Information Sciences 589 (2022) 595–607

Contents lists available at ScienceDirect

Information Sciences
journal homepage: www.elsevier.com/locate/ins

SocialLGN: Light graph convolution network for social

recommendation
Jie Liao a, Wei Zhou a, Fengji Luo b, Junhao Wen a,⇑, Min Gao a, Xiuhua Li a, Jun Zeng a
a
School of Big Data and Software Engineering, Chongqing University, 401331 Chongqing, China
b
School of Civil Engineering, The University of Sydney, Sydney, Australia

a r t i c l e i n f o a b s t r a c t

Article history: Graph Neural Networks have been applied in recommender systems to learn the represen-
Received 15 July 2021 tation of users and items from a user-item graph. In the state-of-the-art, there are two
Received in revised form 30 December 2021 major challenges in applying Graph Neural Networks to social recommendation: (i) how
Accepted 2 January 2022
to accurately learn the representation of users and items from the user-item interaction
Available online 6 January 2022
graph and social graph, and (ii) based on the fact that each user is represented simultane-
ously by the two graphs, how to integrate the user representations learned from these two
Keywords:
graphs. Aiming at addressing these challenges, this paper proposes a new social recom-
Social recommendation
Graph convolution network
mendation system called SocialLGN. In SocialLGN, the representation of each user and item
Recommender system is propagated in the user-item interaction graph with light graph convolutional layers; in
Embedding propagation the meantime, the user’s representation is propagated in the social graph. Based on this, a
graph fusion operation is designed to aggregate user representations during propagation.
The weighted sum is applied to combine the representations learned by each layer.
Comprehensive experiments are conducted on two real-world datasets, and the result
shows that the proposed SocialLGN outperforms the SOTA method, especially in handling
the cold-start problem. Our PyTorch implemented model is available via https://github.c
om/leo0481/SocialLGN.
Ó 2022 Elsevier Inc. All rights reserved.

1. Introduction

Recommender systems provide decision-making assistance to users to alleviate the information overload issue by pre-
dicting their preferences and interests on items. Learning low dimensional latent representations (known as embeddings)
of users and items is an important step in such a prediction process and has been extensively studied [28]. Traditional rec-
ommender systems usually lean the low dimensional latent representations sorely from the user-item interaction data; for a
recommender system, when the user-term interaction data is insufficient, the system’s representation ability would be sig-
nificantly affected. This is referred to as the data sparse problem [9]. References [3,23] and life experiences have shown that
people are easily influenced by friends, and people tend to make friends with other people with similar preferences. These
indicate that social relationships can provide information about the user’s potential interest and can be considered as useful
auxiliary information in recommender system designs [18] (the way of using auxiliary information to enhance the perfor-
mance of expert systems has been widely adopted. For example, reference [36] uses the click information to improve the

⇑ Corresponding author.
E-mail addresses: [email protected] (J. Liao), [email protected] (W. Zhou), [email protected] (F. Luo), [email protected] (J. Wen),
[email protected] (M. Gao), [email protected] (X. Li), [email protected] (J. Zeng).

https://doi.org/10.1016/j.ins.2022.01.001
0020-0255/Ó 2022 Elsevier Inc. All rights reserved.
J. Liao, W. Zhou, F. Luo et al. Information Sciences 589 (2022) 595–607

accuracy of image recognition). Based on this realization, recommender systems are developed by utilizing social informa-
tion to enhance the representations of users and items and mitigate the data sparse problem. These systems are referred to
as the social-based recommender system [31,37].
Both the user-item interaction data and social relationship data used in recommender systems are usually modeled as
two graphs, which are called the user-item interaction graph and the social graph, respectively. For the social platform envi-
ronment that is essentially user-centric, both the user-item interaction data and social relationship data contain the users’
information; as a result, the users’ information is included in both graphs. Generally, there are two approaches of using social
information in recommender systems [33]: One approach is to integrate the social graphs and user-item interaction graphs
into a heterogeneous graph and uniformly learn the representation of each node in the graph. Another approach is to learn
the representation of each user in the two graphs, and then merge the two learned representations into a final representation
(Fig. 1).
Due to its superior capability of modelling graph-structured data, Graph Neural Networks (GNNs) have been recently
applied to social recommendation systems [30,35] and other types of recommendation systems [11,29]. Compared with
non-graph based social recommendation methods, GNN-based social recommendation methods can improve the recom-
mendation accuracy. Graph convolution networks (GCNs) are a subset of GNN and they have been applied in recommender
systems to learn the representation of each node in the graph. Nevertheless, literature [11] has shown that two components
in GCN are redundant for recommendation problems and would even pose a negative impact on the recommendation’s accu-
racy. Based on this realization, a light GCN is developed [11], which is based on GCN but removes the two operations of linear
transformation and nonlinear activation in GCNs. The experiment reported in [11] shows that the light GCN has superior
performance than conventional GCNs in terms of several metrics that are commonly used in recommender systems.
This paper proposes a new social recommendation system based on a light graph convolution network, called ’SocialLGN’.
SocialLGN innovatively extends the user/item representation propagation mechanism in LightGCN to incorporate two graphs
(i.e., the user-item interaction graph and social graph). It propagates the users’ representations in both graphs; based on this,
a new graph fusion component is developed and integrated into SocialLGN to fuse the two user representations obtained
from the propagation into one integrated representation. The system also propagates the item representation in the user-
item interaction graph. By removing the operations of linear transformation and nonlinear activation in GCNs and enabling
representation propagation in the two graphs, the developed system can perform accurate social recommendation in a com-
putationally efficient manner and therefore alleviates the aforementioned limitations identified in the literature (e.g., [6,29]).
More specifically, the main contributions of this paper are summarized as follows:

(1) Propose a SocialLGN system, which extends the LightGCN to make it more adaptable to social recommendation prob-
lems. A light graph convolution network-based representation propagation mechanism is designed for the user-item
interaction graph and social graph simultaneously.
(2) Design a customized graph fusion component that fuses the two user representations learned from the two graphs
into a uniform representation. As an integrated part of the SocialLGN system, the component uses a normalization oper-
ation to avoid the increase of the user’s representation’s scale when the number of the graph convolution layers increases.

Fig. 1. Learning and fusion of user and item representations.

