratings for the target item based on this similarity, facilitating
the computation of the user’s rating for that item. The method for determining user-to-
user similarity is crucial to this concept. Consequently, existing social recommendation
algorithms have explored various designs for graph neural network encoders and adopted
different architectures based on their objectives. Upon determining the similarity between
two entities, the set of users most similar to the target user can be identified based on
their ratings. This, in turn, allows for the recommendation of items that the user may find
interesting and ranks them accordingly.

3.4.2. Deep Social Recommendation Based on Graph Embedding

Traditional social recommender systems only integrate explicit trust relationships
between users and ignore implicit trust relationships. Graph embedding models represent
all entities as nodes and their relationships as edges in a graph. These models transform
nodes into vector representations, allowing them to retain their approximate relationships
from the graph structure in a Euclidean space. Afterwards, vector distances are computed
to predict the scores of candidate items for the target user.
Using the application of graph embedding techniques, it becomes possible to acquire
low-dimensional feature representations of users within a social network. These low-
dimensional representations enable the inference of detailed trust relationships between
users. Consequently, by incorporating the weighted ratings from trusted and similar
users for a specific item, predictions regarding the user’s rating for that item can be
Appl. Sci. 2023, 13, x FOR PEER REVIEW 23 of 32
generated. Figure 8 illustrates the framework of a recommendation algorithm employing
graph embedding models.

Figure
Figure 8.
8. Recommendation
Recommendationframework
framework based
based on
on graph
graph embedding.
embedding.

Prominent graph
Prominent graph embedding
embeddingtechniques,
techniques,such
suchas
asDeep
DeepWalk
Walk[133],
[133],graph
graphfactoriza-
factoriza-
tion [44], and LINE [134], possess distinct characteristics and limitations. For instance,
tion [44], and LINE [134], possess distinct characteristics and limitations. For instance,
graph factorization [45] is suited for performing embedding learning in large-scale
graph factorization [45] is suited for performing embedding learning in large-scale net- net-
works but fails to preserve the network’s global structure and is solely applicable
works but fails to preserve the network s global structure and is solely applicable to un- to
directed information networks. In contrast, while Deep Walk reveals certain attributes of
an information network using a random walk algorithm, it lacks a comprehensive depic-
tion of the preserved network properties and overlooks the connections between feature
vectors. In contrast, LINE effectively maintains both local and global structural aspects of
Appl. Sci. 2023, 13, 11378 22 of 31

undirected information networks. In contrast, while Deep Walk reveals certain attributes
of an information network using a random walk algorithm, it lacks a comprehensive
depiction of the preserved network properties and overlooks the connections between
feature vectors. In contrast, LINE effectively maintains both local and global structural
aspects of information networks and has found widespread application in recommender
systems. Additionally, the prowess of graph embedding techniques in terms of extracting
the semantic relationships within structures is evident in applications such as the work
performed by Xie et al. [135] on POI recommendations and the research conducted by
Yin et al. [136] on low-dimensional vector representations for users and items. Graph
embedding representation learning, with its powerful representational capabilities, low
memory consumption, and efficient computations, has been widely researched and applied
for semantic relationship extraction in network structures [137].
Overall, given the advantages of graph embedding techniques in modelling and
optimization tasks, they present broad prospects for various applications and merit fur-
ther exploration.

3.4.3. Social Recommendation Based on GNN

Over time, neural networks’ applications in the recommendation field have signif-
icantly broadened. Graph neural networks (GNNs) represent innovative extensions of
neural networks, drawing inspiration from convolutional neural networks (CNNs) and the
concept of graph embeddings. They are adept at extracting and representing features in
graph-based data. Characterized by their efficiency and scalability, GNNs have demon-
strated profound capabilities when handling graph data. Users and items are considered
as nodes in the graph, and interactions are represented as edges. This process leads to the
transformation of user-item interactions into embedded representations within the graph.
By integrating these social relationships into recommender systems and subsequently feed-
ing them into neural networks, more comprehensive and personalized recommendations
can be provided.
The goal of a GNN is to acquire a state embedding that captures the neighborhood
information of each node. The embedding is a z-dimensional real-valued vector associated
with node u that can be used to generate a node label output, representing the vector output
of node u at a given network layer. Local transfer functions are shared across all nodes and
update node state based on input neighborhood. Finally, the output is produced. Then,
based on the characteristics of nodes, the characteristics of adjacent edges, and the state
embeddings and features of adjacent fixed points, the social relationship model is further
improved, for example, by considering the impact of various social relationships on user
behaviors. Attention mechanisms, among other techniques, have been introduced [138].
Broadly speaking, the architecture of a social recommendation model comprises three key
components: an encoder, a decoder, and a loss function. The encoder [139] utilizes different
graph neural network encoders to represent users and items as low-dimensional vectors
(i.e., embeddings). The decoder then predicts each user’s preference for each item by
performing different operations on the obtained user and item embeddings. Optimization
is finally achieved using various loss functions [140]. The performance of social recommen-
dation methods based on GNN largely depends on their encoders.
Beyond user behaviors, social recommender systems also take various types of infor-
mation into account, including social interactions and shared interests. They operate on the
premise that users are likely to share similar interests with friends or contacts within their
social networks, utilizing these social relationships to enhance the accuracy and relevance of
recommendations. Additionally, the strength and depth levels of social ties are considered
to enhance the produced recommendation outcomes. For new users, the cold-start problem
can be mitigated via their social networks. Furthermore, recommendations stemming from
friends are more likely to be trusted and accepted by users. However, social recommen-
dations exhibit some limitations. Utilizing users’ social network data may raise privacy
Appl. Sci. 2023, 13, 11378 23 of 31

