Exploiting Structural and Temporal Influence for Dynamic Social-Aware Recommendation

Abstract

Recent years have witnessed the rapid development of online social platforms, which effectively support the business intelligence and provide services for massive users. Along this line, large efforts have been made on the social-aware recommendation task, i.e., leveraging social contextual information to improve recommendation performance. Most existing methods have treated social relations in a static way, but the dynamic influence of social contextual information on users' consumption choices has been largely unexploited. To that end, in this paper, we conduct a comprehensive study to reveal the dynamic social influence on users' preferences, and then we propose a deep model called Dynamic Social-Aware Recommender System (DSRS) to integrate the users' structural and temporal social contexts to address the dynamic social-aware recommendation task. DSRS consists of two main components, i.e., the social influence learning (SIL) and dynamic preference learning (DPL). Specifically, in the SIL module, we arrange social graphs in a sequential order and borrow the power of graph convolution networks (GCNs) to learn social context. Moreover, we design a structural-temporal attention mechanism to discriminatively model the structural social influence and the temporal social influence. Then, in the DPL part, users' individual preferences are learned dynamically by recurrent neural networks (RNNs). Finally, with a prediction layer, we combine the users' social context and dynamic preferences to generate recommendations. We conduct extensive experiments on two real-world datasets, and the experimental results demonstrate the superiority and effectiveness of our proposed model compared with the state-of-the-art methods.

Full text

Liu Y, Li Z, Huang W et al. Exploiting structural and temporal influence for dynamic social-aware recommendation.
JOURNAL OF COMPUTER SCIENCE AND TECHNOLOGY 35(2): 281–294 Mar. 2020. DOI 10.1007/s11390-020-9956-9
Exploiting Structural and Temporal Influence for Dynamic
Social-Aware Recommendation
Yang Liu, Zhi Li, Wei Huang, Tong Xu∗, and En-Hong Chen, Fellow,CCF,Senior Member,IEEE
Anhui Province Key Laboratory of Big Data Analysis and Application, University of Science and Technology of China
Hefei 230027, China
E-mail: {ly0330, zhili03, ustc0411}@mail.ustc.edu.cn; {tongxu, cheneh}@ustc.edu.cn
Received August 20, 2019; revised December 22, 2019.
Abstract Recent years have witnessed the rapid development of online social platforms, which effectively support the
business intelligence and provide services for massive users. Along this line, large efforts have been made on the social-
aware recommendation task, i.e., leveraging social contextual information to improve recommendation performance. Most
existing methods have treated social relations in a static way, but the dynamic influence of social contextual information on
users’ consumption choices has been largely unexploited. To that end, in this paper, we conduct a comprehensive study to
reveal the dynamic social influence on users’ preferences, and then we propose a deep model called Dynamic Social-Aware
Recommender System (DSRS) to integrate the users’ structural and temporal social contexts to address the dynamic social-
aware recommendation task. DSRS consists of two main components, i.e., the social influence learning (SIL) and dynamic
preference learning (DPL). Specifically, in the SIL module, we arrange social graphs in a sequential order and borrow the
power of graph convolution networks (GCNs) to learn social context. Moreover, we design a structural-temporal attention
mechanism to discriminatively model the structural social influence and the temporal social influence. Then, in the DPL
part, users’ individual preferences are learned dynamically by recurrent neural networks (RNNs). Finally, with a prediction
layer, we combine the users’ social context and dynamic preferences to generate recommendations. We conduct extensive
experiments on two real-world datasets, and the experimental results demonstrate the superiority and effectiveness of our
proposed model compared with the state-of-the-art methods.
Keywords recommender system, social influence, sequential recommendation
1 Introduction
In the digital and informational era, people are over-
whelmed by the glut of information. To deal with
this problem, the recommender system has become an
important application. Usually, recommender system
techniques focus on modeling user behaviors, item at-
tributes, and contexts to capture user preference, thus
accomplishing personalized recommendation services.
Along this line, large efforts have been made in both
industry and academia to enhance the performance of
recommender systems.
For a long time, collaborative filtering (CF) is
one of the most popular techniques for building rec-
ommender systems. Among the CF-based algorit-
hms, latent factor models have achieved significant suc-
cess in many recommendation tasks [1–3]. In recent
years, with the rapid development of online social plat-
forms, many researchers have exploited the use of so-
cial contextual information to alleviate the sparsity is-
sue and improve recommendation performance. Ini-
tially, social information is integrated into latent fac-
tor models for social-aware recommendation[4–9]. How-
ever, interactions between users and items are com-
plex, which could be hardly captured by the lin-
ear operation in most latent factor methods. For-
tunately, the deep learning technique has shown its
capability in solving these problems and enhanc-
Regular Paper
Special Section on Learning and Mining in Dynamic Environments
This work was partially supported by the National Key Research and Development Program of China under Grant
No. 2018YFB1402600, the National Natural Science Foundation of China under Grant Nos. 61703386 and U1605251, and the MSRA
(Microsoft Research Asia) Collaborative Research Project.
∗Corresponding Author
©Institute of Computing Technology, Chinese Academy of Sciences 2020