596
J. Liao, W. Zhou, F. Luo et al. Information Sciences 589 (2022) 595–607

(3) Conduct comprehensive experiments on two real-world datasets to validate the proposed system. Comparison study
is performed to demonstrate the effectiveness of SocialLGN by comparing it with several state-of-the-art recommenda-
tion methods, including DiffNet [31], NGCF [29] and LightGCN [11].

The rest of the paper is organized as follows. Section 2 gives an overview of the state-of-the-art. Section 3 presents the
general form of the social-based recommendation problem. Section 4 presents the design details of the proposed SocialLGN
system. Section 5 reports the experiments conducted for validating the proposed system. Section 6 concludes the paper.

2. Related work

This section provides an overview of the representative literature that is related to this work. The related work is
reviewed following the 2 categories as below.

2.1. Social recommendation

With the increasing prevalence of social platforms, it is recognized that social information can be utilized to improve the
performance of recommender systems [31,34]. Most social recommendation models rely on social effects of homophily [23]
and social influence [3] to tackle the data sparse problem and cold start problem that are common in recommendation sys-
tems. Based on the realization that the matrix factorization (MF) [16] technique can well capture social relationships, the
early designs of social recommender systems are mostly based on MF. These recommender systems can be generally divided
into 3 categories in terms of the way they utilize social information [27]: co-factorization methods [21], ensemble methods
[20], and regularization methods [22].
SBPR [39] is a pair-wise Bayesian Personalized Ranking (BPR) model that assumes that users tend to give higher ratings to
the items that their friends prefer. It defines a new class of items called ”social feedback” to represent the information in the
social network. TrustSVD [7,8] is developed to apply SVD++ [15], a MF-based model that has been widely used in recom-
mender systems, in social recommendation. TrustSVD utilizes social relationships in the way of treating the social neighbors’
implicit feedbacks as the user’s auxiliary implicit feedback. ContextMF [12] is a collective MF-based framework with addi-
tional regularization terms, in which the regularization terms are designed for combining Contextual signals and social net-
works. TrustMF [34] is an MF-based model, which factorizes social trust networks to map the users into two low-
dimensional latent feature spaces. In this way, it can capture the influence between the users. While the above work is based
on MF techniques, recent research develops deep neural networks [5] and graph representation [17] approaches to represent
the users and items by graphs. These approaches take advantage of the multi-layer structure of graph convolution neural
networks to better mine the nonlinear relationship between the users and items.

2.2. Applications of graph neural networks and graph convolution networks in social recommendation

The main principle of Graph Neural Networks is to iteratively aggregate information from a target node’s neighboring
nodes’ information and obtain the representation of each node in the graph. Some literature applies GNNs to social recom-
mendation problems. GraphRec [6] is the first research that introduces GNNs into social recommendation by fusing one-hop
neighboring nodes information in the user-item interaction graph and social graph to learn the users’ representation. DiffNet
[31] adds up the user representations from two graphs. For the user-item interaction graph, it performs a mean-pooling
operation on the one-hop neighboring nodes; for the social graph, it uses a GNN to propagate each user’s representation.
DiffNet++ [30] is an extension of DiffNet. It enhances the performance of the original DiffNet by designing a multi-level graph
attention network to simultaneously model the users’ interests on the items and the users’ influences with each other. Both
of DiffNet and DiffNet++ use layer-wise propagation structure to describe the dynamic diffusion process of social influence.
Reference [32] proposes a dual graph attention network to collaboratively learn the representations of twofold social effects.
DGRec [26] uses two recurrent neural networks to model the dynamic user behaviors and the social influence based on the
context.
GCNs is a kind of convolution network that encodes a graph’s structure directly and obtain the nodes’ representations in
the graph by propagating the nodes’ features in the graph. GCNs extend the convolution operation, which is traditionally
applied to the data with Euclidean structures, to the data with non-Euclidean structures. In recent years, GCNs have been
applied to graph representation learning (e.g., [14]) and have shown satisfactory performance. Early graph convolution
approaches [1,4] are defined in the spectral domain and are computationally expensive. Later studies [10,14] redefine the
graph convolution approach in the spatial domain and have gained better interpretability with lower computation cost;
these approaches now have been widely adopted and applied in different areas, such as human body skeleton modelling
[25], multi-label image recognition [2] and text classification [19]. GCNs have also been applied in social recommendation
[6,30], where they iteratively aggregate the embeddings of a target node’s neighboring nodes to refine the target node’s
embedding on the graph that represents the user-item interactions. Yu et al. [35] propose a GCN-based deep adversarial
framework to solve three existing problems in the social recommendation. Literature [38] proposes a multi-channel hyper-
graph convolutional network, which utilizes high-order user relations to enhance the recommendation accuracy.
597
J. Liao, W. Zhou, F. Luo et al. Information Sciences 589 (2022) 595–607

3. Overview of social recommendation problem

Recommendation problems [18,32] are essentially about analyzing a group of users’ interests and preferences on items.
Such preferences can be reflected in explicit measures, such as the explicit ratings the users give to the items (e.g., the star
ratings the users give to the movies). In this way, the relationship between the users and items can be represented as a rating
matrix. In most practical situations, a user only explicitly rates a small set of items; this makes the rating matrix usually very
sparse. Based on the fact that the explicit feedbacks (e.g., explicit ratings) are highly sparse, implicit feedbacks – for example,
the information that can reflect a user ever ordered a service or listened to a song, are used to model the users’ preferences on
the items.
A graph-based social recommendation problem can be expressed as follows. Let U and I denote the sets of N users and M
items, respectively, i.e., jUj ¼ N and jIj ¼ M. The users, items, and their relationships are usually represented by a user-item
interaction graph and a social graph. These two graphs are introduced below.
The user-item interaction graph represents the interactions between the users and items (e.g., clicks, views, and rates). A
user-item interaction graph (denoted as GI ) is defined as fðu; yui ; iju 2 U; i 2 IÞg, where U and I denote the user set and item
set, respectively. An edge yui > 0 indicates an observed interaction between the user u and item i, and yui ¼ 0 indicates there
is no interaction between them. The notation NIu denotes the items that have interactions with the user u in the user-item
interaction graph, and the nation NIi denotes the set of the users who have interactions with the item i.
The social graph represents the social connections between the users, which is usually reflected in the users’ auxiliary side
information (e.g., undirected friendship or direct follower/followee relations). A social graph (denoted as GS ) can be defined

 
as ui ; sij ; uj jui ; uj 2 U , where an edge sij ¼ 1 indicated that the users ui and uj have an observed social connection, while
sij ¼ 0 there is no social connection between them. The notation NSu is used to denote all the users that user u trust or follow
in the social graph.
Based on the above discussion, the input and output of a social recommendation task are formally presented as:
Input: a user-item interaction graph GI and a social graph GS .
Output: a function that predicts the value of the edge yui that connects user u and item i. This value is then used as the
ranking score for recommendation.

