This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the

exception of pagination.

IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS 1

Community Enhanced Knowledge Graph

for Recommendation
Zhen-Yu He , Chang-Dong Wang , Senior Member, IEEE, Jinfeng Wang , Member, IEEE,
Jian-Huang Lai , Senior Member, IEEE, and Yong Tang

Abstract—Due to the capability of encoding auxiliary in- I. INTRODUCTION

formation for alleviating the data sparsity issue, knowledge
ECOMMENDER system plays an important role in
R
graph (KG) has gained an increasing amount of attention in
recent years. With auxiliary knowledge about items, the KG- many online platforms [1], since it is an effective tool to al-
based recommender systems have achieved better performance leviate the information overload problem [2]. To mine user
compared with the existing methods. However, the effectiveness preference from historical interaction data, collaborative fil-
of the KG-based methods highly depends on the quality of
the KG. Unfortunately, KGs are usually with the problem of tering (CF) methods [3], [4], [5], [6], [7] have been applied
incompleteness and sparseness. Besides, the existing KG-based to discover users’ latent interest, and they have shown great
methods could not discriminate the importance of different success. In recent years, the CF-based methods are extended
factors that users consider when making decisions, which may to graph structured data, and LightGCN [8] is one of the most
degrade the interpretability of the methods. In this article, representative works, which linearly propagates embeddings
we propose a recommendation model named community en-
hanced knowledge graph for recommendation (CEKGR). By on the user–item interaction graph. However, the data sparsity
adding entities and relations, the KG is enriched with more issue has restricted the performance of the CF-based
semantic information, which would help mine users’ preference methods [9], [10], [11], [12]. Since knowledge graph (KG)
for better recommendation. With weights of each path, the contains abundant semantic information about items, many
interpretability of the recommendation can be improved. To
efforts have been made in designing various kinds of KG-
validate the effectiveness of the proposed method, we conduct
experiments on three public datasets. Experiment results have based recommender systems [13], [14], [15], [16], which
shown the improvement compared with other state-of-the-art show quite fine performance on the
methods. Besides, case study has illustrated the interpretability mining of user preference.
of the proposed recommendation model. Generally, the KG-based recommender system is a kind of
Index Terms—Interpretability, knowledge graph (KG) recommender system that utilizes the rich knowledge about
enhancement, path representation, recommendation. items in the KG to promote user interest mining. In addition
to KG, other kinds of auxiliary information is also utilized
to improve the recommendation performance, such as review
text [17], [18], image [19], and social networks [20]. Com-
Manuscript received 19 September 2023; revised 19 January 2024 and
11 March 2024; accepted 27 March 2024. This work was supported in pared with the methods utilizing other kinds of information,
part by the National Key Research and Development Program of China the KG-based recommendation methods have the following
under Grant 2021YFF1201200, in part by NSFC under Grant 62276277, advantages. 1) KG contains various aspects of knowledge
in part by Guangdong Basic and Applied Basic Research Foundation under
Grant 2022B1515120059, and in part by Guangdong Provincial Key Labo- about items. Actually, KG is a kind of heterogeneous graph,
ratory of Intellectual Property and Big Data under Grant 2018B030322016. which contains various kinds of entities and relations. For
(Corresponding author: Chang-Dong Wang.) example, in a movie KG, the various aspects of attributes that
Zhen-Yu He and Chang-Dong Wang are with the School of Computer
Science and Engineering, Sun Yat-sen University, Guangzhou 510006, China, a certain movie contains can be expressed by edges between
also with Guangdong Provincial Key Laboratory of Intellectual Property this movie and corresponding entities. In contrast, images only
and Big Data, Guangzhou 510665, China, and also with Key Laboratory provide the visual features for items, and social networks only
of Machine Intelligence and Advanced Computing, Ministry of Education,
Guangzhou 510006, China (e-mail: [email protected]; changdong- supply social relations among users. Although there are many
[email protected]). features of items in the review text, it is challenging to extract
Jinfeng Wang is with the College of Mathematics and Informatics, these features from text. 2) KG enhances the interpretability of
South China Agricultural University, Guangzhou 510642, China (e-mail:
[email protected] recommender systems. Obviously, it is difficult to provide
Jian-Huang Lai is with the School of Computer Science and Engineering, explanations based on the features that are extracted from
Sun Yat-sen University, Guangzhou 510006, China, also with Guangdong images. Social networks can provide reasons for
Key Laboratory of Information Security Technology, Guangzhou 510006,
China, and also with the Key Laboratory of Machine Intelligence and recommendation based on relations among users, but these
Advanced Computing, Ministry of Education, Guangzhou 510006, China (e- explanations only focus on the aspects of users. As for review
mail: [email protected]). text, a typical approach to give ex- planations is to utilize the
Yong Tang is with the School of Computer Science, South China Normal
University, Guangzhou 510631, China (e-mail: [email protected]). existing reviews [17], but it may not be appropriate for a
Digital Object Identifier 10.1109/TCSS.2024.3383603 certain user. For KG, the connections between items and their
attributes can provide explanations with various aspects of user
preference. Besides, the paths linking a

2329-924X © 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
ed licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on August 10,2024 at 05:42:33 UTC from
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the
exception of pagination.

2 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS

recommendation.
TABLE I
INCOMPLETENESS AND SPARSENESS IN
THE FREEBASE

Percentage of Unknown Value

Relation
All 3M Top 100k
Profession 68% 24%
Place of birth 71% 13%
Nationality 75% 21%
Education 91% 63%
Spouses 92% 68%
Parents 94% 77%
Children 94% 80%
Siblings 96% 83%
Ethnicity 99% 86%

certain user and a candidate item guarantee the personalization

of explanations.
As described above, KG is a kind of heterogeneous graph,
which describes the relations among items and their relevant
attributes. Intuitively, these relations can supply rich informa-
tion for the refinement of representations of entities. For ex-
ample, a KG embedding method called TransR [21] is utilized
to construct items’ structural representations, which are then
used to predict users’ preference [19]. On the other hand, there
are edges connecting two relevant entities in the KG. Based
on these edges, we can obtain paths between a user and an
item or high-order connection patterns between two entities.
By mining the semantic information of paths or high-order
connection patterns, latent user preference can be predicted. In
recent years, some methods [22], [23], [24] utilize the
connections in the KG to conduct recommendation. The above
methods achieve quite fine performance due to the utilization
of the KG. Despite the success, the above-mentioned methods
highly depend on the quality of the KG. Unfortunately, the
KG is usually with the problem of incompleteness and
sparseness [25], [26], [27]. For better demonstration of the
above-mentioned situation, we take Freebase [28] as an
example. Freebase is a practical, scalable tuple database used
to organize general human knowledge, and it contains
plenty of information. However, it suffers from
incompleteness and sparseness. For the type Person in
Freebase, there are unknown values for commonly used
relations. Specif- ically, we list the statistic that is provided by
West et al. [29] in Table I. The All 3M column is for all the
roughly three million people in Freebase, and the Top 100k is
for 100 000 most frequently searched-for people. From Table
I, we can see that the percentages of unknown values are quite
high. Since most researchers utilize public available KGs such
as Freebase to construct the KG which is used for KG-based
recommendation, it is inevitable that those KGs used in
recommendation scenario will also be with the above-
mentioned issues. With incomplete KG, the effectiveness of
the KG-based recommendation would degrade. To improve
the effectiveness of the KG-based recom- mendation, it is
necessary to enrich the KG.
Considering the interpretability of the KG-based recom-
mender systems, it is easy to generate explanations based on
the paths connecting users and items. Therefore, it is intuitive
for us to adopt paths when designing KG-based method for
the purpose of easily generating explanations of
ed licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on August 10,2024 at 05:42:33 UTC from
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the
exception of pagination.

The rest of this article is organized as follows. We review

Generally, there are several paths linking a user and an item, relevant research in Section II. Related concepts and the task
and different paths can reflect different aspects of user
preference. In recent years, there are some methods [22], [23]
utilizing paths to conduct recommendation, but these
methods fail to discriminate the importance of different paths
reflecting different factors that users consider when making
decisions, which would degrade the interpretability of the
method. Since users have different preferences for each
aspect of items, it is necessary to differ- entiate the
importance of each aspect if we want to enhance the
interpretability of recommender systems and improve user
satisfaction. For example, it is common that a user likes the
film Fast Five because this user has the preference for the
famous actor Vin Diesel, and the secondary consideration is
the film genre Action. The interpretability and user
satisfaction of a KG- based recommender system is
improved once it can figure out this user’s main preference
for Vin Diesel for the film Fast Five. As a result, it is better
for us to distinguish the importance of the various factors
reflected by paths linking user–item pairs.
In this article, we propose a new model called community
en- hanced knowledge graph for recommendation (CEKGR),
which enriches the original KG and aggregates multipath
with weights derived by the similarity between entities in
path and the target item. To overcome the issue of
incompleteness and sparse- ness, CEKGR first enriches the
original KG with community structure in the KG.
Specifically, entities in the original KG are first clustered,
based on which the corresponding clusters are generated.
These clusters can be considered as the new entities, and
they are then added into the original KG. By this way, the
KG is enhanced with more semantic information. Due to the
fact that the paths linking certain user–item pairs can be
considered as entity sequences, long short-term memory
(LSTM) is adopted to learn the representations of the
paths in the enhanced KG. Since there are several paths
connecting a certain user–item pair and our task is to predict
the user– item interaction probability, a weighting
mechanism is utilized to merge the path representations
concerning a certain user– item interaction into a path set
representation. Finally, the path set representation is
transformed into the interaction probability with a multilayer
perceptron (MLP).
The main contributions of this article are as follows.
1) In this article, we propose a community detection-
based KG enhancement approach, which relies on
adding addi- tional entities that are derived from
clusters in the original KG and corresponding relations
into the original KG. By this way, the KG can be
enriched, which further improves the effectiveness of
mining users’ preference.
2) The representations of the paths that connect a certain
user–item pair are aggregated with a weighting mecha-
nism. The weight of each path can also provide infor-
mation for the identification of the different
importance of factors that affect a user’s decision
process, which is helpful for the improvement on the
interpretability of recommender system.
3) Extensive experiments are conducted on three
datasets, and the results demonstrate the effectiveness of
CEKGR.
ed licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on August 10,2024 at 05:42:33 UTC from
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the
exception of pagination.