and security concerns [141]. Moreover, not all social interactions correlate with genuine
preferences, necessitating the filtering of irrelevant or misleading information.
Graph neural networks offer robust mechanisms for recommender systems to capture
the intricate interactions and relationships between users and items [142], with several
models demonstrating significant influences and innovations. Initially, GraphSAGE [28], a
pioneering spatial GNN model, marked an essential milestone. It moved away from relying
on the entirety of graph information and instead strategically sampled the neighbors of
target nodes and combined their embeddings. This approach achieved efficient target
embedding updates and offered a viable tool for providing social recommendations in
large-scale networks. Following GraphSAGE, the GAT [143] further refined the spatial
GNN concept, addressing the limitations of the previous spectral methods regarding key
issues such as model generalization. Notably, the GAT introduced an attention mechanism,
enabling the model to allocate different weights based on the importance of neighboring
nodes, leading to a more selective neighborhood feature aggregation process. This strategy
enhanced the model’s discriminatory capacity, achieving commendable results in social
recommendation tasks. For dealing with heterogeneous graph structures, HetGNN [144]
emerged as a representative technique. Distinct from traditional GNN models, HetGNN
devises specific aggregation strategies for nodes and edges that are present in heterogeneous
graphs. Initially, it segregates neighbors based on their types and then employs LSTM and
MEAN operations to individually process these subsets, effectively capturing the abundant
structural and attribute information inherent in heterogeneous graphs. Furthermore, it is
noteworthy that HetGNN implements spectral GNN methods on hypergraph structures,
propelling recommender system research in a novel direction. Regarding the exploration of
social network structures, DiffNet and DiffNet++ stand as pivotal milestones. Both methods
emphasize delving into higher-order social structures to achieve enhanced recommendation
accuracy and efficacy. DiffNet accentuates capturing higher-order neighbor information
within the network to boost its recommendation results [145], while DiffNet++ further
refines and extends this base [46]. Following DiffNet and DiffNet++, GraphRec emerged
as an innovative approach, emphasizing not only capturing interactions between users
and items but also jointly capturing opinions about both users and items [146]. This
dual capture strategy is designed to ensure recommendation accuracy while maintaining
system interpretability.
With the continuous evolution of social media and platforms, the significance of social
recommender systems is set to expand further [147]. Advanced technologies such as deep
learning and graph neural networks have notably enhanced the accuracy and efficiency
of social recommendation [148]. Moreover, striking a balance between user privacy and
the quality of recommendations will emerge as a pivotal research direction [149]. In the
future, we may also witness an increase in cross-platform social recommendation solutions,
ensuring that users receive consistent and high-quality recommendation experiences across
diverse platforms.

4. Challenges and Developments

Recommender systems are designed to help users discover items that align with
their preferences from a wide range of potential recommendations. Deep learning can aid
researchers in improving the effectiveness of these recommendations. This paper analyzes
four distinct types of recommender systems, including content-based recommendation,
sequential recommendation, cross-domain recommendation, and social recommender
systems. Although the integration of these recommendation techniques with deep learning
has achieved satisfactory results, challenges remain [150]. Future research efforts can be
pursued in the following areas.
(1) Security enhancements are needed
With the growth of networking sites, accurately recommending items of interest to
users has become one of the key strategies employed by various websites to attract users.
Only by digging deep into multidimensional user data can recommendations be identified
Appl. Sci. 2023, 13, 11378 24 of 31

that truly align with user preferences [151]. In reality, while users expect recommender
systems to suggest items of interest, they do not want their other private details to be
disclosed [152]. Current research primarily involves distorting and obfuscating user data
to ensure privacy [153]. Although this data perturbation strategy does protect the user’s
personal information, it can lead to the inaccurate extraction of user data, significantly
compromising the resulting recommendation accuracy. Therefore, future research could
focus on a methodology that not only safeguards user privacy but also enhances recom-
mendation precision.
(2) Methods for extracting user preference features are lacking
At present, recommendation subjects rely heavily on users’ ratings or feedback on
recommended items. Current research lacks adequate modeling methods that can multi-
dimensionally extract user features and recommended items [154], as well as their linear
and non-linear relationships [155]. Although neural networks can address this issue to
some extent, there remains significant room for improvement. Consequently, forthcoming
research should incorporate a more diverse range of methods to extract the features of both
users and recommended entities.
(3) Evaluation metrics are singular
When evaluating the performance of recommender systems, the existing studies
primarily emphasize the accuracy of results and associated precision rate, considering
accuracy as the key metric for determining the effectiveness of recommender systems.
However, when users interact with these applications in real-world scenarios, they not only
expect the system to provide accurate recommendations for items of interest but also seek a
broader and more innovative range of suggestions [156]. Hence, the novelty and diversity
of the output recommendations should be considered in the future.
In summary, in addition to the aforementioned aspects, there are several directions for
future research. There may be an increased focus on the interpretability of deep learning
models. This is because, in practical applications, there is a growing demand from users and
regulatory bodies for greater transparency and comprehensibility in recommender system
decisions. With the proliferation of multimodal data (text, images, audio, etc.), future
recommender systems may integrate these data sources more extensively to provide a richer
and more diverse recommendation experience. The application of reinforcement learning
in recommender systems may become more widespread, optimizing recommendation
strategies via user interactions for higher long-term returns.
Future research may explore how to transfer deep learning models from one domain to
another to enhance the generalizability of recommender systems. Federated learning, while
protecting user privacy, allows different institutions and platforms to collaboratively train
recommendation models and may become an important direction in future recommender
system research.
Despite improvements in handling sparse data with deep learning methods, data
sparsity and cold-start problems remain challenges. Specifically, providing accurate rec-
ommendations for new users and items remains a challenge. Privacy and fairness have
long been key challenges in the field of recommender systems. Balancing personalized
recommendations with user privacy and ensuring that recommendation system decisions
are fair will be a focus of future research. Deep learning models often require substantial
computational resources. Building scalable and efficient deep learning recommender sys-
tems in practical applications remains a challenge. Deep learning methods typically require
a large amount of training data, so encouraging user participation in feedback collection to
improve the performance of recommender systems remains a question.

5. Conclusions
As technologies like deep learning, data mining, and predictive algorithms mature,
future research will focus on enhancing the accuracy, security, and privacy of recommender
systems. This article explores both traditional recommendation methodologies and those
incorporating various deep learning models. We compare the differences between conven-
Appl. Sci. 2023, 13, 11378 25 of 31

tional recommendation models and deep learning-based approaches, summarizing the

prevalent challenges in recommender systems. We also offer insights into future research
directions for recommender systems, benefiting researchers who are interested in the fields
of recommender systems or deep learning.

Author Contributions: Conceptualization, H.Z.; methodology, H.Z.; validation, H.Z.; formal analysis,
H.Z., F.X. and H.C.; investigation, H.Z.; resources, H.Z.; data curation, H.Z.; writing—original draft
preparation, H.Z.; writing—review and editing, H.Z., F.X. and H.C.; visualization, H.Z., F.X. and
H.C.; supervision, F.X.; project administration, F.X. and H.C.; funding acquisition, H.Z., F.X. and H.C.
All authors have read and agreed to the published version of the manuscript.
Funding: This research was funded by the Fundamental Research Funds for the Central Universities
(No. 2022JBMC005), the National Natural Science Foundation of China under Grant 61872033 and
Grant 72004009, the National Key R&D Program of China under Grant 2018YFC0832304, and the
Beijing Nova Program from Beijing Municipal Science & Technology Commission under Grant
Z201100006820015.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: No new data were created or analyzed in this study. Data sharing is
not applicable to this article.
Conflicts of Interest: The authors declare that they have no conflict of interest or personal relation-
ships that could have appeared to influence the work reported in this paper.