4. Design of SocialLGN

This section presents the design details of the proposed SocialLGN system.

4.1. System architecture

The overall architecture of the proposed system is illustrated in Fig. 2. The system consists of 3 main components: (i) an
embedding layer, which initializes the embeddings of the users and items from their identifiers (IDs); (ii) multiple propaga-
tion layers, which propagate the embeddings in the social graph and user-item interaction graph; and (iii) a prediction layer,
which predicts the value of any edge connecting two nodes based on the final representations of the users and items (the
final representations are obtained from the embedding propagation process in (ii)).
The detailed working mechanism of each component is presented as follows.

4.1.1. Embedding layer

ð0Þ ð0Þ
The embedding layer encodes a user u (or an item i) with a fixed-length vector eu 2 Rd (ei 2 Rd ). Such a vector is called
an embedding. Here d is a pre-specified parameter, denoting the size of an embedding. The superscript ”ðkÞ” is a layer index,
indicating that the embeddings are the output of the kth propagation layer (as later discussed in Section 4.1.2); the super-
script ”ð0Þ” is used to indicate the output of the embedding layer. An embedding can be initialized using the user’/item’s side
features, or it can be initialized following a probabilistic distribution. After being initialized in the embedding layer, the
embeddings will be iteratively optimized in the subsequential training process of the recommendation model (as later dis-
cussed in Section 4.2).
ð0Þ ð0Þ
The embeddings of all the N users can be represented as a matrix EU 2 RNd , where the user u’s embedding eu is the
ð0Þ ð0Þ
transpose of the uth row of EU . Similarly, the embeddings of all the M items can be represented as a matrix EI 2 RMd ,
ð0Þ ð0Þ
where the item i’s embedding ei is the transpose of the ith row of EI .

4.1.2. Propagation layers

Based on the users’ and items’ embeddings generated from the embedding layer, a light graph convolution operation is
performed to the users’ and items’ latent features on the user-item interaction graph and the social graph. As introduced in
Section 1, the light graph convolution operation simplifies the conventional GCN [14] by removing the linear transformation
and nonlinear activation operations.
598
J. Liao, W. Zhou, F. Luo et al. Information Sciences 589 (2022) 595–607

Fig. 2. Architecture design of SocialLGN.

In the kth propagation layer (k ¼ 1 : K, where K is the number of propagation layers), it takes the embeddings from the
ðk  1Þth propagation layer (the 0th propagation layer means the embedding layer) and outputs the updated embeddings.
The update processes of the user embedding and item embedding are introduced as follows.
For a user u, we use both the user-item interaction graph and social graph to update the user embedding outputted from
ðk1Þ
the ðk  1Þth layer (denoted as eu ). To do this, two embeddings are generated for the user u in the user-item interaction
ðkÞ ðkÞ
graph and social graph respectively, denoted as qu and pu :
X 1
qðukÞ ¼ qffiffiffiffiffiffiffiffiffiffiffiqffiffiffiffiffiffiffiffiffiffi eði k1Þ ; ð1Þ
i2NIu jNu j jNIi j
I

X 1
pðukÞ ¼ qffiffiffiffiffiffiffiffiffiffiffiqffiffiffiffiffiffiffiffiffiffiffi eðuk1Þ ; ð2Þ
v 2Nu jNu j jNv j
S S S

where NSu denotes the user u’s neighborhood set in the social graph; NIu denotes the user u’s neighborhood set in the user-
item interaction graph. Eqs. (1) and (2) essentially update the user u’s embedding by aggregating the embeddings of the
user’s neighboring nodes in the two graphs.
ðkÞ ðkÞ ðkÞ
qu and pu are then fused to generate an updated embedding eu . Such fusion process can be generally expressed as
 
ðkÞ ðkÞ ðkÞ
eu ¼ f qu ; pu , where f ðÞ represents the graph fusion operation. In this study, we propose the following graph fusion
ðkÞ
operation to generate eu :




~ðukÞ ¼W3
e r W1 pðukÞ jjr W2 qðukÞ ; ð3Þ
~ðukÞ
e
eðukÞ ¼ ðkÞ ; ð4Þ
~u jj2
jje

where rðÞ is the tanh activation function, W1 ; W2 2 Rdd ; W3 2 Rd2d are trainable matrices. The normalization operation Eq.
ðkÞ
(4) is performed to suppress the increase of the values in eu as increase of k.
ðk1Þ
For the item i’s embedding ei outputted from the ðk  1Þth layer, it is propagated in the user-item interaction graph as
follows:

599
J. Liao, W. Zhou, F. Luo et al. Information Sciences 589 (2022) 595–607

ðkÞ
X 1
ei ¼ qffiffiffiffiffiffiffiffiffiffiqffiffiffiffiffiffiffiffiffiffiffi eðuk1Þ ; ð5Þ
u2NIi jNi j jNIu j
I

where NIu and NIi denote the user u’s neighborhood set and item i’s neighborhood set in the user-item interaction graph,
respectively.

4.1.3. Prediction layer

n o
ð0Þ ðK Þ
After being propagated through the K layers, the embeddings of the user u and item are obtained, i.e., eu ; . . . ; eu and
n o
ð0Þ ðK Þ
ei ; . . . ; ei , respectively. Based on this, the final representations of the user and item are generated by summing up the
outputs of the K layers:

X
K
eu ¼ ak eðukÞ ð6Þ
k¼0
X
K
ei ¼ ak eði kÞ ð7Þ
k¼0

The raking score of the user u on item i is calculated by taking the inner product of their final representations:

^ui ¼ eTu ei
y ð8Þ

^ui is then used to optimize the parameters in the proposed recommendation model. Such an optimization process is called
y
model training, which is introduced as below.