HE et al.: COMMUNITY ENHANCED KNOWLEDGE GRAPH FOR RECOMMENDATION 3

TABLE II method of graph neural network and the attention mechanism

NOTATIONS USED IN THIS ARTICLE is utilized to differentiate the importance of the neighbors.
Notations Descriptions Knowledge graph convolutional networks (KGCNs) [36] sam-
G Knowledge graph ples neighbors of each entity in the KG and incorporates
E = {e1, e2, ..., em} Entity set neigh- borhood information with bias during the calculation of
R = {r1, r2, ..., Relation set
rn } S Community the representation for a given entity. Knowledge graph-based
C = {Cl1 , Cl2 , ..., Clustering result intent network (KGIN) [37] takes into account user–item
Clj } GE Enhanced knowledge graph relationships at the finer granularity of intents and long-range
A User–item interaction set
(u, v) A user–item interaction semantics of relational paths within the GNN paradigm.
P(u,v) = pi .1 ≤ i ≤ n Path set associated with (u, v) KGRec [38] enables the distillation of the useful knowledge
(u,v)
Path representation for path pi connections within the KG for recommendation and align
hi (u,v)
l,(u,v)
H(u,v) Representation of path set P(u,v) them in a noise-free and rationale-aware manner.
yuv Actual implicit feedback for (u, v) The above KG-based methods achieve quite fine performance
y^uv Predicted interaction for (u, v)
Loss function due to the rich information related to items in the KG. In other
L
words, the effect of these methods is highly correlated with
the introduction of KG. As mentioned in Section I,
formulation of KG-based recommendation are presented in incompleteness and sparseness of the existing KGs are
Sec- tion III. Then, the proposed method is described in inevitable. Therefore, actually the effect of the KG-based
Section IV. We then present the experimental results in methods may be limited. To alleviate this, KGs should be
Section V. Finally, we make conclusion in Section VI. enriched for better recommen- dation performance.

II. RELATED WORK

B. Explainable Recommendation Based on KG
A. KG-Based Recommendation Methods In recent years, explainable recommendation has attracted
Since the issue of data sparsity commonly appears in online a large amount of attention. In terms of the goal, explainable
platforms, it is inevitable that the CF methods [30], [31], [32], recommendation not only achieves the satisfactory recommen-
[33], [34] would suffer from performance degradation. As a dation performance but also provides reasonable explanations
result, it is natural to introduce auxiliary information into rec- why a certain product or service has been recommended to a
ommender systems. In recent years, KG-based certain user [39]. Since the KG contains many entities and
recommendation has attracted more and more attention for the rela- tions, it is of great help to conduct recommendation and
rich semantic in- formation contained in a KG. Generally, the provide the corresponding explanations. Intuitively, paths
KG-based methods can be divided into three categories, linking a user and an item can serve as the reasons why this
namely embedding-based methods, path-based methods, and user would select the item. RippleNet [24] can provide
propagation-based methods. The embedding-based methods explanations by finding paths between user’s historical
first learn the embeddings of entities and relations based on interacted item and a candidate item. Temporal metapath
the KG, and then recommen- dation is conducted with the guided explainable recommendation (TMER) [40] utilizes
above learned embeddings. Col- laborative knowledge base paths between user–item and item–item pairs which are
embedding (CKE) [19] first mines feature of items based on constructed based on predefined metapath pat- terns to serve
the structural information of a certain KG together with text as explanations. RKGE [22] helps supply latent explanations
and image about items. Then, CF is used by mining paths between the rated items and the correctly
to predict users’ preference for items. recommended items, and it can be found that most of the
The path-based methods regard paths connecting user–item correctly recommended items are connected with the rated
pairs as basis of exploration of user latent interest. Recurrent items by different types of paths, which to some extent proves
knowledge graph embedding (RKGE) [22] first mines paths that user–item interactions are coinfluenced by different paths.
between user–item pair, and paths are then modeled by recurrent However, RippleNet focuses on the relevance probabil- ity
neural network (RNN) to obtain their representations, which between nodes, and it cannot clearly give the weights of
are utilized to predict the interaction probability. Similar to different paths that connect a user and a candidate item. Then,
RKGE, knowledge-aware path recurrent network (KPRN) RKGE could not clearly distinguish the different contribution
[23] adopts LSTM to model paths and estimates the score with of different paths, so it is hard to discriminate the importance
a pooling layer. Different from RKGE, KPRN encodes entity of different factors that affect the decision process of a user.
type, entity value, and relation in paths before feeding them At the same time, the sampled paths in TMER are based on
into LSTM, while RKGE only adopts entity value. metapath connection patterns. Generally, the patterns of
The propagation-based methods attempt to refine the rep- metapath are manually selected, so its selection is somewhat
resentations of entities and relations through the high-order subjective and the number of it is limited, which would
connection patterns contained in a certain KG. RippleNet [24] possibly degrade the recommendation performance.
propagates user preference over a KG through the links To improve the interpretability of recommender systems,
among entities. Knowledge graph attention network (KGAT) the KG-based methods should be able to distinguish the
[35] prop- agates embeddings from a node’s neighbors based different importance of different paths for the purpose of the
on the analysis of
ed licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on August 10,2024 at 05:42:33 UTC from
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the
exception of pagination.

4 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS

Fig. 1. Overall framework of the proposed model. The symbols beginning with u,i,e, and s represent the users, the items, the entities, and the communities,
respectively. Due to the space limitation, the module of KG enhancement is illustrated in a separate figure.

user preference, which could improve the persuasiveness and then lead to the poor performance in the KG-based
user satisfaction of the recommender systems. recommen- dation. As a result, it is necessary to enrich the
KG for better recommendation performance.
III. PRELIMINARIES Intuitively, a possible way to enrich KG is to add additional
entities and relations. With this scheme, the semantic infor-
In this section, we introduce related concepts and the task
mation of KG would be enriched, and more paths between
formulation of KG-based recommendation.
user–item pairs could be found. It is obvious that there are
For clarity, we summarize the notations used throughout
similar entities for a certain type in a KG. For example, in a
this article in Table II. Then, the related definitions would
movie KG, there are similar movies or actors. With clustering
be given as follows for the purpose of better description
techniques, these similar entities can be clustered, and each
of task.
generated cluster can represent the set of entities that share
Definition 1 (User–Item Interaction Information): User–item
interaction information can be denoted as = (u, v) . Specif- similar features. In other words, these generated clusters can
A { } be considered as entities that represent items’ latent attributes,
ically, if a user u has interacted with an item v, then we use
such as the mixture of Action and Crime for the film genre,
(u, v) to denote an interaction between u and v. As a result,
A which does not exist in the original movie KG. For the
can be viewed as the set of all observed user–item
interactions. purpose of enriching KG, those above-mentioned clusters can
Definition 2 (KG): A KG is a kind of directed graph, which be added into the original KG to offer additional semantic
information. To our best of knowledge, it is the first time to
can be denoted as = ( , ). Specifically, = e1, e2, ..., em
G E R E { } add these clusters into KG. To better illustrate this scheme, we
represents the entity set, and = r1, r2, ..., rn
R { } describe this enhancement process in Fig. 2.
represents the relation set.
In the field of network mining, the nodes in the same cluster
Given the user–item interaction information and KG ,
A G constitute a community. To make the terminology clear and
the task is to predict the interaction probability y^
uv for the user
understandable, we adopt the terminology community to refer
u and item v, where v is the item that u has not interacted
to the node that would be added to the KG. The use of this
with before.
terminology indicates that this kind of nodes is relevant to the
result of the clustering. Next, we will give the definition of
IV. COMMUNITY ENHANCED KG FOR RECOMMENDATION community in this article.
A. Framework Definition 3 (Community): The community is the node to
S
be added into the KG, which can be regarded as an
In this section, the proposed method will be presented in ordinary entity in a KG. A community represents a cluster
detail. The proposed method consists of three parts: 1) KG of similar entities Cl in the original KG . In addition, there are
enhancement; 2) path representation learning; and 3) interac- two types of relations associated G
with community, namely the
tion probability prediction. The overall framework is relations rSe and reS between community and entity
illustrated in Fig. 1. belonging to the
In what follows, we will describe the proposed model community, and the relations rS S , rS S between commu-
l1 l2 l2 l1
in detail. nity l1 and l2 .
S S
As described above, a community represents a cluster Cl
S
B. KG Enhancement in . As a result, it is necessary to perform entity clustering
G
in the original KG . Actually, the KG can be considered as
As described above, the KG is usually with the problem the heterogeneous graph. G
The representations of nodes in a
of incompleteness and sparseness. If a KG is incomplete and heterogeneous graph can be learned by graph embedding
sparse, little semantic information is contained in it, which meth- ods, which can promote diverse downstream tasks such
may
ed licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on August 10,2024 at 05:42:33 UTC from
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the
exception of pagination.

as node

ed licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on August 10,2024 at 05:42:33 UTC from
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the
exception of pagination.