282 J. Comput. Sci. & Technol., Mar. 2020, Vol.35, No.2
ing the performance of recommendation task, such
as CTR prediction [10,11], top-krecommendation [12,13]
and session- based recommendation [14–16]. Existing
studies have leveraged rich contextual information in
deep learning based models, such as visual [17,18]and
textual [19,20]contents of items. But the dynamic so-
cial influence, as an important information that reflects
the evolving of the user profiles, has been largely unex-
ploited.
Actually, there are still some unique challenges in-
herent in introducing dynamic social influence into the
process of user preference modeling. Firstly, different
from common auxiliary information such as pictures
and texts, which are associated with users or items,
social context refers to the correlation among users.
Thereby most existing context-aware recommendation
methods cannot be directly applied to model social
contextual information. Secondly, interactions between
users and items change over time, so do the social re-
lations. It could be a difficult task to bridge dynamic
information from these two domains. Thirdly, the so-
cial influence is complex and evolving over time. For
example, Fig.1shows a case that a user’s preference
is affected by his/her social relations. As illustrated
in Fig.1, among all current social relations, different
friends have impacts on the center user in different lev-
els. Meanwhile, when the user builds a new social rela-
tion, he/she may show more interest in the items that
his/her new friend likes. Therefore, the dynamic social
context could help us to better understand the evolving
user preference. However, how to capture dynamic in-
fluence from evolving social relations effectively is still
an under-explored question.
(b)(a)
Fig.1. Illustration of a case that users’ preferences are dynami-
cally affected by their social relations.
Based on these intuitions, in this paper, we conduct
a comprehensive study to exploit the dynamic social
influence to enhance the performance of the sequen-
tial recommendations. We propose a hybrid model
called Dynamic Social-Aware Recommender System
(DSRS) to deeply explore the dynamic social influ-
ence and integrate social context to address the se-
quential recommendation task. DSRS is composed of
two components, i.e., the social influence learning (SIL)
and dynamic preference learning (DPL). Specifically,
in the SIL module, we arrange social graphs in a se-
quential order and develop graph convolution networks
(GCNs) to learn social context. Moreover, we design
a structural-temporal attention mechanism to discrim-
inatively model the social influence on structural and
temporal aspects. Then, in the DPL part, we use recur-
rent neural networks (RNNs) to capture users’ dynamic
preferences. Finally, with a prediction layer, we inte-
grate users’ social context and dynamic preferences to
generate recommendations. We conduct experiments
on two real-world datasets, and the experimental re-
sults have clearly demonstrated the superiority of our
proposed model. The contributions of this paper could
be summarized as follows.
•We introduce the problem of dynamic social-aware
recommendation, which focuses on revealing the im-
pacts of dynamic social influence on user profiles.
•We propose a novel model DSRS to jointly capture
dynamic social context and user preference. Moreover,
we design a structural-temporal attention mechanism
to discriminatively model social effects for the target
users on structural and temporal aspects.
•We conduct comprehensive experiments on two
real-world datasets, and the results demonstrate the ef-
fectiveness and superiority of our proposed model com-
pared with several state-of-the-art methods.
2 Related Work
In this section, we will briefly review the studies that
are related to our work. As we focus on exploring dy-
namic property and leveraging social contextual infor-
mation in the recommendation task, the related studies
mainly fall into two parts, i.e., social recommendation
and sequential recommendation.
2.1 Social Recommendation
In social media platforms, users’ behaviors can be
divided into two types, i.e., consumption behaviors and
social behaviors. With the knowledge of social influ-
ence theory that users’ consumption behaviors are af-
fected by their social r elations [21], a large amount of
efforts have been made to exploit social information
for recommendation tasks. Traditionally, latent fac-
tor models have been widely used to model social con-
textual information [4–8]. For example, Jamali and Es-