4.2. Training of the recommendation model

h i
ð0Þ ð0Þ
In this study, we optimize the recommendation model’s parameters (Eð0Þ ¼ EU ; EI ; W1 ; W2 ,and W3 ). This is achieved
by minimizing the Bayesian Personalized Ranking (BPR) loss [24] (Eq. (9)). The BPR loss is based on the assumption that a
user-item interaction that has already existed in the dataset should be assigned a higher ranking score than an interaction
that does not exist in the dataset.
X

LBPR ¼  lnr y ^uj þ kjjEð0Þ jj22
^ui  y ð9Þ
ðu;iÞ2NIu [ðu;jÞRNIu

In Eq. (9), NIu represents the neighborhood set of the user u in the user-item interaction graph; Eð0Þ is the embeddings of the
N users and M items outputted from embedding layer.

4.3. Matrix formulation of SocialLGN

The proposed SocialLGN can also be expressed in a matrix form. To do this, we denote S 2 RNN as the adjacency matrix of
the social graph GS . The rating matrix is denoted as R 2 RNM and each element r ui (u ¼ 1 : N; m ¼ 1 : M) in R represents the
existence of the interaction between the user u and item i. Since in this study we focus on implicit feedbacks, the value of r ui
is binary: 1-there is an interaction; 0-no interaction.
e is a symmetrically normalized matrix and DR 2 RNN is a diagonal matrix, each entry dii
e ¼ D2 RD2 . R
1 1
We further define R R R
~
in DR represents the number of non-zero elements in the ith row of R. We also define RT ¼ DRT2 RT DRT2 and e
1 1 1 1
S ¼ DS 2 SDS 2 . Based
on these notations, the matrix form of the propagation operation in the kth propagation layer is formulated as:
h    i
e ðkÞ ¼ W3
E e ðk1Þ jjr W2 e
r W1 RE ðk1Þ
SEU
U I

ðkÞ eE ðUkÞ
EU ¼ ð10Þ
e ðkÞ
jj E U jj2

ðk Þ ~ ðk1Þ
EI ¼ R T EU

4.4. Discussion

This section analyzes the space complexity and time complexity of training the proposed SocialLGN.
600
J. Liao, W. Zhou, F. Luo et al. Information Sciences 589 (2022) 595–607

4.4.1. Space complexity

The trainable model parameters in SocialLGN consist of two parts (denoted as H1 and H2 ): (a) the users’ and items’
embeddings in the 0th layer use, i.e., H1 ¼ Eð0Þ ; and (b) the parameter set in the graph fusion operation (Eq. (3)), which
includes 3 matrices: H2 ¼ fW1 ; W2 ; W3 g. For part (a), since most of the embedding-based models (e.g., LightGCN [11]) need
to store the latent vectors of each node in the user-item interaction graph, the number of the trainable parameters in H1 is
the same as that in those models. This number increases linearly with the increase of the number of nodes in the user-item
interaction graph. For part (b), the three trainable matrices are shared among all nodes. Both W1 and W1 are d  d matrices
and W3 is a 2d  d matrix. Therefore, the total space storage needed for these trainable parameters approximates
2
ðN þ M þ 4dÞd. Since there is minðN; MÞ  d, the addition constant storage cost 4d can be neglected. This indicates the space
complexity of the proposed SocialLGN is reasonable.

4.4.2. Time complexity

The computation time cost of SocialLGN mainly consists of two parts: (a) the embedding propagation processes in both
the user-item interaction graph and social graph; and (b) the graph fusion operation in each propagation layer.
For the first part, denote the average numbers of the edges across all the N user in the social graph and in the user-item
interaction graph are Ls and Li , respectively; denote the average number of the edges across all the M items in the user-item
interaction graph is Lu . In each propagation layer, we apply the graph convolution operation in the 2 graphs separately (Eqs.
(1), (2), and (5)). The time cost for performing the graph convolution operations in one propagation layer is approximately
 
2
OðN ðLs þ Li Þd þ MLu dÞ. For the second part, the time cost for the graph fusion operation (Eq. (3)) is about O Nd .
Based on the above discussion, the total time cost in one propagation layer is OðN ðLs þ Li þ dÞd þ MLu dÞ. The time cost in
the K propagation layers thus is OðK ðN ðLs þ Li þ dÞd þ MLu dÞÞ. minðN; M Þ  K; Ls ; Li ; Lu , and d, the cost of the additional time
 
2
linear increases with the increase of the number of the users, i.e., O KNd . Since the linear transformation and nonlinear
activation operations are not included, SocialLGN is with lower time complexity than the existing GCN-based social recom-
mendation models.

5. Experiments

This section reports the experiments conducted for validating the proposed SocialLGN system as well as the key findings.
All the programs are executed on a computer with dual Intel Xeon E5-2678 v3 processors and an RTX 3090 GPU with 24-
Gigabyte memory. The SocialLGN is implemented using PyTorch.

5.1. Experimental settings

The main settings of the experiments in this study are reported as below.

5.1.1. Datasets
The experiments are conducted on two real-world datasets that are with different sizes and sparsity degrees: LastFM1 and
Ciao2. LastFM is a music dataset, which contains the friend relationships between 1,892 users and the interactions between the
users and music-related items. Ciao is an online shopping dataset containing the ratings given by 7,375 users on a larger num-
ber of items as well as the friend relationships among the users. Table 1 shows the statistics of the datasets used.
For each dataset, we randomly select 80% of the interaction data to form the training set, and the rest is used as the test
set. To unbiasedly tune the hyper-parameters and avoid over-fitting, we randomly select 10% of the interaction data from the
training set as the validation set.

5.1.2. Benchmark cases

To validate the effectiveness of SocialLGN, we compare it with the following 5 state-of-the-art methods:
 BPR [24] – a classical pair-wise collaborative filtering method. It maximizes the difference between the probability that
the users interact with positive samples and the probability that the users interact with negative samples.
 SBPR [39] – a MF-based social recommendation model that integrates BPR.
 DiffNet [31] – a graph-based social recommendation model. It generates the final user representation by directly taking
the vector sum of the user representations from the user-item interaction graph and the social graph.
 NGCF [29] – a GCN-based model that smooths the representation of each user or item in the user-item interaction
graph.
 LightGCN [11] – a recently prevalent model based on NGCF.