HE et al.: COMMUNITY ENHANCED KNOWLEDGE GRAPH FOR RECOMMENDATION 5

Fig. 2. Process of KG enhancement.

classification and clustering. Intuitively, heterogeneous graph Definition 4 (Enhanced KG): The enhanced KG GE is
embedding methods can be applied to obtain node representa- a special KG which is composed of the original KG G, the
tions s1, s 2 ,... , sm , which are then utilized to cluster enti- set of communities S = {S1 , S2 , ..., Sj }, relations rSe , reS ,
rSl1 Sl2 , and rSl2 Sl1 . Here, reS is a kind of relation indicating a
ties in {
a KG . Among}various kinds of heterogeneous graph
G certain node ek in G has the attribute that a certain
embedding methods, JUST [41] is a heterogeneous graph em- r eS
community Sl represents, which can be denoted as e →
bedding technique based on random walks. Different from S ; r represents k l
those random walks-based heterogeneous graph embedding
the relation indicating the cluster corresponding to
techniques which utilize metapaths to assist random walks, S Seincludes the node e , which can be denoted as S
r l
JUST does not need any metapaths. In most situations, due → e , where 1 ≤ k ≤k m and 1 ≤ l ≤ j. Besides, rSl1 Slk2
to the lack of domain knowledge, it is difficult to design and rSl2 Sl1 denote the relations between two communities Sl1
effective metapaths, which would degrade the performance and Sl2 , where 1 ≤ l1, l2 ≤ j and l1 l2.
}
of representation learning. As a result, heterogeneous ,For anyncluster
where i denotes 1 2
ei , eni ,

≤
embedding techniques without metapaths would be a better number of entities contained in Cli and 1 i j, it contains
choice. Therefore, in this article we choose JUST as the graph a certain type of entities. To enrich the original KG G
, we first
embedding method. add all communities S = 1, 2, ..., j into the KG . Then,
{S S S } G
With the help of JUST, the embeddings of entities emd = new relations should be involved in the graph. Specifically,
E for community Sl, the relation rSe and reS are added
s1, s 2 ,... , sm in the KG are learned. Based on the em-
{ }
beddings of entities, we then cluster the entities into between
clusters. Among many clustering methods, k-means is a Sl and each entity in Cll. Then, we consider adding relation
commonly used among communities. For communities,Sl1 and Sl2 , w,e can
clustering method [42]. The k-means algorithm splits the
, corresponding, clusters
find the
1 1
=n el1 , el1 , ...,
2
a
into clusters with the goal of having small difference within l2 l2 l2
Cl = e , e , ..., e . If there is an edge linking an entity
the l 1 2 n
same cluster and large difference between different clusters.
achieve this goal, sum of squared error (SSE) is adopted as the pair el1 , el2 , where el1 ∈ Cl and el2 ∈ Cl , we then add
l l
objective function, which is defined as
k q k q
Σ Σ 1 2
relations rSl1 Sl2 and rSl2 Sl1 between communities Sl1 and Sl2 .
SSE = D E (C i ,X) (1) With the above process, the enhanced KG GE is formed. Based
K
on GE , we then mine user preference for items, as will be
i=1 X∈Ci described later.
where Ci denotes the ith cluster center, X denotes a certain
sample, and DE denotes the Euclidean distance. Then, SSE
C. Path Representation Learning
needs to be minimized to achieve the above goal. Here, we
utilize k-means to cluster the entities based on their Since paths serve as the recommendation context, it is
embeddings learned by JUST. After the process of neces- sary to find paths concerning all the interactions in the
clustering, clusters are then obtained. Note that the KG is a user–item interaction set . For each user–item interaction (u, v)
kind of heterogeneous , we mine paths A connecting u and v in the enhanced ∈ KG.
,
The ob-
graph, so a cluster may contain various kinds of entities. As i ,
a result, we further divide each cluster into several subclusters, tained path set can be denoted as P (u,v) (u,v) .
= 1 ≤ i ≤ n ,
each of which is composed of entities of the same type. By (u
the above process, the final clusters are formed, which can be mapped onto a certain community, that is, Cli i, where
denoted as = Cl 1 ,Cl 2 , ..., Clj . Each cluster in can be 1 i j. Based on the concept of community, we give the
≤ C { } C
ed licensed use limited to: Vignan's Foundation for Science Technology
→ & Research (Deemed to be University). Downloaded on August 10,2024 at 05:42:33 UTC from
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the
exception of pagination.
where p i
is a path connecting u and v, and n is the path
definition of the new type of KG, namely enhanced KG. sampling size for the pair (u, v). For the reason that the entities
in the path connecting a certain user and item can be regarded
as a sequence, the RNN is naturally used to learn the repre-
sentation for a certain path. Among various kinds of network,
LSTM is a popular network architecture to model sequence

ed licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on August 10,2024 at 05:42:33 UTC from
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the
exception of pagination.

6 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS

data. The principle of LSTM can be formalized by the embedding size of entities. Based on the similarity weight, we
following equations: can obtain the path set representation H(u,v) for a certain user–
item interaction (u, v) as follows:
ft = σ (Wf · [ht−1 , xt ] + bf ) n
Σ
it = σ (Wi · [ht−1 , xt ] + bi ) H(u,v) = i l,
h(u,v) (4)
(u
˜ swi
Ct = tanh (WC · [ht−1 , xt ] + bC ) i=1

Ct = ft ∗ Ct−1 + it ∗ C˜t wher

i l,
(u,v) denote the path representation of path pi .
ot = σ (Wo · [ht−1, x t ]+ bo) After the above path representation aggregation, the (u repre-
sentation for path set P(u,v) is obtained. Since our purpose is
ht = ot ∗ tanh (Ct) (2) to predict interaction probability for user–item pairs, it is nec-
where ft, it, and ot represent the forget, input, and output gates; essary to map the path set representation H(u,v) to the value of
Ct˜and Ct denote candidate cell state and current cell state; probability. Generally, MLP is an effective tool to
ht and ht−1 are the final output of current step and last step; approximate any function. As a result, we adopt a two-layer
and xt is the input of current step. MLP to learn the mapping from the path set representation to
the interaction
For a path pi ∈ P(u,v), it can be denoted as an entity probability. We formulate this process as follows:
(u
sequence [e1, e 2 ,... , el−1, el], where e1 = u, el = v and l is the
y^uv = σ W2 ReLU W1 H(u,v) + b1 + b2 (5)
The entities
.number of entities
in contained in
are first represented by embedding(uvectors that are obtained
(u
where σ and ReLU denote sigmoid function and linear
∈ ∈
after the above-mentioned graph embedding process, and the recti- fication function, W1 R(dh/2)×dh and W2 R1×(dh/2)
above entity sequence can be transformed into the correspond- are the weight matrix of the first and second layer,
ing embedding sequence [s1, s 2 ,... , sl−1, sl]. Then, each respectively, b1 R(dh/2)×1 and b2 R represent the bias terms
entity embedding in the above embedding sequence is
∈ first and second∈layer, and dh denotes the embedding
for the
sequentially fed into LSTM. If the step begins at t = 1 for size of hidden state in LSTM. Here, (dh/2) is the dimension
LSTM, the input of LSTM can be represented as xt = st, of the hidden layer, following the design of tower pattern,
≤
where 1 t that is, halving the layer size for each successive higher layer
l. After the above process, the final output hl which can be [31]. In this way, we have obtained the predicted interaction
viewed as a path representation is obtained for the subsequent probability yuv. ^
prediction task.
E. Model Optimization
D. Interaction Probability Prediction
Since the recommendation task can be treated as a binary
As there are several paths related to a user–item interaction, classification problem [22], where an observed user–item in-
several path representations are obtained after the process of teraction is assigned 1 while 0 for an unobserved interaction,
path representation learning. Since our task is to predict the we adopt the binary cross-entropy (BCE) loss for the model
interaction probability, it is necessary to aggregate these rep- optimization, which is formulated as follows:
resentations appropriately.
Σ
A possible strategy for the path representation aggregation L =− (yuv log y^uv + (1 − yuv )log (1 − y^uv ))
is to merge them with weights. Intuitively, each path in a path (u,v)∈τ (6)
set that is associated with a certain user–item pair contributes where yuv and yuv^denote the actual implicit feedback (0 or
differently to the interacted item due to a user’s different 1) and the predicted interaction probability, respectively, and τ
aspects of preference for this item. Here, we assume that a denotes the training set. For clarity, Algorithm 1 summarizes
path will contribute more to the target item if the entities in the main procedure of the proposed CEKGR method.
this path
are more similar to the above item. For a path set P(u,v) =
,
,i .1 ≤ i ≤ n , a path in it can be denoted by an
(u,v) F. Time Complexity
p
embed-
ding sequence [s1, s 2 ,... , sl−1, sl]. As mentioned above, The time consumption of CEKGR mainly comes from
the KG enhancement and path representation learning. In the
similarity between entities in a path and the target item can KG enhancement, the computational complexity of JUST is
represent the importance of this path for recommendation to
O(r L(w + 1)), where r, , L, and w denote the number
some extent. Based on this, we define the similarity weight
of random walks for each entity, the number of entities in
sw, which can reflect the contribution of a certain path.
the |E| |
Concretely,
we formalize the similarity weight with respect to path pi
ed licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on August 10,2024 at 05:42:33 UTC from
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the
exception of pagination.

as follows: (u,v) KG, the length of random walk, and window size of Skip-
l 1 Gram, respectively. The time cost of k-means and the addition
Σ −
i T of communities and corresponding relations are O(k |E|) and
sw(u,v) sj v
j
O( |2 + k2), respectively, where k is the number of
where sj R di ×1
and v R di ×1
are the embedding vectors clusters. In the path representation learning, the time cost of
and calculation of path representation are O((|E| + B) |A|) and
for entity e∈ ∈
j and interacted item v, respectively, and di denotes
O(P |A|), respectively, where B, |A|, and P denote the number
the

ed licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on August 10,2024 at 05:42:33 UTC from
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the
exception of pagination.