Yang Liu et al.: Exploiting Structural and Temporal Influence for DSRS 283
ter proposed SocialMF, a social influence propagation
mechanism that enables user latent representation to
depend on p ossibly the entire so cial network [6]. Ma
et al. designed SocialReg which adopts regularization
method to force similar users to have similar latent
representations [5]. Guo et al. proposed a trust-based
matrix factorization model, which incorporates both
rating and trust information [7]. With the advance of
online social networks, latent factor based methods
show its insufficiency in handling large-scale complex
data. Lately, researchers have adopted neural networks,
which have particular advantage in processing com-
plex data with their deep non-linear operation, to im-
plement so cial-aware rec ommendation [22–28]. Among
those methods, Sun et al. attempted to incorporate
social information into RNNs architecture [22], and Wu
et al. devised an autoencoder-based approach for so-
cial embedding learning [23]. Considering that social re-
lations are more suitable to be expressed as a graph,
some studies take advantage of GCNs to extract social
contextual information [24–26]. For example, Wu et al.
proposed SocialGCN to model the diffusion process of
social influence [24]. Qiu et al. proposed DeepInf [26],
a graph-based learning framework for predicting social
influence. Most previous studies process social rela-
tions with a static treatment, while we differ from these
studies by considering sequential social behaviors and
dynamic social influence.
2.2 Sequential Recommendation
In realistic scenarios, various user behaviors are
recorded with timestamps, which results in the se-
quence form that may reflect dynamic user preference.
The importance of sequential information has gradu-
ally attracted researchers’ attention. Initially, Markov-
based methods are widely used in modeling temporal
data, such as FPMC [29]and HRM [30]. These methods
expose their deficiency in capturing complex long-term
sequence dependency. To overcome the limitation of
Markov-based methods and provide more effective so-
lution to time series data mining, researchers have ap-
plied RNNs to sequential recommendation in various
application scenarios [13,14,31–33]. For example, Hidasi
et al. attempted to make gated recurrent unit (GRU) fit
session-based recommendation by introducing session-
parallel mini-batches [14]. Yu et al. proposed a recurrent
neural model (DREAM) for next basket prediction[13].
Wu et al. provided RRN, a method based on recur-
rent neural networks that can model the user and the
item dynamics synchronously [31]. Some scholars also
attempted to alleviate the data sparsity problem and
improve prediction accuracy by integrating rich contex-
tual information, such as visual and textual content of
items and external situatio ns[34–36]. Under the back-
ground of dynamic modeling, social contextual infor-
mation, as an important driven force of the evolving
of user preference, is rarely used compared with other
context. To fill this gap, in our work we dynamically
model not only user preference but also social context.
3 Methodology
In this section, we will firstly give a problem def-
inition of the dynamic social-aware recommendation
task. Then, we will describe the technical details of the
proposed Dynamic Social-Aware Recommender System
(DSRS) model, which consists of two main components:
an attentive GCNs part named Social Influence Learn-
ing (SIL) to acquire dynamic influence in social do-
main, and a Dynamic Preference Learning (DPL) mod-
ule to capture dynamic preference of the target user in
the consumption domain. Finally, we will analyze the
model complexity of DSRS. The overview of our model
architecture is presented in Fig.2, where Fig.2(a) is the
SIL module and Fig.2(b) is the DPL module. The finial
predictions come from the two modules.
3.1 Preliminaries
3.1.1 Problem Definition
In our recommendation scenario, there are a set
of users U={u1, u2,...,u|U|}and a set of items
V={v1, v2,...,v|V|}. Without confusion, we use in-
dexes u, u′to denote users and v , v′to denote items.
Users can interact with items and build relationship
with each other. As we focus on sequential behavior
and implicit feedback, the transactions are recorded
with timestamps and the rating values are transformed
to binary. For user-item interactions, we use Bu
tto
represent the set of items that uconsumed at time t.
For user-user relationships, let Nu
tbe the set of neigh-
bors that is established at time t. Our problem can be
formulated as follows.
Definition 1 (Dynamic Social-Aware Recommen-
dation). Given a user u, the corresponding temporal
consumption set Bu={Bu
1, Bu
2,...,Bu
T}, and the
temporal social records Nu={Nu
1, N u
2,...,Nu
T}, our
task is to predict the user’s future choices (Bu
T+1)by
leveraging both consumption and social information.