References
1. Zhu, Y.; Lin, Q.; Lu, H. Recommending learning objects through attentive heterogeneous graph convolution and operation-aware
neural network. IEEE Trans. Knowl. Data Eng. 2021, 35, 4178–4189. [CrossRef]
2. Leiva, M.; Budán, M.C.D.; Simari, G.I. Guidelines for the analysis and design of argumentation-based recommendation systems.
IEEE Intell. Syst. 2020, 35, 28–37. [CrossRef]
3. Goldberg, K.; Roeder, T.; Gupta, D.; Perkins, C. Eigentaste: A constant time collaborative filtering algorithm. Inf. Retr. 2001, 4,
133–151. [CrossRef]
4. Abiodun, O.I.; Jantan, A.; Omolara, A.E.; Dada, K.V.; Umar, A.M.; Linus, O.U.; Arshad, H.; Kazaure, A.A.; Gana, U.; Kiru,
M.U. Comprehensive review of artificial neural network applications to pattern recognition. IEEE Access 2019, 7, 158820–158846.
[CrossRef]
5. Gheisari, M.; Wang, G.; Bhuiyan, M.Z.A. A survey on deep learning in big data. In Proceedings of the 2017 IEEE International
Conference on Computational Science and Engineering (CSE) and IEEE International Conference on Embedded and Ubiquitous
Computing (EUC), Guangzhou, China, 21–24 July 2017; IEEE: Piscataway, NJ, USA, 2017; Volume 2, pp. 173–180.
6. Covington, P.; Adams, J.; Sargin, E. Deep neural networks for youtube recommendations. In Proceedings of the 10th ACM
Conference on Recommender Systems, Boston, MA, USA, 19 September 2016; pp. 191–198.
7. Sun, K.; Wang, L.; Xu, B.; Zhao, W.; Teng, S.W.; Xia, F. Network representation learning: From traditional feature learning to deep
learning. IEEE Access 2020, 8, 205600–205617. [CrossRef]
8. Han, S.; Qiao, Y.; Zhang, Y.; Lin, W.; Yang, J. Analyze users’ online shopping behavior using interconnected online interest-product
network. In Proceedings of the 2018 IEEE Wireless Communications and Networking Conference (WCNC), Barcelona, Spain,
15–18 April 2018; IEEE: Piscataway, NJ, USA, 2018; pp. 1–6.
9. Zhang, Y.; Qian, Y.; Gan, M.; Tang, X.; Lin, Z. Service Recommendation Based on User Dynamic Preference Extraction and
Prediction. In Proceedings of the 2019 IEEE World Congress on Services (SERVICES), Milan, Italy, 8–13 July 2019; IEEE: Piscataway,
NJ, USA, 2019; Volume 2642, pp. 121–126.
10. Skarding, J.; Gabrys, B.; Musial, K. Foundations and modeling of dynamic networks using dynamic graph neural networks: A
survey. IEEE Access 2021, 9, 79143–79168. [CrossRef]
11. Ouyang, Y.; Guo, B.; Wang, Q.; Yu, Z. Cross-domain recommendation with cross-graph knowledge transfer network. In
Proceedings of the ICC 2021-IEEE International Conference on Communications, Montreal, QC, Canada, 14–23 June 2021; IEEE:
Piscataway, NJ, USA, 2021; pp. 1–6.
12. Jiang, M.; Cui, P.; Wang, F.; Zhu, W.; Yang, S. Scalable recommendation with social contextual information. IEEE Trans. Knowl.
Data Eng. 2014, 26, 2789–2802. [CrossRef]
13. Guo, G. Resolving data sparsity and cold start in recommender systems. In Proceedings of the User Modeling, Adaptation, and
Personalization: 20th International Conference, UMAP 2012, Montreal, QC, Canada, 16–20 July 2012; Proceedings 20. Springer:
Berlin/Heidelberg, Germany, 2012; pp. 361–364.
14. Tang, J.; Hu, X.; Liu, H. Social recommendation: A review. Soc. Netw. Anal. Min. 2013, 3, 1113–1133. [CrossRef]
Appl. Sci. 2023, 13, 11378 26 of 31