HE et al.: COMMUNITY ENHANCED KNOWLEDGE GRAPH FOR RECOMMENDATION 7

Algorithm 1 Community Enhanced Knowledge Graph for TABLE III

STATISTICS OF DATASETS
Rec- ommendation (CEKGR)
Input: The knowledge graph G, the user-item interaction set A Datasets IM-1M Yelp Book
1: /* Knowledge graph enhancement */ #Users 943 37 940 17 860
= User-item
2: Learn node representations E
, s{s , s ,... } in G interaction
#Items 1675 11 516 14 967
#Ratings 99 975 229 178 69 873
with JUST. emd 1 2 m #Entities 7740 46 361 95 772
#Entity Types 5 4 16
3: Based on emd, obtain clusters = Cl 1 ,Cl 2 , ..., Clj by Knowledge
#Links 100 065 219 133 228 562
E C { } graph
clustering entities in with k-means. #Link Types 7 5 36
G Graph Density 0.1671% 0.0102% 0.0025%
4: Obtain enhanced knowledge graph E by adding commu-
G
nities and corresponding relations to .
G
5: /* Path representation learning and interaction probability
prediction */ A. Experimental Settings
6: = . /* Define an empty path set */
P ∅ 1) Datasets: To evaluate the proposed method, we adopt
7: for (u, v) do
∈ three widely used datasets: IM-1M, Yelp, and Book.
8: Obtain path set P(u,v) by mining paths linking u and v.
9: Add P(u,v) to . 1) IM-1M is constructed by combining the MovieLens 1M1
P and IMDB2 datasets. MovieLens 1M contains ratings,
10: end for
11: for epoch = 1 to n_epochs do
while IMDB provides movie auxiliary information in-
12: for batch do cluding director, actor, and genre.
⊂ 2) Yelp3 contains user check-in information and business
13: for P(u,v) batch do
∈ related information, such as city and genre.
14: Obtain n embedding sequences based on the entity
sequences derived from n paths in P(u,v). 3) Book4 contains the ratings of books in the Book-
15: Feed the above embedding sequences into LSTM Crossing community. As for the auxiliary information of
and obtain n path representations based on Eq. (2). the books, we directly utilize the triplets that are
provided by authors who conduct experiments on the
16: Calculate the similarity weights for n paths based
Book dataset [24].
on Eq. (3).
The statistics of these three datasets are listed in Table III.
17: Obtain the path set representation H(u,v) based on
Eq. (4). For the IM-1M and Book datasets, we split each dataset
18: Calculate the predicted interaction probability yuv into the training set and the testing set. Specifically, the
^ training set and the testing set account for 80% and 20% of
the original
based on Eq. (5). dataset, respectively. Particularly, the interactions on the IM-
19: end for 1M dataset are split in the order of their timestamps, that is,
20: Obtain the loss based on Eq. (6). the first 80% of interactions serve as the training data, and the
L
21: Update model parameters Θ based on . rest is for testing. As for the Book dataset, the interactions are
L
22: end for randomly split. For the Yelp dataset, we directly use the
23: end for training set and the testing set in the original published
Output: The model parameters Θ. version. To facilitate the model training, it is necessary to
utilize a negative sampling strategy. Specifically, we sample
an unobserved user–item in- teraction for each positive user–
item interaction in the training set for all the three datasets.
Since the time cost of finding paths
of edges in the KG, the number of user–item interaction, and at https://pan.baidu.com/s/1QHcMnLoJjaLDYpcgpRuGIg. The
the number of epochs, respectively. As a result, the time extraction code is tcss, and the unzipping password for zip file is
complexity of CEKGR is as follows: O(r |E| L(w + 1)+ k tcss2024.
2
|E| + |E| +
k2 + (|E| + B + P ) |A|).

V. EXPERIMENTS
In this section, extensive experiments will be conducted to
confirm the effectiveness of the proposed CEKGR method.
First of all, we will describe the experimental settings, in-
cluding datasets, evaluation measures, and baselines. Then,
we will report and analyze the comparison results. In addi-
tion, ablation study and parameter analysis will be conducted.
Finally, case study will be conducted. The code is available
ed licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on August 10,2024 at 05:42:33 UTC from
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the
exception of pagination.

from a certain user to a certain item in a KG is quite

high, it is also better to utilize a negative sampling strategy
in the testing set. Specifically, for each positive user–item
pair, we sample ten unobserved user–item interactions as
the negative samples for the Yelp and Book datasets,
considering the high- time cost of path search. As for the
IM-1M dataset, we only sample five negative samples,
because there are many positive user–item interactions for
many users, it is impossible to sample ten negative samples.
2) Evaluation Measures: For the purpose of evaluating
the proposed method, we conduct top-K recommendation in
our ex- periments. Generally, Recall@K and NDCG@K are
commonly used in the top-K recommendation, so the above
measures are

1
http://grouplens.org/datasets/movielens/
2
http://www.imdb.com/
3
https://www.kaggle.com/c/yelp-recsys-2013/data
4
http://www2.informatik.uni-freiburg.de/cziegler/BX/

ed licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on August 10,2024 at 05:42:33 UTC from
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the
exception of pagination.

8 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS

TABLE IV
TOP-K PERFORMANCE OBTAINED BY ALL THE METHODS ON IM-1M IN TERMS OF RECALL@K AND NDCG@K

Recall NDCG
Recall@1 Recall@5 Recall@10 Recall@20 NDCG@1 NDCG@5 NDCG@10 NDCG@20
RKGE 0.0567 0.2417 0.4024 0.5909 0.5589 0.5198 0.5382 0.5839
Ripple Net 0.0979 0.4000 0.5448 0.6919 0.9989 0.9447 0.9041 0.8829
KGAT 0.0764 0.3413 0.5239 0.6841 0.7476 0.7002 0.7111 0.7331
CKE 0.0771 0.3354 0.5148 0.6761 0.7423 0.6949 0.7030 0.7232
ECFKG 0.0620 0.2395 0.3597 0.5001 0.5546 0.4904 0.4888 0.5126
NFM 0.0771 0.3282 0.4999 0.6595 0.7434 0.6824 0.6838 0.7063
KGCL 0.0727 0.3341 0.5134 0.6786 0.7137 0.6912 0.7018 0.7257
CEKGR 0.0980 0.4685 0.6747 0.8394 0.9981 0.9601 0.9653 0.9705
Improvement percentage 0.10% 17.13% 23.67% 21.32% -0.08% 1.63% 6.77% 9.92%

TABLE V
TOP-K PERFORMANCE OBTAINED BY ALL THE METHODS ON YELP IN TERMS OF RECALL@K AND NDCG@K

Recall NDCG
Recall@1 Recall@5 Recall@10 Recall@20 NDCG@1 NDCG@5 NDCG@10 NDCG@20
RKGE 0.3242 0.6980 0.8426 0.9243 0.5010 0.6148 0.6622 0.6881
Ripple Net 0.6743 0.9078 0.9480 0.9779 0.9760 0.9663 0.9677 0.9710
KGAT 0.3191 0.6492 0.8212 0.9304 0.4861 0.5811 0.6339 0.6659
CKE 0.2523 0.5836 0.7976 0.9183 0.4009 0.5005 0.5679 0.6039
ECFKG 0.3409 0.7193 0.8642 0.9381 0.4986 0.6225 0.6694 0.6926
NFM 0.3671 0.7446 0.8872 0.9498 0.5359 0.6555 0.7019 0.7216
KGCL 0.3350 0.6563 0.8162 0.9292 0.5121 0.5967 0.6446 0.6769
CEKGR 0.6928 0.9430 0.9766 0.9931 0.9788 0.9791 0.9802 0.9811
Improvement percentage 2.74% 3.88% 3.02% 1.55% 0.29% 1.32% 1.29% 1.04%