1
https://grouplens.org/datasets/hetrec-2011/
2
https://www.cse.msu.edu/tangjili/datasetcode/truststudy.htm

601
J. Liao, W. Zhou, F. Luo et al. Information Sciences 589 (2022) 595–607

Table 1
Statistics of two datasets.

Dataset LastFM Ciao

User # 1,892 7,375
Item # 17,632 105,114
Interaction # 92,834 284,086
Density(Interaction) 0.278% 0.037%
Relation # 25,434 57,544
Density(Relation) 0.711% 0.106%

5.1.3. Metrics
In this study, we use 3 widely used metrics [11,30] to evaluate the top-N ranking recommendation performance of the
proposed SocialLGN and the above 5 benchmark models: Precision, Recall, and Normalized Discounted Cumulative Gain
(NDCG). The metrics of Precision and Recall are expressed as Eqs. (11) and (12), respectively.
#tp
Precision ¼ ð11Þ
#tp þ #fp
#tp
Recall ¼ ð12Þ
#tp þ #fn

where #tp; #fp; #fn denote the number of true-positive, false-positive, and false-negative recommended items, respectively.
The true-positive and false-positive recommended items are the result of interacted recommendation. The false-negative
recommended items are the result of non-interacted recommendation.

X
k
r ð iÞ
r ð 1Þ þ
log i2
i¼2
NDCG@N ¼ ð13Þ
X
jRELj
r ð iÞ
ðiþ1Þ
log 2
i¼1

where rðiÞ is the relevance score. r ðiÞ ¼ 1 indicates the recommended item has interaction with the user; rðiÞ ¼ 0 indicates
there is no interaction. jRELj represents the sum of the top-N recommended items that are sorted in descending order by the
relevant score.
For each of the 3 metrics, a larger value indicates a better recommendation performance. Due to the sparsity nature of the
user-item interaction records, for each user, we repeatedly and randomly sample one unrated item as a negative sample; we
then combine the negative sample with an item that has interaction with the user in the training process. To achieve unbi-
ased evaluation, we rank all the items rather than only ranking the sampled items. For each experiment that is performed on
a specific model and one of the two datasets, we repeat the experiment 5 times and calculate the average result for
comparison.

5.1.4. Parameter settings

For all the SocialLGN and the 5 benchmark models, the number of the representation factor (i.e., the dimension of embed-
ding) is set to be fixed at 64, and the training batch size is set to be fixed at 1,024. For other settings, we follow the literature
[11]: Adam [13] with an initial learning rate of 1e3 is used for optimizing the parameters of SocialLGN. A Gaussian distri-
bution with a mean value of 0 and a standard deviation of 0.01 is used to initialize Eð0Þ . The value of k is set to be 1e4 by trials
 
in the range of 1e6 ; 1e5 ; . . . ; 1e2 . The combination weight (aK ) for the kth propagation layer’s output is set to be ðKþ1
1
Þ
. In
this simulation, the number of propagation layers (K) is set to be 3.

5.2. Recommendation performance evaluation

We evaluate the recommendation performance of SocialLNG and the benchmark models on the 2 datasets of LastFM and
Ciao. Considering the cold-start problem is an important issue in most recommendation system designs, we particularly
design a case to evaluate the recommendation performance of the models subjected to the cold-start problem. In the social
recommendation scenario studied in this paper, the cold-start problem corresponds to generating recommendation results
for the users who only interact with a few items. We treat all the users in the test set who has less than 20 interactions as
cold-start users; based on this, we construct a separate cold-start test set that only contains these identified cold-start users.
Table 2 reports the values of the 3 metrics of SocialLGN and the 5 benchmark models, which are calculated from the recom-
mendation results on the original test set with top-10 and top-20, respectively. Table 3 shows the results on the cold-start
test set.
From the results, it can be seen that the MF-based models (BPR and SBPR) perform unsatisfactorily on both the original
and cold-start test sets – much worse than GNN-based models (DiffNet, NGCF, LightGCN, and SocialLGN). LightGCN signif-
602
J. Liao, W. Zhou, F. Luo et al. Information Sciences 589 (2022) 595–607

Table 2
Comparison of SocialLGN and benchmark cases on the original LastFM and Ciao datasets.

Dataset Metric BPR SBPR DiffNet NGCF LightGCN SocialLGN Improve.

LastFM Precision@10 0.0922 0.1398 0.1727 0.1766 0.1961 0.1972 0.5609%
Precision@20 0.0720 0.1010 0.1215 0.1269 0.1358 0.1368 0.7364%
Recall@10 0.0962 0.1442 0.1779 0.1796 0.2003 0.2026 1.1483%
Recall@20 0.1499 0.2070 0.2488 0.2576 0.2769 0.2794 0.9029%
NDCG@10 0.1099 0.1749 0.2219 0.2287 0.2536 0.2566 1.1830%
NDCG@20 0.1321 0.1978 0.2474 0.2563 0.2788 0.2883 3.4075%
Ciao Precision@10 0.0145 0.0179 0.0238 0.0228 0.0271 0.0276 1.8450%
Precision@20 0.0111 0.0141 0.0182 0.0179 0.0202 0.0205 1.4851%
Recall@10 0.0220 0.0259 0.0341 0.0343 0.0410 0.0430 2.4390%
Recall@20 0.0339 0.0412 0.0527 0.0531 0.0591 0.0618 4.5685%
NDCG@10 0.0229 0.0266 0.0359 0.0359 0.0437 0.0441 0.9153%
NDCG@20 0.0260 0.0307 0.0403 0.0407 0.0478 0.0486 1.6736%

Table 3
Comparison of SocialLGN and benchmark cases on the cold-start LastFM and Ciao datasets.

Dataset Metric BPR SBPR DiffNet NGCF LightGCN SocialLGN Improve.