15. Wang, H.; Kou, G.; Peng, Y. An iterative algorithm to derive priority from large-scale sparse pairwise comparison matrix. IEEE
Trans. Syst. Man Cybern. Syst. 2021, 52, 3038–3051. [CrossRef]
16. Zhang, S.; Yao, L.; Sun, A.; Tay, Y. Deep learning based recommender system: A survey and new perspectives. ACM Comput.
Surv. (CSUR) 2019, 52, 1–38. [CrossRef]
17. Yang, Z.; Ding, M.; Zou, X.; Tang, J.; Xu, B.; Zhou, C.; Yang, H. Region or global a principle for negative sampling in graph-based
recommendation. IEEE Trans. Knowl. Data Eng. 2022, 35, 6264–6277. [CrossRef]
18. Li, H.; Zhang, H.; Wang, S.; Hassan, A.E. Studying the Practices of Logging Exception Stack Traces in Open-Source Software
Projects. IEEE Trans. Softw. Eng. 2021, 48, 4907–4924. [CrossRef]
19. Gantner, Z.; Drumond, L.; Freudenthaler, C.; Rendle, S.; Schmidt-Thieme, L. Learning attribute-to-feature mappings for cold-start
recommendations. In Proceedings of the 2010 IEEE International Conference on Data Mining, Sydney, NSW, Australia, 13–17
December 2010; IEEE: Piscataway, NJ, USA, 2010; pp. 176–185.
20. Ma, H.; Yang, H.; Lyu, M.R.; King, I. Sorec: Social recommendation using probabilistic matrix factorization. In Proceedings of the
17th ACM Conference on Information and Knowledge Man-Agement, Napa Valley, CA, USA, 26–30 October 2008; pp. 931–940.
21. Mnih, A.; Salakhutdinov, R.R. Probabilistic matrix factorization. In Advances in Neural Information Processing Systems; 2007; Volume
20, Available online: https://proceedings.neurips.cc/paper/2007 (accessed on 18 September 2023).
22. Jamali, M.; Ester, M. A matrix factorization technique with trust propagation for recommendation in social networks. In
Proceedings of the Fourth ACM Conference on Recommender Systems, Barcelona, Spain, 26–30 September 2010; pp. 135–142.
23. Ma, H.; Zhou, D.; Liu, C.; Lyu, M.R.; King, I. Recommender systems with social regularization. In Proceedings of the Fourth
ACM International Conference on Web Search and Data Mining, Hong Kong, China, 9–12 February 2011; pp. 287–296.
24. Sedhain, S.; Menon, A.K.; Sanner, S.; Xie, L. Autorec: Autoencoders meet collaborative filtering. In Proceedings of the 24th
Intermnational Conference on World Wide Web, Florence, Italy, 18–22 May 2015; pp. 111–112.
25. Devooght, R.; Bersini, H. Long and short-term recommendations with recurrent neural networks. In Proceedings of the 25th
Conference on User Modeling, Adaptation and Personalization, Bratislava, Slovakia, 9–12 July 2017; pp. 13–21.
26. Wu, S.; Sun, F.; Zhang, W.; Xie, X.; Cui, B. Graph neural networks in recommender systems: A survey. arXiv 2020, arXiv:2011.02260.
[CrossRef]
27. Wang, X.; He, X.; Wang, M.; Feng, F.; Chua, T.S. Neural graph collaborative filtering. In Proceedings of the 42nd International
ACM SIGIR Conference on Research and Development in Information Retrieval, Paris, France, 21–25 July 2019; pp. 165–174.
28. Hamilton, W.; Ying, Z.; Leskovec, J. Inductive representation learning on large graphs. arXiv 2017, arXiv:1706.02216.
29. Ying, R.; He, R.; Chen, K.; Eksombatchai, P.; Hamilton, W.L.; Leskovec, J. Graph convolutional neural networks for web-scale
recommender systems. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data
Mining, London, UK, 19–23 August 2018; pp. 974–983.
30. He, X.; Liao, L.; Zhang, H.; Nie, L.; Hu, X.; Chua, T.S. Neural collaborative filtering. In Proceedings of the 26th International
Conference on World Wide Web, Perth, Australia, 3–7 April 2017; pp. 173–182.
31. Guo, H.; Tang, R.; Ye, Y.; Li, Z.; He, X. DeepFM: A factorization-machine based neural network for CTR prediction. arXiv 2017,
arXiv:1703.04247.
32. Wu, Y.; Su, L.; Wu, L.; Xiong, W. FedDeepFM: A Factorization Machine-Based Neural Network for Recommendation in Federated
Learning. IEEE Access 2023, 11, 74182–74190. [CrossRef]
33. Rendle, S.; Freudenthaler, C.; Schmidt-Thieme, L. Factorizing personalized markov chains for next-basket recommendation. In
Proceedings of the 19th International Conference on World Wide Web, Raleigh, NC, USA, 26–30 April 2010; pp. 811–820.
34. Kang, W.C.; McAuley, J. Self-attentive sequential recommendation. In Proceedings of the 2018 IEEE International Conference on
Data Mining (ICDM), Singapore, 17–20 November 2018; IEEE: Piscataway, NJ, USA, 2018; pp. 197–206.
35. Wang, K.; Wang, X.; Lu, X. POI recommendation method using LSTM-attention in LBSN considering privacy protection. Complex
Intell. Syst. 2023, 9, 2801–2812. [CrossRef]
36. Lv, F.; Jin, T.; Yu, C.; Sun, F.; Lin, Q.; Yang, K.; Ng, W. SDM: Sequential Deep Matching Model for Online Large-scale Recommender
System. In Proceedings of the CIKM ’19: The 28th ACM International Conference on Information and Knowledge Management,
Beijing, China, 3–7 November 2019.
37. Hidasi, B.; Karatzoglou, A.; Baltrunas, L.; Tikk, D. Session-based recommendations with recurrent neural networks. arXiv 2015,
arXiv:1511.06939.
38. Natarajan, S.; Vairavasundaram, S.; Kotecha, K.; Indragandhi, V.; Palani, S.; Saini, J.R.; Ravi, L. CD-SemMF: Cross-Domain
Semantic Relatedness Based Matrix Factorization Model Enabled With Linked Open Data for User Cold Start Issue. IEEE Access
2022, 10, 52955–52970. [CrossRef]
39. Zhu, F.; Chen, C.; Wang, Y.; Liu, G.; Zheng, X. Dtcdr: A framework for dual-target cross-domain recommendation. In Proceedings
of the 28th ACM International Conference on Information and Knowledge Management, Beijing, China, 3–7 November 2019; pp.
1533–1542.
40. Li, P.; Tuzhilin, A. Ddtcdr: Deep dual transfer cross domain recommendation. In Proceedings of the 13th International Conference
on Web Search and Data Mining, Houston, TX, USA, 3–7 February 2020; pp. 331–339.
41. Krishnan, A.; Das, M.; Bendre, M.; Yang, H.; Sundaram, H. Transfer learning via contextual invariants for one-to-many cross-
domain recommendation. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in
Information Retrieval, Virtual Event, China, 25–30 July 2020; pp. 1081–1090.
Appl. Sci. 2023, 13, 11378 27 of 31