adopted to assess the performance. Here, we set K = 1, 5, nonlinearity of neural network in modeling the higher
10, 20. order feature interactions.
3) Baselines: To validate the effectiveness of the pro- 7) KGCL [45]: This method takes inspirations from the KG
posed CEKGR method, we compare CEKGR with the follow- learning and self-supervised data augmentation, to in-
ing methods. corporate the KG context to guide the model in refin-
1) RKGE [22]: This method utilizes recurrent network to ing user/item representations with new knowledge-
model paths connecting entity pairs in a KG, and the aware contrastive objectives.
embeddings of entities are used to obtain proximity
score between user–item pairs.
2) RippleNet [24]: To represent users’ preference, this B. Comparison Results
model propagates users’ potential interests along the The comparison results on the three datasets are listed in
connection in a KG. In the end, the clicking probability Tables IV, V, and VI, respectively. For convenience, the best
is predicted based on the item representation and user results are boldfaced, and the row “improvement percentage”
representation. shows the performance improvement obtained by the pro-
3) KGAT [35]: This method propagates the embeddings posed method compared with the best performance of base-
from the current node’s neighbors in a KG, and lines. Based on the results in three tables, we have the follow-
attention mechanism is adopted for the purpose of ing observations.
distinguishing the importance of neighbor nodes. 1) Compared with the most similar baseline, namely
4) CKE [19]: This model obtains the item representation RKGE, which also utilizes paths as recommendation
from structural content, textual content, and visual con- context, the proposed CEKGR method has obtained
tent with respect to a certain knowledge base and the significant performance improvement in terms of recall,
item offset vector learned from historical user–item i.e., the average improvement percentage of 69.10%,
interac- tions. Then, the collaborative joint learning is 43.04%, and 187.73% on IM-1M, Yelp, and Book,
conducted to predict users’ implicit feedback. respectively. This may be due to the following reasons:
5) ECFKG [43]: Through the KG containing users and 1) RKGE uses paths’ representations in the training
items, the representations of users and items can be process, but it con- ducts recommendation with only the
learned by maximizing the generative probability of the representations of users and items that are refined by
observed relation triplets. Then, CF can be conducted to encoding the con- nected paths in the test process. In
predict users’ preference for items. contrast, CEKGR directly utilizes path representation
6) NFM [44]: This method combines the linearity of FM in for the test process; and 2) CEKGR has enriched the
modeling the second-order feature interactions and the KG for more semantic

ed licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on August 10,2024 at 05:42:33 UTC from
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the
exception of pagination.

HE et al.: COMMUNITY ENHANCED KNOWLEDGE GRAPH FOR RECOMMENDATION 9

TABLE VI
TOP-K PERFORMANCE OBTAINED BY ALL THE METHODS ON BOOK IN TERMS OF RECALL@K AND NDCG@K
is more effective than
Recall NDCG
Recall@1 Recall@5 Recall@10 Recall@20 NDCG@1 NDCG@5 NDCG@10 NDCG@20
RKGE 0.1837 0.3593 0.4012 0.4257 0.3368 0.3535 0.3610 0.3673
Ripple Net 0.1086 0.4839 0.7274 0.8657 0.1294 0.3144 0.3985 0.4436
KGAT 0.3022 0.5775 0.7686 0.8824 0.4019 0.4892 0.5509 0.5865
CKE 0.1937 0.4359 0.6929 0.8675 0.2855 0.3541 0.4356 0.4882
ECFKG 0.3382 0.6022 0.7914 0.8920 0.4765 0.5312 0.5900 0.6205
NFM 0.3382 0.6126 0.7957 0.8959 0.4692 0.5369 0.5940 0.6249
KGCL 0.2981 0.5486 0.7493 0.8754 0.3960 0.4668 0.5321 0.5712
CEKGR 0.7482 0.9523 0.9822 0.9952 0.9011 0.9156 0.9277 0.9325
Improvement percentage 121.23% 55.45% 23.44% 11.08% 89.11% 70.53% 56.18% 49.22%

information, which enhances the effectiveness of path

representation learning.
2) The embedding-based methods (i.e., CKE and ECFKG)
achieve better performance than RKGE in most cases.
In particular, both methods are better than RKGE on the
Book dataset. Specifically, the average recall improve-
ment percentage of CKE and ECFKG over RKGE on
the Book dataset is 50.81% and 89.63%, respectively.
Since the KG for the Book dataset is relatively sparse,
we can conclude that the embedding-based methods
would be better than the path-based methods when the
KG is with low graph density. The most possible reason
is that the paths connecting user–item pairs in the sparse
KG contain little semantic information.
3) The propagation-based methods (i.e., RippleNet and
KGAT) have better performance than the embedding-
based methods on the IM-1M dataset. However, the per-
formance of ECFKG is better than that of KGAT on
the Yelp and Book datasets with the average Recall im-
provement percentage of 5.93% and 5.06%,
respectively, and ECFKG performs better than
RippleNet on the Book dataset with the average Recall
improvement percentage of 61.93%. According to Table
III, the graph densities of KG for the Yelp and Book
datasets are both lower than that of the IM-1M dataset.
Since the propagation-based methods depend on the
high-order connection patterns in a KG, the sparsity of
graph may harm the propagation of the representations
of entities.
4) NFM performs better than RKGE on all the three
datasets, and particularly it outperforms the path-based,
embedding-based, and propagation-based methods on
the Book dataset with the minimum recall improvement
per- centage of 0.44%. As mentioned above, the Book
dataset is with a relatively sparse KG. With neural
network, NFM can model the higher order feature
interactions, which is effective for prediction under the
sparse settings.
5) KGCL outperforms RKGE and CKE in most of metrics,
which indicates that this method is superior to some
path- based and embedding-based methods. Compared
with the results of NFM, it is obvious that KGCL
outperforms NFM in some metrics on the IM-1M
dataset only. Since the Yelp and Book datasets are both
with relatively sparse KGs, it is most likely that NFM
ed licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on August 10,2024 at 05:42:33 UTC from
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the
exception of pagination.

the paradigm of contrastive learning in the KG-

based recommendation.
6) It is clear that the proposed CEKGR method is the
best in terms of all the measures on the three
datasets. For the values of recall, CEKGR improves
over the best baseline by the average percentage of
15.56%, 2.80%, and 52.80% on the IM-1M, Yelp,
and Book, respectively. The reason for this
phenomenon may be that the proposed method can
effectively learn the aggregated representation of
multipath for the certain user–item pair. Obviously,
on the Book dataset, CEKGR acquires more
improvement than on the other two datasets.
Actually, the KG for the Book dataset is relatively
sparse, that is, the graph density is relatively low.
Despite this situation, CEKGR still achieves
relatively high performance thanks to the addition of
communities to the original KG. We can con- clude
that the advantage of our approach is evident when
the original KG is relatively sparse.

C. Ablation Study
In this section, ablation study will be conducted to
demon- strate the effect of some key components of the
proposed method, namely community and aggregation
method.
1) Effectiveness Analysis on Community: The
comparison results have shown that CEKGR achieves
better performance than the path-based recommendation
method RKGE, which proves the effectiveness of the
introduction of community to a certain extent. To further
validate the effectiveness of the introduction of
community, we compare the proposed model with a
variant which does not introduce the communities into
original KG. We report the results in Table VII. Note
that the row “CEKGR-RS” shows the results of the variant
that is without communities, and the row “improvement
percentage” shows the performance improvement of the
complete model over the variant. Generally, CEKGR
achieves the higher recall on three datasets than the variant
which is without communities. In particular, on the IM-1M
dataset, CEKGR shows a small performance improvement
over the variant without communi- ties, and even the value
of Recall@1 decreases. Based on the statistics of datasets
in Table III, the original KG of the IM-1M dataset is
relatively dense, so the semantic information of it is
relatively abundant. Therefore, the performance
improvement of adding communities is relatively limited.
The possible reason

ed licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on August 10,2024 at 05:42:33 UTC from
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the
exception of pagination.

10 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS

TABLE VII
ABLATION STUDY: IMPACT OF COMMUNITY

Recall
Recall@1 Recall@5 Recall@10 Recall@20
CEKGR-RS 0.1048 0.4665 0.6685 0.8345
IM-1M CEKGR 0.0980 0.4685 0.6747 0.8394
Improvement percentage -6.49% 0.43% 0.93% 0.59%
CEKGR-RS 0.5022 0.7682 0.8613 0.9430
Yelp CEKGR 0.6928 0.9430 0.9766 0.9931
Improvement percentage 37.95% 22.75% 13.39% 5.31%
CEKGR-RS 0.5997 0.9067 0.9721 0.9901
Book CEKGR 0.7482 0.9523 0.9822 0.9952
Improvement percentage 24.76% 5.03% 1.04% 0.52%
Note: Bold results in each column are the best.

TABLE VIII
ABLATION STUDY: IMPACT OF AGGREGATION METHOD

Recall
Recall@1 Recall@5 Recall@10 Recall@20
Max pooling 0.0980 0.4685 0.6747 0.8394
IM-1M Average pooling 0.0980 0.4685 0.6747 0.8394
Similarity weight 0.0980 0.4685 0.6747 0.8394
Max pooling 0.6928 0.9430 0.9766 0.9931
Yelp Average pooling 0.6928 0.9430 0.9766 0.9931
Similarity weight 0.6928 0.9430 0.9766 0.9931
Max pooling 0.3586 0.7950 0.8995 0.9509
Book Average pooling 0.3420 0.7423 0.8669 0.9323
Similarity weight 0.7482 0.9523 0.9822 0.9952
Note: Bold results in each column are the best.