LastFM Precision@10 0.0282 0.0292 0.0417 0.0333 0.0417 0.0458 9.8321%
Precision@20 0.0209 0.0333 0.0271 0.0292 0.0313 0.0333 6.3898%
Recall@10 0.1151 0.1123 0.1713 0.1169 0.1727 0.1974 14.3023%
Recall@20 0.1615 0.2467 0.2407 0.2141 0.2416 0.2663 10.2235%
NDCG@10 0.0828 0.0709 0.1107 0.1074 0.1374 0.1419 3.2751%
NDCG@20 0.0989 0.1159 0.1309 0.1411 0.1560 0.1643 5.3205%
Ciao Precision@10 0.0061 0.0070 0.0104 0.0104 0.0131 0.0134 2.2901%
Precision@20 0.0047 0.0060 0.0081 0.0085 0.0096 0.0097 1.0417%
Recall@10 0.0208 0.0234 0.0339 0.0341 0.0429 0.0441 2.7972%
Recall@20 0.0328 0.0384 0.0539 0.0557 0.0616 0.0630 2.2727%
NDCG@10 0.0138 0.0165 0.0248 0.0245 0.0319 0.0328 2.8213%
NDCG@20 0.0179 0.0219 0.0316 0.0319 0.0384 0.0394 2.6042%

icantly outperforms BPR, SBPR, DiffNet, and NGCF in almost all the cases. This could be explained that LightGCN excludes the
two conventional operations in GCN, i.e., the nonlinear activation and linear transformation. As is pointed out by [11], these
two components in GCN would negatively affect the recommendation’s accuracy.
The results clearly demonstrate that the proposed SocialLGN shows satisfactory performance in all the cases. For example,
compared with LightGCN, SocialLGN achieves 4.57% improvement in terms of Recall@20 metric on the original Ciao dataset
and 3.41% improvement in terms of NDCG@20 metric on the original LastFM dataset, respectively. Since LightGCN does not
include the propagation process in the social graph, by comparing the results of SocialLGN and LightGCN in Table 3, it can be
seen that the propagation mechanism in the social graph can effectively help the whole system to alleviate the cold-start
problem. It can be seen that compared with LightGCN, SocialLGN achieves 14.30% and 10.22% improvements in terms of
Recall@10 and Recall@20 metrics respectively on the cold-start LastFM dataset. Further, SocialLGN performs better than
LightGCN on the Ciao cold-start dataset in terms of the NDCG@10 metric (2.82% improvement). This well demonstrates
SocialLGN’s capability in alleviating the cold-start problem in the social recommendation. Also, the results reflect that the
denser the social graph is, the more such improvements can be achieved. The social graph in the LastFM dataset is denser
than that in the Ciao dataset (0.71% and 0.11% densities, respectively). The experiment results show that for each of all
the 3 metrics, even the minimum improvement magnitude achieved by SocialLGN on the LastFM dataset is larger than
the maximum improvement magnitude achieved on the Ciao dataset.

5.3. Ablation and effectiveness analyses

Ablation study is performed to investigate the impact of the 2 critical operations in SocialLGN (i.e., graph fusion and layer
combination) on the final recommendation performance.

5.3.1. Impact of graph fusion

To validate the effectiveness of the proposed graph fusion operation (Eq. (3)), we compare the recommendation perfor-
mance of SocialLGN with two variations:(i) Replace the proposed graph fusion operation with the fusion operation in GCN
[14], which adds up the user’s representations in the 2 graphs (Eq. (14)))); and(ii) Replace the proposed graph fusion oper-
ation with the GraphSage fusion operation [10], which concatenate the user’s representations in the 2 graphs (Eq. (15)).

603
J. Liao, W. Zhou, F. Luo et al. Information Sciences 589 (2022) 595–607


f GCN ¼ r W pðukþ1Þ þ qðukþ1Þ ; ð14Þ



f GraphSage ¼ r W pðukþ1Þ jjqðukþ1Þ ; ð15Þ

In Eqs. (14) and (15), rðÞ is the activation function; W is a trainable transformation matrix; d denotes the latent vector’s
dimension; jj represents the concatenation operation.
The comparison results are shown in Fig. 3. It can be seen that the proposed graph fusion operation obtains the best per-
formance among the three fusion operations in both datasets, while the GCN fusion model produces the worst performance.
The reason for the superior performance of SocialLGN is that different from the fusion operation in GCN and the GraphSage
fusion operation, the proposed fusion operation separately applies the feature transformation and nonlinear activation oper-
ðkÞ ðkÞ
ations to qu and pu respectively. In this way, the user’s latent performances learned from the social graph and the user-item
interaction graph can be transformed into a union space. Besides, the normalization operation can suppress the increase of
ðkÞ
eu with the increase of k, which is also helpful to improve the recommendation performance.

5.3.2. Impact of layer combination

Next, we conduct an ablation study to investigate the effectiveness of the layer combination operation in SocialLGN. We
compare SocialLGN with a variation, which directly uses the output of the last layer (EðK Þ ) as the final representation of the
users and items.
The comparison results are shown in Table 4. The results reflect that the variation shows much poorer performance in
modeling the users’ and items’ representations than SocialLGN. This is because without performing layer combination,
the over-smoothing problem is more obvious in the variation, as it only utilizes the highest-order neighbors and neglects
the low-order neighbors.

Fig. 3. Impact of graph fusion operation on recommendation performance.

Table 4
Performance of SocialLGN and the variant.

Dataset LastFM Ciao

Method Recall@10 NDCG@10 Recall@10 NDCG@10
variant 0.0726 0.0974 0.0259 0.0293
SocialLGN 0.2026 0.2566 0.0430 0.0441

604
J. Liao, W. Zhou, F. Luo et al. Information Sciences 589 (2022) 595–607

5.4. Sensitivity analysis for hyper-parameters

In this section, we report the results of sensitivity analysis conducted for 2 important hyper-parameters in SocialLGN: the
number of the propagation layers (K) and the L2 regularization coefficient (k). K plays an important role when applying
SocialLGN to new datasets, as it determines the propagation depth of different graphs. In this experiment, we investigate
the system’s performance under: (i) varying values of K from 1 to 6. In this case, the value of k is fixed at 1e4 . And (ii) varying
 
values of k in 0; 1e6 ; 1e5 ; . . . ; 1e2 , while the value of K is fixed at 3. The sensitivity study results are shown in Figs. 4 and
5.
Fig. 4 shows that on the LastFM dataset, the performance of SocialLGN improves rapidly when the value of K increases
from 1 to 3. At K ¼ 3, the system achieves the best performance. The system’s performance slightly drops when K is larger
than 5. A similar trend is also observed on the Ciao dataset, where SocialLGN achieves the best performance at K ¼ 4. These
observations indicate that the over-large value of K would lead to the over-smoothing effect [11,30], which would conse-
quentially deteriorate the recommendation performance.
From Fig. 5, it can be seen that when the value of k is relatively small (from 0 to 1e5 ), increasing its value only generates
marginal improvement for SocialLGN’s recommendation performance. When k ¼ 1e4 , SocialLGN achieves the best perfor-
mance: the values of Precision@10, Recall@10, and NDCG@10 are 0.1972, 0.2026, 0.2566 on LastFM dataset and 0.0276,
0,0430, 0.0441 on Ciao dataset, respectively. When k is larger than 1e4 (on LastFM) or 1e3 (on Ciao), the system’s perfor-
mance sharply drops. These results show that SocialLGN is less prone to overfitting; therefore, over regularization would
negatively affect its recommendation accuracy.