42. He, X.; Deng, K.; Wang, X.; Li, Y.; Zhang, Y.; Wang, M. 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, Virtual Event, China, 25–30 July 2020; pp. 639–648.
43. Chen, L.; Wu, L.; Hong, R.; Zhang, K.; Wang, M. Revisiting graph based collaborative filtering: A linear residual graph
convolutional network approach. Proc. AAAI Conf. Artif. Intell. 2020, 34, 27–34. [CrossRef]
44. Ahmed, A.; Shervashidze, N.; Narayanamurthy, S. Distributed large-scale natural graph factorization. In Proceedings of the 22nd
International Conference on World Wide Web, Rio de Janeiro, Brazil, 13–17 May 2013; pp. 37–48.
45. Li, Z.; Wu, S.; Cui, Z.; Zhang, X. GraphFM: Graph factorization machines for feature interaction modeling. arXiv 2021,
arXiv:2105.11866.
46. Wu, L.; Li, J.; Sun, P.; Hong, R.; Ge, Y.; Wang, M. Diffnet++: A neural influence and interest diffusion network for social
recommendation. IEEE Trans. Knowl. Data Eng. 2020, 34, 4753–4766. [CrossRef]
47. Goldberg, D.; Nichols, D.; Oki, B.M.; Terry, D. Using collaborative filtering to weave an information tapestry. Commun. ACM 1992,
35, 61–70. [CrossRef]
48. Wu, C.Y.; Ahmed, A.; Beutel, A.; Smola, A.J.; Jing, H. Recurrent recommender networks. In Proceedings of the Tenth ACM
International Conference on Web Search and Data Mining, Cambridge, UK, 6–10 February 2017; pp. 495–503.
49. Kamishima, T.; Akaho, S.; Asoh, H.; Sakuma, J. Efficiency Improvement of Neutrality-Enhanced Recommendation. In Proceedings
of the Decisions@ RecSys, Hong Kong, China, 12–16 October 2013; pp. 1–8.
50. Rendle, S. Factorization machines. In Proceedings of the 2010 IEEE International Conference on Data Mining, Sydney, NSW,
Australia, 13–17 December 2010; IEEE: Piscataway, NJ, USA, 2010; pp. 995–1000.
51. Zhou, G.; Zhu, X.; Song, C.; Fan, Y.; Zhu, H.; Ma, X.; Yan, Y.; Jin, J.; Li, H.; Gai, K. Deep interest network for click-through rate
prediction. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, London,
UK, 19–23 August 2018; pp. 1059–1068.
52. Zhou, G.; Mou, N.; Fan, Y.; Pi, Q.; Bian, W.; Zhou, C.; Zhu, X.; Gai, K. Deep interest evolution network for click-through rate
prediction. Proc. AAAI Conf. Artif. Intell. 2019, 33, 5941–5948. [CrossRef]
53. Li, C.; Liu, Z.; Wu, M.; Xu, Y.; Zhao, H.; Huang, P.; Kang, G.; Chen, Q.; Li, W.; Lee, D.L. Multi-interest network with dynamic
routing for recommendation at Tmall. In Proceedings of the 28th ACM International Conference on Information and Knowledge
Management, Beijing, China, 3–7 November 2019; pp. 2615–2623.
54. Marcuzzo, M.; Zangari, A.; Albarelli, A.; Gasparetto, A. Recommendation systems: An insight into current development and
future research challenges. IEEE Access 2022, 10, 86578–86623. [CrossRef]
55. De Gemmis, M.; Lops, P.; Semeraro, G.; Basile, P. Integrating tags in a semantic content-based recommender. In Proceedings of
the 2008 ACM Conference on Recommender Systems, Lausanne, Switzerland, 23–25 October 2008; pp. 163–170.
56. Mooney, R.J.; Roy, L. Content-based book recommending using learning for text categorization. In Proceedings of the Fifth ACM
Conference on Digital Libraries, San Antonio, TX, USA, 2–7 June 2000; pp. 195–204.
57. Waila, P.; Singh, V.; Singh, M. A Scientometric Analysis of Research in Recommender Systems. J. Scientometr. Res. 2016, 4, 71–84.
[CrossRef]
58. Wang, H.; Czerminski, R.; Jamieson, A.C. Neural networks and deep learning. In The Machine Age of Customer Insight; Emerald
Publishing Limited: Bingley, UK, 2021; pp. 91–101.
59. Lu, Y.T.; Yu, S.I.; Chang, T.C.; Hsu, J.Y.J. A content-based method to enhance tag recommendation. In Proceedings of the
Twenty-First International Joint Conference on Artificial Intelligence, Pasadena, CA, USA, 11–17 July 2009.
60. Peng, J.; Zeng, D.D.; Zhao, H.; Wang, F.Y. Collaborative filtering in social tagging systems based on joint item-tag recommendations.
In Proceedings of the 19th ACM International Conference on Information and Knowledge Management, Toronto, ON, Canada,
26–30 October 2010; pp. 809–818.
61. Vasile, F.; Smirnova, E.; Conneau, A. Meta-prod2vec: Product embeddings using side-information for recommendation. In
Proceedings of the 10th ACM Conference on Recommender Systems, Boston, MA, USA, 15–19 September 2016; pp. 225–232.
62. Wu, L.; He, X.; Wang, X.; Zhang, K.; Wang, M. A survey on accuracy-oriented neural recommendation: From collaborative
filtering to information-rich recommendation. IEEE Trans. Knowl. Data Eng. 2022, 35, 4425–4445. [CrossRef]
63. LeCun, Y.; Bengio, Y.; Hinton, G. Deep learning. Nature 2015, 521, 436–444. [CrossRef]
64. Xu, M.; Liu, F.; Xu, W. A survey on sequential recommendation. In Proceedings of the 2019 6th International Conference on
Information Science and Control Engineering (ICISCE), Shanghai, China, 20–22 December 2019; IEEE: Piscataway, NJ, USA, 2019;
pp. 106–111.
65. Yang, Y.; Xiang, Y.; Xiong, L. Collaborative filtering and recommendation algorithm based on matrix factorization and user
nearest neighbor model. J. Comput. Appl. 2012, 32, 395. [CrossRef]
66. Wang, J.F.; Liu, R.D.; Liu, Y. Non-negative matrix factorization algorithm with bias in recommender system. J. Chin. Comput. Syst.
2018, 39, 69–73.
67. Yoon, J.H.; Jang, B. Evolution of Deep Learning-Based Sequential Recommender Systems: From Current Trends to New
Perspectives. IEEE Access 2023, 11, 54265–54279. [CrossRef]
68. Park, K.; Lee, J.; Choi, J. Deep neural networks for news recommendations. In Proceedings of the 2017 ACM on Conference on
Information and Knowledge Management, Singapore, 6–10 November 2017; pp. 2255–2258.
Appl. Sci. 2023, 13, 11378 28 of 31