for the decrease of Recall@1 is that the increase of semantic pooling. As revealed in the literature [46], there is a tradeoff
information is limited after adding communities to the KG of between interpretability and recommendation effect. Specifi-
IM-1M, and when the length of the list of candidate recom- cally, the recommendation performance may degrade when
mendations is too short (e.g., when the length is 1), the value the interpretability is improved. However, the results show
of recall is more susceptible to unobserved interactions. While that our model becomes more explainable without any
on the Yelp and Book datasets, the improvement of Recall is sacrifice of per- formance, which demonstrates the superiority
obvious. As shown in Table III, these two datasets are both of our proposed aggregation method. In a word, the utilization
with a relatively sparse original KG. Therefore, there is of similarity weight can bring improvement in
relatively little semantic information in KGs. As a result, the recommendation effect when the KG is relatively sparse.
introduction of communities can enhance the semantic Although there is no performance improvement with a
information to a greater extent, so the performance relatively dense KG when using similarity weight, the
improvement will be more obvious. In summary, the results proposed aggregation method can distinguish the importance
have validated the effective- ness of introducing communities, of each path without compromising performance.
especially for the KG with low graph density.
2) Effectiveness Analysis on Aggregation Method: Consid- D. Parameter Analysis
ering that there are several paths between a certain user–item
In this section, parameter analysis will be conducted to
pair, so several path representations should be appropriately
show how some parameters affect the performance of the
aggregated for the following interaction probability
proposed method, namely, path length, path sample size,
prediction. In this part, we compare our method with two
embedding size, and the number of clusters. When analyzing
variants that aggregate path representations by average
one parameter, the other parameters are kept fixed as the
pooling and max pooling, respectively. From the results in
default setting.
Table VIII, it can be seen that the recall value obtained by the
1) Impact of Path Length: When mining paths connecting
proposed method are higher than those of two variants on the
user–item pairs, the length of paths can play an important role
Book dataset. It can also be observed that the recall value on
in the prediction of users’ preference. To study the effect of
the other two datasets keep the same among three kinds of
path length on the recommendation performance, we vary the
aggregation methods, which shows no improvement on the
path length to see how the performance changes with it. Due
other two aggregation methods. However, note that either the
to the huge time complexity consumed by the path-mining
average pooling or max pool- ing cannot indicate the
task in a KG, we restrict the path length to 5. Specifically, the
importance of each path. The method with similarity weight
path length is set to 3 and 5. Since the number of interactions
distinguishes the diverse contribution of each path for a
and the scale of the KG are both relatively large for the Yelp
certain user–item pair without any perfor- mance degradation
dataset, it would take quite a long time to find paths for all the
compared with the average pooling or max
interactions on the Yelp dataset when the path length is set to
5, of which the cost is too high to be acceptable. As a result,
ed licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on August 10,2024 at 05:42:33 UTC from
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the
exception of pagination.

we adopt the IM-1M

ed licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on August 10,2024 at 05:42:33 UTC from
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the
exception of pagination.

HE et al.: COMMUNITY ENHANCED KNOWLEDGE GRAPH FOR RECOMMENDATION 11

TABLE IX TABLE XI
PARAMETER ANALYSIS: THE IMPACT OF PATH LENGTH PARAMETER ANALYSIS: THE IMPACT OF THE EMBEDDING SIZE

Recall Recall
Recall@1 Recall@5 Recall@10 Recall@20 Recall@1 Recall@5 Recall@10 Recall@20
L3 0.0980 0.4685 0.6747 0.8394 8 0.0980 0.4685 0.6746 0.8394
IM-1M L5 0.0979 0.4684 0.6746 0.8394 16 0.0982 0.4687 0.6748 0.8394
L3 0.7482 0.9523 0.9822 0.9952 IM-1M 32 0.0980 0.4685 0.6747 0.8394
Book L5 0.7561 0.9568 0.9840 0.9958 64 0.0982 0.4685 0.6747 0.8395
128 0.0978 0.4679 0.6743 0.8392
Note: Bold results in each column are the best.
8 0.7464 0.9535 0.9828 0.9953
16 0.7454 0.9506 0.9812 0.9948
TABLE X Book 32 0.7482 0.9523 0.9822 0.9952
PARAMETER ANALYSIS: THE IMPACT OF PATH SAMPLE SIZE 64 0.7429 0.9511 0.9810 0.9944
128 0.7490 0.9529 0.9825 0.9954
Recall
Note: Bold results in each column are the best.
Recall@1 Recall@5 Recall@10 Recall@20
S2 0.0980 0.4685 0.6747 0.8394
IM-1M S5 0.0980 0.4685 0.6747 0.8394 TABLE XII
S8 0.0980 0.4685 0.6747 0.8394 PARAMETER ANALYSIS: THE IMPACT OF THE NUMBER OF CLUSTERS
S2 0.7479 0.9523 0.9822 0.9952
Book S5 0.7482 0.9523 0.9822 0.9952 Recall
S8 0.7482 0.9523 0.9822 0.9952 Recall@1 Recall@5 Recall@10 Recall@20
10 0.0979 0.4684 0.6746 0.8394
Note: Bold results in each column are the best. 50 0.0979 0.4684 0.6746 0.8394
IM-1M 100 0.0980 0.4685 0.6747 0.8394
and Book datasets for the parameter analysis. The results are 500 0.0993 0.4668 0.6724 0.8380
1000 0.1014 0.4648 0.6700 0.8362
shown in Table IX, where the row “L3” denotes that the path 10 0.3830 0.8482 0.9346 0.9783
length is 3 and the other rows are with similar meanings. For 50 0.3922 0.6371 0.8168 0.9034
the IM-1M dataset, we can observe that the performance Book 100 0.7482 0.9523 0.9822 0.9952
500 0.6773 0.9294 0.9709 0.9884
remains almost unchanged. However, the situation is different 1000 0.6602 0.9259 0.9686 0.9873
on the Book dataset. As shown in Table IX, the performance
Note: Bold results in each column are the best.
increases when the path length increases. The reason why
the situation
is different on these two datasets is that the graph density of size can be set to 8.
KG of the IM-1M dataset is relatively large, which provides
rich semantic information, so a shorter path would not bring
performance degradation. Nevertheless, the KG of the Book
dataset is so sparse that it contains little information. As a result,
the path should be longer for the purpose of completely
mining user preference.
2) Impact of the Path Sample Size: Since there are many
paths linking a user and an item in a KG, the time cost for
model training would be quite high if we utilize all paths
between users and items. As a result, the strategy of
sampling is necessary. In this part, we will study the impact
of sampling size on rec- ommendation performance. The
results are shown in Table X, where the row “S2” denotes that
the path sample size is 2 and the other rows are with similar
meanings. Obviously, the results stay almost unchanged as the
sample size changes. This fact can confirm that our method is
insensitive to the path sample size, which would be helpful to
reduce the time and space cost for the model training process
since we can adopt a small sample size without performance
degradation.
3) Impact of the Embedding Size: In this part, we inves-
tigate the impact of the embedding size on recommendation
performance. The results are shown in Table XI. For the IM-
1M dataset, it can be seen that the values of recall increase at
first and then decrease as the embedding size rises. The
possible reason is that the model with too small dimension
cannot encode more features, and too large dimension leads to
overfitting. As for the Book dataset, the recommendation
performance is quite fine when the embedding size is 8, 32,
and 128. To reduce the cost of model training, the embedding
ed licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on August 10,2024 at 05:42:33 UTC from
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the
exception of pagination.

4) Impact of the Number of Clusters: In the

process of KG enhancement, we utilize k-means to cluster
entities. Since the number of communities is relevant to the
number of clusters, we investigate the impact of the number
of clusters on recommendation performance. The results
are shown in Table XII. From the results on the two
datasets, we can observe that the recommendation
performance is relatively fine when the number of clusters k
is set to 100. When k is too small, there would be few
communities to be added to the original KG, so the
additionally introduced semantic information may be little.
In this situation, the performance would be worse than
the scenario with k = 100. While k is much larger, it is
possible to introduce much noise to the original KG, so the
recommendation performance may degrade.

E. Case Study
By mining the connection between a user and an item, the
proposed method yields a better performance compared
with other state-of-the-art methods. Besides, the paths can
provide explanations why the recommender system
recommends a cer- tain product or service to a user. To
better illustrate it, we randomly select a user–item pair from
the IM-1M dataset. We present the paths between this user–
item pair in Fig. 3. From the figure, we can see that there are
three paths linking the user u941 and the movie Heat, whose
genre covers Action, Crime, and Drama. First, the lower
two paths share the same linking pattern. Based on these
two paths, explanation can be generated as “The movie Heat
is recommended to you since you have watched the movie
Face/Off and Star Wars: Return of the

ed licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on August 10,2024 at 05:42:33 UTC from
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the
exception of pagination.