Fig. 4. Sensitivity analysis for K.

Fig. 5. Sensitivity analysis for k.

605
J. Liao, W. Zhou, F. Luo et al. Information Sciences 589 (2022) 595–607

6. Conclusion

This paper proposes a new social recommendation system based on the light graph convolution network, namely
SocialLGN. The system propagates the users’ representations in the user-item interaction graph and the social graph simul-
taneously. Based on this, a new graph fusion operation is designed to fuse the two user representations obtained from the
propagation into an integrated representation. The system also propagates the item representation in the user-item interac-
tion graph.
The space and time complexities of SocialLGN are analyzed, showing that the model training process of SocialLGN leads to
reasonable memory and computation time costs. The social recommendation accuracy of SocialLGN is further validated by
comprehensive experiments. We compare SocialLGN with 5 state-of-the-art recommendation models on two real-world
datasets, and the results show SocialLGN can achieve better recommendation performance than the other models. The
results also demonstrate that SocialLGN can effectively alleviate the cold-start problem in the social recommendation. Sen-
sitivity studies are performed for critical hyper-parameters; this is expected to provide parameter tuning guidance to
researchers or engineers who are interested in implementing this work.

CRediT authorship contribution statement

Jie Liao: Conceptualization, Methodology, Data curation, Writing - original draft. Wei Zhou: Project administration, Visu-
alization, Investigation, Funding acquisition. Fengji Luo: Writing - review & editing. Junhao Wen: Investigation, Supervision,
Validation, Formal analysis, Funding acquisition. Min Gao: Data curation, Supervision. Xiuhua Li: Validation, Supervision.
Jun Zeng: Validation, Supervision.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have
appeared to influence the work reported in this paper.

Acknowledgments

This work is in part supported by the National Natural Science Foundation of China (Grant No. 62072060, 72074036) and
in part supported by a Discovery Project from the Australian Research Council (DP220103881).

References

[1] Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. Spectral networks and locally connected networks on graphs. In 2nd International
Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14–16, 2014, Conference Track Proceedings, 2014..
[2] Zhao-Min Chen, Xiu-Shen Wei, Peng Wang, Yanwen Guo, Multi-label image recognition with graph convolutional networks, in: In 2019 IEEE/CVF
Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 5172–5181.
[3] Robert B. Cialdini, Noah J. Goldstein, Social influence: Compliance and conformity, Annu. Rev. Psychol. 55 (1) (2004) 591–621, PMID:14744228.
[4] Michaël Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. In
Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, December 5–10, 2016,
Barcelona, Spain, pages 3837–3845, 2016..
[5] Wenqi Fan, Qing Li, and Min Cheng. Deep modeling of social relations for recommendation. In Proceedings of the AAAI Conference on Artificial
Intelligence, volume 32, 2018..
[6] Wenqi Fan, Yao Ma, Qing Li, Yuan He, Eric Zhao, Jiliang Tang, and Dawei Yin. Graph Neural Networks for Social Recommendation. In The World Wide
Web Conference on - WWW ’19, pages 417–426, San Francisco, CA, USA, 2019. ACM Press..
[7] G. Guo, J. Zhang, N. Yorke-Smith, A novel recommendation model regularized with user trust and item ratings, IEEE Trans. Knowl. Data Eng. 28 (7)
(2016) 1607–1620.
[8] Guibing Guo, Jie Zhang, Neil Yorke-Smith, Trustsvd: Collaborative filtering with both the explicit and implicit influence of user trust and of item ratings,
Proceedings of the Twenty-Ninth AAAI Conference on Artificial Intelligence, AAAI’15, AAAI Press, 2015, pp. 123–129.
[9] Lei Guo, Hongzhi Yin, Qinyong Wang, Tong Chen, Alexander Zhou, and Nguyen Quoc Viet Hung. Streaming session-based recommendation. In
Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD ’19, page 1569-1577, New York, NY, USA,
2019. Association for Computing Machinery..
[10] William L. Hamilton, Rex Ying, Jure Leskovec, Inductive representation learning on large graphs, in: Proceedings of the 31st International Conference
on Neural Information Processing Systems, Curran Associates Inc., Red Hook, NY, USA, 2017.
[11] Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yong-Dong Zhang, Meng Wang, 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, SIGIR 2020,
Virtual Event ACM, July 25–30, 2020, China, 2020, pp. 639–648.
[12] M. Jiang, P. Cui, F. Wang, W. Zhu, S. Yang, Scalable recommendation with social contextual information, IEEE Trans. Knowl. Data Eng. 26 (11) (Nov
2014) 2789–2802.
[13] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR
2015, San Diego, CA, USA, May 7–9, 2015, Conference Track Proceedings, 2015..
[14] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In 5th International Conference on Learning
Representations, ICLR 2017, Toulon, France, April 24–26, 2017, Conference Track Proceedings. OpenReview.net, 2017..
[15] Yehuda Koren, in: Factorization meets the neighborhood: A multifaceted collaborative filtering model In Proceedings of the 14th ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining, Association for Computing Machinery, New York, NY, USA, 2008.
[16] Yehuda Koren, Robert M. Bell, Chris Volinsky, Matrix factorization techniques for recommender systems, Computer 42 (8) (2009) 30–37.