69. Rendle, S.; Freudenthaler, C.; Gantner, Z.; Schmidt-Thieme, L. BPR: Bayesian personalized ranking from implicit feedback. arXiv
2012, arXiv:1205.2618.
70. Kurgan, L.A.; Musilek, P. A survey of knowledge discovery and data mining process models. Knowl. Eng. Rev. 2006, 21, 1–24.
[CrossRef]
71. Villatel, K.; Smirnova, E.; Mary, J.; Preux, P. Recurrent neural networks for long and short-term sequential recommendation. arXiv
2018, arXiv:1807.09142.
72. Thaipisutikul, T.; Shih, T.K.; Enkhbat, A.; Aditya, W. Exploiting long-and short-term preferences for deep context-aware
recommendations. IEEE Trans. Comput. Soc. Syst. 2021, 9, 1237–1248. [CrossRef]
73. Fattah, S.M.M.; Bouguettaya, A.; Mistry, S. Long-term IaaS selection using performance discovery. IEEE Trans. Serv. Comput. 2020,
15, 2129–2143. [CrossRef]
74. Zheng, C.; Tao, D.; Wang, J.; Cui, L.; Ruan, W.; Yu, S. Memory augmented hierarchical attention network for next point-of-interest
recommendation. IEEE Trans. Comput. Soc. Syst. 2020, 8, 489–499. [CrossRef]
75. Ying, H.; Zhuang, F.; Zhang, F.; Liu, Y.; Xu, G.; Xie, X.; Xiong, H.; Wu, J. Sequential recommender system based on hierarchical
attention network. In Proceedings of the IJCAI International Joint Conference on Artificial Intelligence, Vienna, Austria, 23–29
July 2018.
76. Gers, F.A.; Schmidhuber, J.; Cummins, F. Learning to forget: Continual prediction with LSTM. Neural Comput. 2000, 12, 2451–2471.
[CrossRef]
77. Bahdanau, D.; Cho, K.; Bengio, Y. Neural machine translation by jointly learning to align and translate. arXiv 2014, arXiv:1409.0473.
78. Li, J.; Ren, P.; Chen, Z.; Ren, Z.; Lian, T.; Ma, J. Neural attentive session-based recommendation. In Proceedings of the 2017 ACM
on Conference on Information and Knowledge Management, Singapore, 6–10 November 2017; pp. 1419–1428.
79. Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A.N.; Kaiser, Ł.; Polosukhin, I. Attention is all you need. In
Advances in Neural Information Processing Systems; 2017; Volume 30, Available online: https://proceedings.neurips.cc/paper/2017
(accessed on 18 September 2023).
80. Zhang, S.; Tay, Y.; Yao, L.; Sun, A.; An, J. Next item recommendation with self-attentive metric learning. In Proceedings of the
Thirty-Third AAAI Conference on Artificial Intelligence, Honolulu, HI, USA, 27 January–1 February 2019; Volume 9.
81. Cho, K.; Van Merrienboer, B.; Bahdanau, D.; Bengio, Y. On the properties of neural machine translation: Encoder-decoder
approaches. arXiv 2014, arXiv:1409.1259.
82. Tsai, Y.H.H.; Bai, S.; Liang, P.P.; Kolter, J.Z.; Morency, L.P.; Salakhutdinov, R. Multimodal transformer for unaligned multimodal
language sequences. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, Florence, Italy,
28 July–2 August 2019; Volume 2019, p. 6558.
83. Azhir, E.; Hosseinzadeh, M.; Khan, F.; Mosavi, A. Performance Evaluation of Query Plan Recommendation with Apache Hadoop
and Apache Spark. Mathematics 2022, 10, 3517. [CrossRef]
84. Singer, U.; Roitman, H.; Eshel, Y.; Nus, A.; Guy, I.; Levi, O.; Hasson, I.; Kiperwasser, E. Sequential modeling with multiple
attributes for watchlist recommendation in e-commerce. In Proceedings of the Fifteenth ACM International Conference on Web
Search and Data Mining, Tempe, AZ, USA, 21–25 February 2022; pp. 937–946.
85. Tuan, T.X.; Phuong, T.M. 3D convolutional networks for session-based recommendation with content features. In Proceedings of
the Eleventh ACM Conference on Recommender Systems, Como, Italy, 27–31 August 2017; pp. 138–146.
86. Yan, A.; Cheng, S.; Kang, W.C.; Wan, M.; McAuley, J. CosRec: 2D convolutional neural networks for sequential recommendation.
In Proceedings of the 28th ACM International Conference on Information and Knowledge Management, Beijing, China, 3–7
November 2019; ACM: New York, NY, USA, 2019; pp. 2173–2176.
87. Tang, J.X.; Wang, K. Personalized top-n sequential recommendation via convolutional sequence embedding. In Proceedings of
the Eleventh ACM International Conference on Web Search and Data Mining. Association for Computing Machinery, Marina Del
Rey, CA, USA, 5–9 February 2018; pp. 565–573.
88. Tan, Y.K.; Xu, X.; Liu, Y. Improved recurrent neural networks for session-based recommendations. In Proceedings of the 1st
Workshop on Deep Learning for Recommender Systems, Boston, MA, USA, 15 September 2016; pp. 17–22.
89. Bogina, V.; Kuflik, T. Incorporating Dwell Time in Session-Based Recommendations with Recurrent Neural Networks. In
Proceedings of the RecTemp@ RecSys, Como, Italy, 27–31 August 2017; pp. 57–59.
90. Hidasi, B.; Karatzoglou, A. Recurrent neural networks with top-k gains for session-based recommendations. In Proceedings of the
27th ACM International Conference on Information and Knowledge Management, Torino, Italy, 22–26 October 2018; pp. 843–852.
91. Hidasi, B.; Quadrana, M.; Karatzoglou, A.; Tikk, D. Parallel recurrent neural network architectures for feature-rich session-based
recommendations. In Proceedings of the 10th ACM Conference on Recommender Systems, Boston, MA, USA, 15–19 September
2016; pp. 241–248.
92. Jannach, D.; Ludewig, M. When recurrent neural networks meet the neighborhood for session-based recommendation. In
Proceedings of the Eleventh ACM Conference on Recommender Systems, Como, Italy, 27–31 August1 2017; pp. 306–310.
93. Chen, X.; Xu, H.; Zhang, Y.; Tang, J.; Cao, Y.; Qin, Z.; Zha, H. Sequential recommendation with user memory networks. In
Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining, Marina Del Rey, CA, USA, 5–9
February 2018; pp. 108–116.
Appl. Sci. 2023, 13, 11378 29 of 31

94. Sun, F.; Liu, J.; Wu, J.; Pei, C.; Lin, X.; Ou, W.; Jiang, P. BERT4Rec: Sequential recommendation with bidirectional encoder
representations from transformer. In Proceedings of the 28th ACM International Conference on Information and Knowledge
Management, Beijing, China, 3–7 November 2019; pp. 1441–1450.
95. Zhou, W.; Liu, Y.; Li, M.; Wang, Y.; Shen, Z.; Feng, L.; Zhu, Z. Dynamic Multi-Objective Optimization Framework with Interactive
Evolution for Sequential Recommendation. IEEE Trans. Emerg. Top. Comput. Intell. 2023, 7, 1228–1241. [CrossRef]
96. Cen, Y.; Zhang, J.; Zou, X.; Zhou, C.; Yang, H.; Tang, J. Controllable multi-interest framework for recommendation. In Proceedings
of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, Virtual Event, CA, USA, 6–10 July
2020; pp. 2942–2951.
97. Meng, W.; Yang, D.; Xiao, Y. 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, Virtual Event, China, 25–30 July 2020; pp. 1091–1100.
98. Yin, D.; Feng, S. Enhanced Attention Framework for Multi-Interest Sequential Recommendation. IEEE Access 2022, 10, 67703–
67712. [CrossRef]
99. Huang, X.; Qian, S.; Fang, Q.; Sang, J.; Xu, C. Csan: Contextual self-attention network for user sequential recommendation.
In Proceedings of the 26th ACM International Conference on Multimedia, Seoul, Republic of Korea, 22–26 October 2018; pp.
447–455.
100. Hoppe, A.; Nicolle, C.; Roxin, A. Automatic ontology-based user profile learning from heterogeneous web resources in a big data
context. Proc. VLDB Endow. 2013, 6, 1428–1433. [CrossRef]
101. Gao, C.; He, X.; Gan, D.; Chen, X.; Feng, F.; Li, Y.; Chua, T.-S.; Yao, L.; Song, Y.; Jin, D. Learning to recommend with multiple
cascading behaviors. IEEE Trans. Knowl. Data Eng. 2019, 33, 2588–2601. [CrossRef]
102. Le, D.T.; Lauw, H.W.; Fang, Y. Modeling contemporaneous basket sequences with twin networks for next-item recommendation.
In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence (IJCAI-18), Stockholm, Sweden,
13–19 July 2018.
103. Li, Z.; Zhao, H.; Liu, Q.; Huang, Z.; Mei, T.; Chen, E. Learning from history and present: Next-item recommendation via
discriminatively exploiting user behaviors. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge
Discovery & Data Mining, London, UK, 19–23 August 2018; pp. 1734–1743.
104. Zhou, C.; Bai, J.; Song, J.; Liu, X.; Zhao, Z.; Chen, X.; Gao, J. Atrank: An attention-based user behavior modeling framework for
recommendation. In Proceedings of the AAAI Conference on Artificial Intelligence, New Orleans, LA, USA, 2–7 February 2018;
Volume 32.
105. Fan, Z.; Liu, Z.; Zhang, J.; Xiong, Y.; Zheng, L.; Yu, P.S. Continuous-time sequential recommendation with temporal graph
collaborative transformer. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management,
Gold Coast, QLD, Australia, 1–5 November 2021; pp. 433–442.
106. Liu, Z.; Chen, Y.; Li, J.; Yu, P.S.; McAuley, J.; Xiong, C. Contrastive self-supervised sequential recommendation with robust
augmentation. arXiv 2021, arXiv:2108.06479.
107. Memmel, C. What drives the short-term fluctuations of banks’ exposure to interest rate risk? Rev. Financ. Econ. 2020, 38, 674–686.
[CrossRef]
108. Luo, H.; Yang, N.; Philip, S.Y. Hybrid deep embedding for recommendations with dynamic aspect-level explanations. In
Proceedings of the 2019 IEEE International Conference on Big Data (Big Data), Los Angeles, CA, USA, 9–12 December 2019; IEEE:
Piscataway, NJ, USA, 2019; pp. 870–879.
109. Choe, B.; Kang, T.; Jung, K. Recommendation system with hierarchical recurrent neural network for long-term time series. IEEE
Access 2021, 9, 72033–72039. [CrossRef]
110. Wang, S.; Cao, L.; Wang, Y.; Sheng, Q.Z.; Orgun, M.A.; Lian, D. A survey on session-based recommender systems. ACM Comput.
Surv. (CSUR) 2021, 54, 1–38. [CrossRef]
111. Elkahky, A.M.; Song, Y.; He, X. A multi-view deep learning approach for cross domain user modeling in recommendation systems.
In Proceedings of the 24th International Conference on World Wide Web, Florence, Italy, 18–22 May 2015; pp. 278–288.
112. Man, T.; Shen, H.; Jin, X.; Cheng, X. Cross-domain recommendation: An embedding and mapping approach. Int. Jt. Conf. Artif.
Intell. 2017, 17, 2464–2470.
113. Wang, C.-D.; Chen, Y.-H.; Xi, W.-D.; Huang, L.; Xie, G. Cross-Domain Explicit–Implicit-Mixed Collaborative Filtering Neural
Network. IEEE Trans. Syst. Man Cybern. Syst. 2021, 52, 6983–6997. [CrossRef]
114. Li, B. Cross-Domain Collaborative Filtering: A Brief Survey. In Proceedings of the 2011 IEEE 23rd International Conference on
Tools with Artificial Intelligence, Boca Raton, FL, USA, 7–9 November 2111; IEEE: Piscataway, NJ, USA, 2011; pp. 1085–1086.
115. Gupta, A.; Budania, H.; Singh, P.; Singh, P.K. Facebook based choice filtering. In Proceedings of the 2017 IEEE 7th International
Advance Computing Conference (IACC), Hyderabad, India, 5–7 January 2017; IEEE: Piscataway, NJ, USA, 2017; pp. 875–879.
116. Chen, H.; Wang, X.; Xie, R.; Zhou, Y.; Zhu, W. Cross-domain Recommendation with Behavioral Importance Perception. Proc.
ACM Web Conf. 2023, 2023, 1294–1304.
117. Zhu, Y.; Tang, Z.; Liu, Y.; Zhuang, F.; Xie, R.; Zhang, X.; Lin, L.; He, Q. Personalized transfer of user preferences for cross-domain
recommendation. In Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining, Tempe, AZ,
USA, 21–25 February 2022; pp. 1507–1515.
Appl. Sci. 2023, 13, 11378 30 of 31