12 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS

[2] Z.-R. Yang, Z.-Y. He, C.-D. Wang, J.-H. Lai, and Z. Tian,
“Collaborative meta-path modeling for explainable recommendation,”
IEEE Trans. Comput. Social Syst., vol. 11, no. 2, pp. 1805–1815, Apr.
2024.
[3] P. Resnick, N. Iacovou, M. Suchak, P. Bergstrom, and J. Riedl, “Grou-
pLens: An open architecture for collaborative filtering of Netnews,” in
Proc. Conf. Comput. Supported Cooperative Work (CSCW), Chapel
Hill, NC, USA, Oct. 22–26, 1994, pp. 175–186.
[4] U. Shardanand and P. Maes, “Social information filtering: Algorithms
for automating ‘word of mouth’,” in Proc. Human Factors Comput.
Syst. (CHI) Conf. Proc., Denver, CO, USA, May 7–11, 1995, pp. 210–
217.
[5] B. M. Sarwar, G. Karypis, J. A. Konstan, and J. Riedl, “Item-based
collaborative filtering recommendation algorithms,” in Proc. 10th Int.
Fig. 3. Paths connecting the user u941 and the movie Heat. World Wide Web Conf. (WWW), Hong Kong, China, May 1–5, 2001,
pp. 285–295.
[6] X. He, H. Zhang, M.-Y. Kan, and T.-S. Chua, “Fast matrix factorization
Jedi, which are both Action films.” As for the path on the top, for online recommendation with implicit feedback,” in Proc. 39th Int.
the community s100 is included in this path. According to the ACM SIGIR Conf. Res. Develop. Inf. Retrieval, Pisa, Italy, Jul. 17–21,
definition of community, s100 represents a cluster of a certain 2016, pp. 549–558.
[7] C.-D. Wang, W.-D. Xi, L. Huang, Y.-Y. Zheng, Z.-Y. Hu, and J.-H. Lai,
kind of users. In the scenario of movie recommendation, s100 “A BP neural network based recommender framework with attention
can reflect a certain user group that shares similar preference mechanism,” IEEE Trans. Knowl. Data Eng., vol. 34, no. 7, pp. 3029–
for movies. To better illustrate this, we investigate the 3043, Jul. 2022.
[8] X. He, K. Deng, X. Wang, Y. Li, Y. Zhang, and M. Wang, “LightGCN:
interacted movies for u941 and u42. For user u941, we find Simplifying and powering graph convolution network for recommenda-
that more than a quarter of interacted movies belong to the tion,” in Proc. 43rd Int. ACM SIGIR Conf. Res. Develop. Inf. Retrieval,
genre of Drama, while over a half of interacted movies are Virtual Event, China, Jul. 25–30, 2020, pp. 639–648.
[9] M. Mao, J. Lu, G. Zhang, and J. Zhang, “Multirelational social rec-
with the genre of Drama for u42. The above observation ommendations via multigraph ranking,” IEEE Trans. Cybern., vol. 47,
implies their common preference for the movies with the no. 12, pp. 4049–4061, Dec. 2017.
genre of Drama. As a result, the explanation based on this [10] J. Ni, Z. Huang, C. Yu, D. Lv, and C. Wang, “Comparative
path can be generated as follows, “We recommend the movie convolutional dynamic multi-attention recommendation model,” IEEE
Trans. Neural Netw. Learn. Syst., vol. 33, no. 8, pp. 3510–3521, Aug.
Heat to you because another user who shares common interest 2022.
with you has watched it.” In the end, we can see that different [11] Q. Zhang, J. Lu, D. Wu, and G. Zhang, “A cross-domain recommender
factors contribute diversely for user’s interacted movies. The system with kernel-induced knowledge transfer for overlapping
difference can be reflected by the weight of the corresponding entities,” IEEE Trans. Neural Netw. Learn. Syst., vol. 30, no. 7, pp.
1998–2012, Jul. 2019.
path. From the weight of each path, we can see that the reason [12] K. Liu, F. Xue, X. He, D. Guo, and R. Hong, “Joint multi-grained
for u941 to choose movies tend to consider the movies with popularity-aware graph convolution collaborative filtering for recom-
the genre of Action rather than the selecting movies of similar mendation,” IEEE Trans. Comput. Social Syst., vol. 10, no. 1, pp. 72–
83, Feb. 2023.
users. [13] L. Sang, M. Xu, S. Qian, and X. Wu, “Knowledge graph enhanced
neural collaborative recommendation,” Expert Syst. Appl., vol. 164,
VI. CONCLUSION 2021, Art. no. 113992.
[14] Q. Zhu, X. Zhou, J. Wu, J. Tan, and L. Guo, “A knowledge-aware
KG-based recommender systems have shown to achieve the attentional reasoning network for recommendation,” in Proc. 34th
fairly fine performance compared with other kinds of recom- AAAI Conf. Artif. Intell. (AAAI), 32nd Innovative Appl. Artif. Intell.
mender systems with the help of KGs containing rich Conf. (IAAI), 10th AAAI Symp. Educational Adv. Artif. Intell. (EAAI),
New York, NY, USA, Feb. 7–12, 2020, pp. 6999–7006.
semantic information about items. Because of this, the quality [15] K. Tu et al., “Conditional graph attention networks for distilling and
of KG plays an important role in recommendation accuracy of refining knowledge graphs in recommendation,” in Proc. 30th ACM
KG- based recommender systems. For addressing the issues of Int. Conf. Inf. Knowl. Manage. (CIKM), Virtual Event, Queensland,
Australia, Nov. 1–5, 2021, pp. 1834–1843.
in- completeness and sparseness in the KG, we enrich the KG [16] H. Sheu, Z. Chu, D. Qi, and S. Li, “Knowledge-guided article em-
with entities’ clusters and the corresponding relations. To bedding refinement for session-based news recommendation,” IEEE
clearly distinguish the importance of different paths which Trans. Neural Networks Learn. Syst., vol. 33, no. 12, pp. 7921–7927,
Dec. 2022.
reflect the corresponding factors that impact users’ decision,
[17] C. Chen, M. Zhang, Y. Liu, and S. Ma, “Neural attentional rating
we devise a weight mechanism to aggregate those paths, and regression with review-level explanations,” in Proc. World Wide Web
weights themselves can provide better explanations. Based on Conf. (WWW), Lyon, France, Apr. 23–27, 2018, pp. 1583–1592.
the above goals, we propose a recommendation model named [18] Z. Ren, S. Liang, P. Li, S. Wang, and M. de Rijke, “Social collaborative
viewpoint regression with explainable recommendations,” in Proc. 10th
CEKGR. Experiments on three datasets have been conducted, ACM Int. Conf. Web Search Data Mining (WSDM), Cambridge, U.K.,
and the results show the superiority of our method over other Feb. 6–10, 2017, pp. 485–494.
state- of-the-art methods. Furthermore, the case study has [19] F. Zhang, N. J. Yuan, D. Lian, X. Xie, and W.-Y. Ma, “Collaborative
knowledge base embedding for recommender systems,” in Proc. 22nd
shown the interpretability of the proposed method. ACM SIGKDD Int. Conf. Knowl. Discovery Data Mining, San
Francisco, CA, USA, Aug. 13–17, 2016, pp. 353–362.
REFERENCES [20] M. Jamali and M. Ester, “A matrix factorization technique with trust
propagation for recommendation in social networks,” in Proc. ACM
[1] O. Tal, Y. Liu, J. X. Huang, X. Yu, and B. Aljbawi, “Neural attention
Conf. Recommender Syst. (RecSys), Barcelona, Spain, Sep. 26–30, 2010,
frameworks for explainable recommendation,” IEEE Trans. Knowl.
Data Eng., vol. 33, no. 5, pp. 2137–2150, May 2021. pp. 135–142.
[21] Y. Lin, Z. Liu, M. Sun, Y. Liu, and X. Zhu, “Learning entity and

ed licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on August 10,2024 at 05:42:33 UTC from
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the
exception of pagination.
relation embeddings for knowledge graph completion,” in Proc. 29th
AAAI Conf. Artif. Intell., Jan. 25–30, Austin, TX, USA, 2015, pp.
2181–2187.

ed licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on August 10,2024 at 05:42:33 UTC from
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the
exception of pagination.