606
J. Liao, W. Zhou, F. Luo et al. Information Sciences 589 (2022) 595–607

[17] Chun-Yi Liu, Chuan Zhou, Jia Wu, Yue Hu, and Li Guo. Social recommendation with an essential preference space. In Proceedings of the Thirty-Second
AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI Symposium
on Educational Advances in Artificial Intelligence (EAAI-18), New Orleans, Louisiana, USA, February 2–7, 2018, pages 346–353. AAAI Press, 2018..
[18] Huafeng Liu, Liping Jing, Yu. Jian, K. Ng. Michael, Social Recommendation with Learning Personal and Social Latent Factors, IEEE Trans. Knowl. Data Eng.
33 (7) (2021) 2956–2970.
[19] Xien Liu, Xinxin You, Xiao Zhang, Ji Wu, and Ping Lv. Tensor graph convolutional networks for text classification. In The Thirty-Fourth AAAI Conference
on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI
Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7–12, 2020, pages 8409–8416. AAAI Press,
2020..
[20] Hao Ma, Irwin King, Michael R. Lyu, Learning to recommend with social trust ensemble, in: Proceedings of the 32nd International ACM SIGIR
Conference on Research and Development in Information Retrieval, SIGIR ’09, Association for Computing Machinery, New York, NY, USA, 2009.
[21] Hao Ma, Haixuan Yang, Michael R. Lyu, Irwin King, Sorec: social recommendation using probabilistic matrix factorization, Proceedings of the 17th ACM
Conference on Information and Knowledge Management, CIKM 2008, 26–30, Napa Valley, California, USA, October 2008, p. 2008.
[22] Hao Ma, Dengyong Zhou, Chao Liu, Michael R. Lyu, Irwin King, Recommender systems with social regularization, in: Proceedings of the Fourth ACM
International Conference on Web Search and Data Mining, Association for Computing Machinery, New York, NY, USA, 2011.
[23] Miller McPherson, Lynn Smith-Lovin, and James M Cook. Birds of a Feather: Homophily in Social Networks. Annual Review of Sociology, 27(1), 415–
444, August 2001..
[24] Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, Lars Schmidt-Thieme, Bpr: Bayesian personalized ranking from implicit feedback, UAI 2009,
Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence, 18–21, Montreal, QC, Canada, June, 2009, p. 2009.
[25] Lei Shi, Yifan Zhang, Jian Cheng, Lu. Hanqing, Two-stream adaptive graph convolutional networks for skeleton-based action recognition, in: In 2019
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 12018–12027.
[26] Weiping Song, Zhiping Xiao, Yifan Wang, Laurent Charlin, Ming Zhang, and Jian Tang. Session-based social recommendation via dynamic graph
attention networks. In Proceedings of the Twelfth ACM International Conference on Web Search and Data Mining, pages 555–563, 2019..
[27] Jiliang Tang, Hu. Xia, Huan Liu, Social recommendation: a review, Social Network Analysis and Mining 3 (4) (2013) 1113–1133.
[28] Jiachun Wang, Fajie Yuan, Jian Chen, Qingyao Wu, Min Yang, Yang Sun, and Guoxiao Zhang. Stackrec: Efficient training of very deep sequential
recommender models by iterative stacking. In Fernando Diaz, Chirag Shah, Torsten Suel, Pablo Castells, Rosie Jones, and Tetsuya Sakai, editors, SIGIR
’21: The 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, Virtual Event, Canada, July 11–15, 2021,
pages 357–366. ACM, 2021..
[29] Xiang Wang, Xiangnan He, Meng Wang, Fuli Feng, and Tat-Seng Chua. Neural Graph Collaborative Filtering. In Proceedings of the 42nd International
ACM SIGIR Conference on Research and Development in Information Retrieval, pages 165–174, Paris France, July 2019. ACM..
[30] L. Wu, J. Li, P. Sun, R. Hong, Y. Ge, M. Wang, DiffNet++: A Neural Influence and Interest Diffusion Network for Social Recommendation, IEEE Trans.
Knowl. Data Eng. (2020), 1-1.
[31] Wu. Le, Peijie Sun, Fu. Yanjie, Richang Hong, Xiting Wang, Meng Wang, A neural influence diffusion model for social recommendation, in: Benjamin
Piwowarski, Max Chevalier, Éric Gaussier, Yoelle Maarek, Jian-Yun Nie, Falk Scholer (Eds.), Proceedings of the 42nd International ACM SIGIR
Conference on Research and Development in Information Retrieval, SIGIR 2019 ACM, July 21–25, 2019, Paris, France, 2019, pp. 235–244.
[32] Qitian Wu, Hengrui Zhang, Xiaofeng Gao, Peng He, Paul Weng, Han Gao, and Guihai Chen. Dual graph attention networks for deep latent representation
of multifaceted social effects in recommender systems. In The World Wide Web Conference, pages 2091–2102, 2019..
[33] Shiwen Wu, Wentao Zhang, Fei Sun, and Bin Cui. Graph Neural Networks in Recommender Systems: A Survey. arXiv:2011.02260 [cs], November 2020.
arXiv: 2011.02260.
[34] B. Yang, Y. Lei, J. Liu, W. Li, Social collaborative filtering by trust, IEEE Trans. Pattern Anal. Mach. Intell. 39 (8) (Aug 2017) 1633–1647.
[35] J. Yu, H. Yin, J. Li, M. Gao, Z. Huang, L. Cui, Enhance social recommendation with adversarial graph convolutional networks, IEEE Trans. Knowl. Data Eng.
(2020), 1-1.
[36] Jun Yu, Min Tan, Hongyuan Zhang, Dacheng Tao, and Yong Rui. Hierarchical deep click feature prediction for fine-grained image recognition. IEEE
Transactions on Pattern Analysis and Machine Intelligence, pages 1–1, 2019..
[37] Junliang Yu, Min Gao, Hongzhi Yin, Jundong Li, Chongming Gao, and Qinyong Wang. Generating reliable friends via adversarial training to improve
social recommendation. In 2019 IEEE International Conference on Data Mining (ICDM), pages 768–777. IEEE, 2019..
[38] Junliang Yu, Hongzhi Yin, Jundong Li, Q. Wang, N. Hung, and X. Zhang. Self-supervised multi-channel hypergraph convolutional network for social
recommendation. ArXiv, abs/2101.06448, 2021..
[39] Tong Zhao, Julian McAuley, Irwin King, Leveraging social connections to improve personalized ranking for collaborative filtering, in: Proceedings of the
23rd ACM International Conference on Conference on Information and Knowledge Management, Association for Computing Machinery, New York, NY,
USA, 2014.

607