118. Vodungbo, B.; Gautier, J.; Lambert, G.; Sardinha, A.B.; Lozano, M.; Sebban, S.; Ducousso, M.; Boutu, W.; Li, K.; Tudu, B.; et al.
Laser-induced ultrafast demagnetization in the presence of a nanoscale magnetic domain network. Nat. Commun. 2012, 3, 999.
[CrossRef]
119. Shapira, B.; Rokach, L.; Freilikhman, S. Facebook single and cross domain data for recommendation systems. User Model.
User-Adapt. Interact. 2013, 23, 211–247. [CrossRef]
120. Zhou, K.; Wang, H.; Zhao, W.X.; Zhu, Y.; Wang, S.; Zhang, F.; Wang, Z.; Wen, J.R. S3-rec: Self-supervised learning for sequen-
tial recommendation with mutual information maximization. In Proceedings of the 29th ACM International Conference on
Information & Knowledge Management, Virtual Event, Ireland, 19–23 October 2020; pp. 1893–1902.
121. Tiroshi, A.; Kuflik, T. Domain ranking for cross domain collaborative filtering. In Proceedings of the User Modeling, Adaptation,
and Personalization: 20th International Conference, UMAP 2012, Montreal, QC, Canada, 16–20 July 2012; Proceedings 20.
Springer: Berlin/Heidelberg, Germany, 2012; pp. 328–333.
122. Jiang, M.; Cui, P.; Yuan, N.J.; Xie, X.; Yang, S. Little is much: Bridging cross-platform behaviors through overlapped crowds. In
Proceedings of the AAAI Conference on Artificial Intelligence, Phoenix, AZ, USA, 12–17 February 2016; Volume 30.
123. Moreno, O.; Shapira, B.; Rokach, L.; Shani, G. Talmud: Transfer learning for multiple domains. In Proceedings of the 21st
ACM International Conference on Information and Knowledge Management, Maui, HI, USA, 29 October–2 November 2012; pp.
425–434.
124. Pan, W.; Yang, Q. Transfer learning in heterogeneous collaborative filtering domains. Artif. Intell. 2013, 197, 39–55. [CrossRef]
125. Zhang, Z.; Jin, X.; Li, L.; Ding, G.; Yang, Q. Multi-domain active learning for recommendation. In Proceedings of the AAAI
Conference on Artificial Intelligence, Phoenix, AZ, USA, 12–17 February 2016; Volume 30.
126. Zhu, F.; Wang, Y.; Chen, C.; Liu, G.; Zheng, X. A graphical and attentional framework for dual-target cross-domain recommenda-
tion. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence (IJCAI-20), Yokohama, Japan,
7–15 January 2020; pp. 3001–3008.
127. Liu, M.; Li, J.; Li, G.; Pan, P. Cross domain recommendation via bi-directional transfer graph collaborative filtering networks. In
Proceedings of the 29th ACM International Conference on Information & Knowledge Management, Virtual Event, Ireland, 19–23
October 2020; pp. 885–894.
128. Cui, Q.; Wei, T.; Zhang, Y.; Zhang, Q. HeroGRAPH: A Heterogeneous Graph Framework for Multi-Target Cross-Domain
Recommendation. In Proceedings of the ORSUM@ RecSys, Online, 25–26 September 2020.
129. Jamali, M.; Ester, M. Trustwalker: A random walk model for combining trust-based and item-based recommendation. In
Proceedings of the 15th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Paris, France, 28
June 2009–1 July 2009; pp. 397–406.
130. Wang, X.; Lu, W.; Ester, M.; Wang, C.; Chen, C. Social recommendation with strong and weak ties. In Proceedings of the 25th
ACM International on Conference on Information and Knowledge Management, Indianapolis, IN, USA, 24–28 October 2016; pp.
5–14.
131. Yang, X.; Steck, H.; Liu, Y. Circle-based recommendation in online social networks. In Proceedings of the 18th ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining, Beijing China, 12–16 August 2012; pp. 1267–1275.
132. Berg, R.; Kipf, T.N.; Welling, M. Graph convolutional matrix completion. arXiv 2017, arXiv:1706.02263.
133. Perozzi, B.; Al-Rfou, R.; Skiena, S. Deepwalk: Online learning of social representations. In Proceedings of the 20th ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining, New York, NY, USA, 24–27 August 2014; pp. 701–710.
134. Tang, J.; Qu, M.; Wang, M.; Zhang, M.; Yan, J.; Mei, Q. Line: Large-scale information network embedding. In Proceedings of the
24th International Conference on World Wide Web, Florence, Italy, 18–22 May 2015; pp. 1067–1077.
135. Xie, M.; Yin, H.; Wang, H.; Xu, F.; Chen, W.; Wang, S. Learning graph-based poi embedding for location-based recommendation.
In Proceedings of the 25th ACM International on Conference on Information and Knowledge Management, Indianapolis, IN,
USA, 24–28 October 2016; pp. 15–24.
136. Yin, H.; Zou, L.; Nguyen, Q.V.H.; Huang, Z.; Zhou, X. Joint event-partner recommendation in event-based social networks. In
Proceedings of the 2018 IEEE 34th International Conference on Data Engineering (ICDE), Paris, France, 16–19 April 2018; IEEE:
Piscataway, NJ, USA, 2018; pp. 929–940.
137. Tang, J.; Qu, M.; Mei, Q. Pte: Predictive text embedding through large-scale heterogeneous text networks. In Proceedings of the
21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Sydney, NSW, Australia, 10–13 August
2015; pp. 1165–1174.
138. Ma, R.; Qiu, X.; Zhang, Q. Co-attention memory network for multimodal microblog’s hashtag recommendation. IEEE Trans.
Knowl. Data Eng. 2019, 33, 388–400. [CrossRef]
139. Wang, Q.; Oàeilly-morgan, D.; Tragos, E.Z.; Hurley, N.; Smyth, B.; Lawlor, A.; Dong, R. Learning Domain-Independent
Representations via Shared Weight Auto-Encoder for Transfer Learning in Recommender Systems. IEEE Access 2022, 10, 71961–
71972. [CrossRef]
140. Ye, L.; Xie, H.; Lin, Y.; Lui, J.C. Rewarding Social Recommendation in OSNs: Empirical Evidences, Modeling and Optimization.
IEEE Trans. Knowl. Data Eng. 2020, 34, 4410–4424. [CrossRef]
141. Ye, Q.; Cao, Y.; Chen, Y. Deep Learning-Based User Privacy Settings Recommendation in Online Social Networks. In Proceedings
of the 2022 International Joint Conference on Neural Networks (IJCNN), Padua, Italy, 18–23 July 2022; IEEE: Piscataway, NJ, USA,
2022; pp. 1–9.
Appl. Sci. 2023, 13, 11378 31 of 31