HE et al.: COMMUNITY ENHANCED KNOWLEDGE GRAPH FOR RECOMMENDATION 13

sary?: Revisiting heterogeneous graph embeddings,” in Proc. 27th ACM

[22] Z. Sun, J. Yang, J. Zhang, A. Bozzon, L.-K. Huang, and C. Xu,

“Recurrent knowledge graph embedding for effective recommendation,”
in Proc. 12th ACM Conf. Recommender Syst. (RecSys), Vancouver, BC,
Canada, Oct. 2–7, 2018, pp. 297–305.
[23] X. Wang, D. Wang, C. Xu, X. He, Y. Cao, and T.-S. Chua, “Explainable
reasoning over knowledge graphs for recommendation,” in Proc. 33rd
AAAI Conf. Artif. Intell. (AAAI), 31st Innovative Appl. Artif. Intell. Conf.
(IAAI), 9th AAAI Symp. Educational Adv. Artif. Intell. (EAAI),
Honolulu, HI, USA, Jan. 27–Feb. 1, 2019, pp. 5329–5336.
[24] H. Wang et al., “RippleNet: Propagating user preferences on the knowl-
edge graph for recommender systems,” in Proc. 27th ACM Int. Conf.
Inf. Knowl. Manage. (CIKM), Torino, Italy, Oct. 22–26, 2018, pp. 417–
426.
[25] J. Messner, R. Abboud, and I. I. Ceylan, “Temporal knowledge graph
completion using box embeddings,” in Proc. 36th AAAI Conf. Artif.
Intell. (AAAI), 34th Conf. Innovative Appl. Artif. Intell. (IAAI), 12th
Symp. Educational Adv. Artif. Intell. (EAAI), Virtual Event, Feb. 22–
Mar. 1, 2022, pp. 7779–7787.
[26] Z. Cao, Q. Xu, Z. Yang, and Q. Huang, “ER: Equivariance regularizer
for knowledge graph completion,” in Proc. 36th AAAI Conf. Artif. Intell.
(AAAI), 34th Conf. Innovative Appl. Artif. Intell. (IAAI), 12th Symp.
Educational Adv. Artif. Intell. (EAAI), Virtual Event, Feb. 22–Mar. 1,
2022, pp. 5512–5520.
[27] Y. Shen, D. Li, and D. Nan, “Modeling path information for knowledge
graph completion,” Neural Comput. Appl., vol. 34, no. 3, pp. 1951–
1961, 2022.
[28] K. D. Bollacker, C. Evans, P. K. Paritosh, T. Sturge, and J. Taylor,
“Freebase: A collaboratively created graph database for structuring
human knowledge,” in Proc. ACM SIGMOD Int. Conf. Manage. Data,
Vancouver, BC, Canada, Jun. 10–12, 2008, pp. 1247–1250.
[29] R. West, E. Gabrilovich, K. Murphy, S. Sun, R. Gupta, and D. Lin,
“Knowledge base completion via search-based question answering,” in
Proc. 23rd Int. World Wide Web Conf. (WWW), Seoul, Republic of
Korea, Apr. 7–11, 2014, pp. 515–526.
[30] Y. Koren, “Factorization meets the neighborhood: A multifaceted col-
laborative filtering model,” in Proc. 14th ACM SIGKDD Int. Conf.
Knowl. Discovery Data Mining, Las Vegas, NV, USA, Aug. 24–27, 2008,
pp. 426–434.
[31] X. He, L. Liao, H. Zhang, L. Nie, X. Hu, and T.-S. Chua, “Neural
collaborative filtering,” in Proc. 26th Int. Conf. World Wide Web
(WWW), Perth, Australia, Apr. 3–7, 2017, pp. 173–182.
[32] X. Wang, X. He, M. Wang, F. Feng, and T.-S. Chua, “Neural graph
collaborative filtering,” in Proc. 42nd Int. ACM SIGIR Conf. Res.
Develop. Inf. Retrieval, Paris, France, Jul. 21–25, 2019, pp. 165–174.
[33] S. Dai, Y. Yu, H. Fan, and J. Dong, “Spatio-temporal representation
learning with social tie for personalized POI recommendation,” Data
Sci. Eng., vol. 7, no. 1, pp. 44–56, 2022.
[34] C.-D. Wang, Z.-H. Deng, J.-H. Lai, and P. S. Yu, “Serendipitous
recommendation in e-commerce using innovator-based collaborative
filtering,” IEEE Trans. Cybern., vol. 49, no. 7, pp. 2678–2692, Jul.
2019.
[35] X. Wang, X. He, Y. Cao, M. Liu, and T.-S. Chua, “KGAT: Knowledge
graph attention network for recommendation,” in Proc. 25th ACM
SIGKDD Int. Conf. Knowl. Discovery Data Mining (KDD), Anchorage,
AK, USA, Aug. 4–8, 2019, pp. 950–958.
[36] H. Wang, M. Zhao, X. Xie, W. Li, and M. Guo, “Knowledge graph
convolutional networks for recommender systems,” in Proc. World
Wide Web Conf. (WWW), San Francisco, CA, USA, May 13–17, 2019,
pp. 3307–3313.
[37] X. Wang et al., “Learning intents behind interactions with knowledge
graph for recommendation,” in Proc. Web Conf. (WWW), Virtual Event,
Ljubljana, Slovenia, Apr. 19–23, 2021, pp. 878–887.
[38] Y. Yang, C. Huang, L. Xia, and C. Huang, “Knowledge graph self-
supervised rationalization for recommendation,” in Proc. 29th ACM
SIGKDD Conf. Knowl. Discovery Data Mining (KDD), Long Beach,
CA, USA, Aug. 6–10, 2023, pp. 3046–3056.
[39] Y. Lin, P. Ren, Z. Chen, Z. Ren, J. Ma, and M. de Rijke, “Explainable
outfit recommendation with joint outfit matching and comment gener-
ation,” IEEE Trans. Knowl. Data Eng., vol. 32, no. 8, pp. 1502–1516,
Aug. 2020.
[40] H. Chen, Y. Li, X. Sun, G. Xu, and H. Yin, “Temporal meta-path
guided explainable recommendation,” in Proc. 14th ACM Int. Conf.
Web Search Data Mining (WSDM), Virtual Event, Israel, Mar. 8–12,
2021, pp. 1056–1064.
[41] R. Hussein, D. Yang, and P. Cudré-Mauroux, “Are meta-paths neces-

ed licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on August 10,2024 at 05:42:33 UTC from
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the
exception of pagination.

Int. Conf. Inf. Knowl. Manage. (CIKM), Torino, Italy, Oct. 22–26,
2018, pp. 437–446.
[42] J. MacQueen, “Some methods for classification and analysis of mul-
tivariate observations,” in Proc. 15th Berkeley Symp. Math. Stat.
Probab., vol. 1. Berkeley, CA, USA: Univ. of California Press, 1967,
pp. 281–297.
[43] Q. Ai, V. Azizi, X. Chen, and Y. Zhang, “Learning heterogeneous
knowl- edge base embeddings for explainable recommendation,”
Algorithms, vol. 11, no. 9, p. 137, 2018.
[44] X. He and T.-S. Chua, “Neural factorization machines for sparse
predictive analytics,” in Proc. 40th Int. ACM SIGIR Conf. Res.
Dev. Inf. Retrieval, Shinjuku, Tokyo, Japan, Aug. 7–11, 2017, pp.
355–364.
[45] Y. Yang, C. Huang, L. Xia, and C. Li, “Knowledge graph contrastive
learning for recommendation,” in Proc. 45th Int. ACM SIGIR
Conf. Res. Dev. Inf. Retrieval (SIGIR), Madrid, Spain, Jul. 11–15,
2022, pp. 1434–1443.
[46] Y. Zhang and X. Chen, “Explainable recommendation: A survey and
new perspectives,” Found. Trends Inf. Retrieval, vol. 14, no. 1, pp. 1–
101, 2020.

Zhen-Yu He received the bachelor’s and

master’s degrees in computer science from South
China Agricultural University, Guangzhou,
China, in 2018 and 2021, respectively. He is
currently working toward the Ph.D. degree in
computer science with Sun Yat-sen University,
Guangzhou, China.
His current research interest includes data mining.

Chang-Dong Wang (Senior Member, IEEE) re-

ceived the Ph.D. degree in computer science
from Sun Yat-sen University, Guangzhou,
China, in 2013.
He joined Sun Yat-sen University in 2013,
where he is currently an Associate Professor with
the School of Computer Science and Engineer-
ing. His current research interests include ma-
chine learning and data mining. He has published
over 80 scientific papers in international journals
and conferences such as IEEE TRANSACTIONS
ON
PATTERN ANALYSIS AND MACHINE INTELLIGENCE, IEEE
TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, IEEE
TRANSACTIONS ON CYBERNETICS, IEEE TRANSACTIONS ON NEURAL
NETWORKS AND LEARNING
SYSTEMS, ACM SIGKDD Conference on Knowledge Discovery and Data
Mining (KDD), AAAI Conference on Artificial Intelligence (AAAI), In-
ternational Joint Conference on Artificial Intelligence (IJCAI), IEEE/CVF
Computer Vision and Pattern Recognition Conference (CVPR), IEEE
Interna- tional Conference on Data Mining (ICDM), ACM International
Conference on Information and Knowledge Management (CIKM), and
SIAM International Conference on Data Mining (SDM).
Dr. Wang won the Honorable Mention for Best Research Paper Awards
at ICDM 2010. He won the 2012 Microsoft Research Fellowship
Nomination Award. He was awarded 2015 Chinese Association for
Artificial Intelligence (CAAI) Outstanding Dissertation. He is an Associate
Editor of Journal of Artificial Intelligence Research (JAIR) and Neural
Networks.

Jinfeng Wang (Member, IEEE) received the

Ph.D. degree in computer science from the
Chinese Uni- versity of Hong Kong, Hong Kong,
China, in 2010. She is currently an Associate
Professor with the College of Mathematics and
Informatics, South China Agricultural
University, Guangzhou, China.
Her research interests include machine learning
and data mining.

ed licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on August 10,2024 at 05:42:33 UTC from
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the
exception of pagination.

14 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS

Jian-Huang Lai (Senior Member, IEEE) received Yong Tang received the B.S. and M.Sc. degrees
the M.Sc. degree in applied mathematics and the from Wuhan University, Wuhan, China, in 1985
Ph.D. degree in mathematics from Sun Yat-sen and 1990, respectively, and the Ph.D. degree from
University, Guangzhou, China, in 1989 and 1999, the University of Science and Technology of China,
respectively. Hefei, China, in 2001, all in computer science.
In 1989, he joined Sun Yat-sen University as He is the founder of SCHOLAT, a kind of
an Assistant Professor, where he is currently a scholar social network. He is currently a Professor
Professor with the School of Computer Science and the Dean of the School of Computer Science,
and Engineering. His current research interests in- South China Normal University, Guangzhou,
clude the areas of digital image processing, pattern China. Be- fore joining the South China Normal
recognition, multimedia communication, wavelet, University in 2009, he was the Vice Dean of the
and its applications. He has published more than 200 scientific papers in School of
the international journals and conferences on image processing and pattern Information of Science and Technology, Sun Yat-Sen University, Guangzhou,
recognition, such as IEEE TRANSACTIONS ON PATTERN ANALYSIS AND China. He has published more than 200 papers and books. He has supervised
MACHINE INTELLIGENCE, IEEE TRANSACTIONS ON KNOWLEDGE AND DATA more than 40 Ph.D. students since 2003 and more than 100 master’s students
ENGINEERING, IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING since 1996. His research interests include the area of data and knowledge
SYSTEMS, IEEE TRANSACTIONS ON IMAGE PROCESSING, IEEE engineering, social networking, and collaborative computing.
TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, PART B, Pattern Dr. Tang currently serves as the Director of Technical Committee on Col-
Recognition, In- laborative Computing at China Computer Federation (CCF) and the Executive
ternational Conference on Computer Vision (ICCV), CVPR, IJCAI, ICDM, Vice President of Guangdong Computer Academy. For more information,
and SDM. please visit https://scholat.com/ytang.
Prof. Lai serves as a Standing Member of the Image and Graphics
Association of China, and as the Standing Director of the Image and Graphics
Association of Guangdong.

ed licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on August 10,2024 at 05:42:33 UTC from