142. Farmaki, A.; Olya, H.; Taheri, B. Unpacking the complex interactions among customers in online fan pages. J. Bus. Res. 2021, 125,
164–176. [CrossRef]
143. Velickovic, P.; Cucurull, G.; Casanova, A. Graph attention networks. Stat 2017, 1050, 10–48550.
144. Zhang, C.; Song, D.; Huang, C.; Swami, A.; Chawla, N.V. Heterogeneous graph neural network. In Proceedings of the 25th ACM
SIGKDD International Conference on Knowledge Discovery & Data Mining, Anchorage, AK, USA, 4–8 August 2019; pp. 793–803.
145. Wu, L.; Sun, P.; Fu, Y.; Hong, R.; Wang, X.; Wang, M. A neural influence diffusion model for social recommendation. In
Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval, Paris,
France, 21–25 July 2019; pp. 235–244.
146. Fan, W.; Ma, Y.; Li, Q.; He, Y.; Zhao, E.; Tang, J.; Yin, D. Graph neural networks for social recommendation. In Proceedings of the
The World Wide Web Conference, Lyon, France, 23–27 April 2019; pp. 417–426.
147. Abdollahpouri, H.; Adomavicius, G.; Burke, R.; Guy, I.; Jannach, D.; Kamishima, T.; Krasnodebski, J.; Pizzato, L. Multistakeholder
recommendation: Survey and research directions. User Model. User-Adapt. Interact. 2020, 30, 127–158. [CrossRef]
148. Guo, Z.; Wang, H. A deep graph neural network-based mechanism for social recommendations. IEEE Trans. Ind. Inform. 2020, 17,
2776–2783. [CrossRef]
149. Lina, L.F.; Setiyanto, A. Privacy concerns in personalized advertising effectiveness on social media. Sriwij. Int. J. Dyn. Econ. Bus.
2021, 5, 147–156. [CrossRef]
150. Koohang, A.; Sargent, C.S.; Nord, J.H.; Paliszkiewicz, J. Internet of Things (IoT): From awareness to continued use. Int. J. Inf.
Manag. 2022, 62, 102442. [CrossRef]
151. Zhong, W.; Yin, X.; Zhang, X.; Li, S.; Dou, W.; Wang, R.; Qi, L. Multi-dimensional quality-driven service recommendation with
privacy-preservation in mobile edge environment. Comput. Commun. 2020, 157, 116–123. [CrossRef]
152. Hasan, M.K.; Alkhalifah, A.; Islam, S.; Babiker, N.B.; Habib, A.A.; Aman, A.H.M.; Hossain, M.A. Blockchain technology on smart
grid, energy trading, and big data: Security issues, challenges, and recommendations. Wirel. Commun. Mob. Comput. 2022, 2022,
9065768. [CrossRef]
153. Zhang, J.; Askari, H.; Psounis, K.; Shafiq, Z. A Utility-Preserving Obfuscation Approach for YouTube Recommendations. Proc.
Priv. Enhancing Technol. 2023, 4, 522–539. [CrossRef]
154. Seaton, D.B.; Caspi, A.; Casini, R.; Downs, C.; Gibson, S.E.; Gilbert, H.; Glesener, L.; Guidoni, S.E.; Hughes, J.M.; McKenzie, D.;
et al. Improving Multi-Dimensional Data Formats, Access, and Assimilation Tools for the Twenty-First Century. arXiv 2023,
arXiv:2305.16535.
155. Hallifax, D.; Houston, J.B. Binding of drugs to hepatic microsomes: Comment and assessment of current prediction methodology
with recommendation for improvement. Drug Metab. Dispos. 2006, 34, 724–726. [CrossRef]
156. Castells, P.; Hurley, N.; Vargas, S. Novelty and diversity in recommender systems. In Recommender Systems Handbook; Springer:
New York, NY, USA, 2021; pp. 603–646.

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.