284 J. Comput. Sci. & Technol., Mar. 2020, Vol.35, No.2
Recurrent
Neural Networks
Item
Embedding
Average
Temporal
Attention
Structural
Attention
Average
Temporal
Attention
Structural
Attention
Average
Temporal
Attention
Structural
Attention
Average
Average Average
Attentive
Convolution
Networks
Dynamic
Social Graph
Input
Pooling
SIL
DPL
...
...
...
...

h
t֓
u
h
t֓
u
B
t֓
u
B
t֓
u
B
t
u
x
t֓
u
x
t֓
u
x
t
u
q
v
h
t֓
u
h
t
u

h
t֓
u

h
t
u

r
t
uv

h
t
u
(a)
(b)
Fig.2. Overview of the proposed DSRS architecture. (a) Social Information Learning (SIL) module. (b) Dynamic Preference Learning
(DPL).
3.1.2 Dynamic Social Context
In a social media platform, users can build relation-
ships with each other. Social relations can be expressed
as a directed graph, in which users act as nodes and so-
cial links act as edges. For a specific node, the nodes
that it points to are regarded as its neighbors. As users
can remove and add their social relations, the edges in
the social graph are not static, but change over time.
Therefore, for each specific user, the corresponding so-
cial context information is dynamic.
In our scenario, we consider the dynamic character-
istics of social context information from two aspects:
structural context and temporal context. As shown
in Fig.3(a), links among users are changing over time,
which evolves into a sequence of graphs (structural con-
text). And Fig.3(b) illustrates the difference among all
current neighbor nodes, which refers to that the dura-
tions of current social relations are different (temporal
context). The evolving structure of social graph re-
flects that the contextual information in the social do-
main is dynamic. And the reason why we pay attention
to temporal context is that time factor can reflect the
peculiarity of social relations to a certain extent. In-
tuitively, newly-added social links affect users’ current
tastes more than the pre-existing ones. We take the two
kinds of dynamic context into consideration and model
them respectively in our Social Influence Learning (SIL)
module. Next, we will discuss the design details of the
SIL module.
........................(a)
(b)
t֓ t֓ t
t֓ t֓ t
Fig. 3. Illustration of dynamic social context. (a) Structural
context. Social structure changes over time, which formulates a
sequence of social graphs. (b) Temporal context. Current social
links come from different time steps. Links within a same small
circle are established at the same time step.

Yang Liu et al.: Exploiting Structural and Temporal Influence for DSRS 285
3.2 Social Influence Learning
For learning social context, it is naturally to treat
users’ social networks as a graph. Indeed, there has
been a surge of research into GCNs [37], which adapts
neural networks to graph-structured data. The core
idea of GCNs is to learn node embeddings by consi-
dering both node contents and the topology structure
in a graph. To differentiate the importance degrees of
neighbor nodes, some researchers have combined atten-
tion mechanism with GCNs [38–41]. Most graph-based
methods process a large and static social graph, which
leads to computational cost issues. Besides, the dy-
namic property of social relations is neglected among
these methods. Instead of static treatment, we im-
plement GCNs on temporal social graphs. In each
step, the convolution operation is performed on a graph
formed in one time interval (Gt), rather than on an
entire static graph. Moreover, we design an attention
approach in our convolution step by considering both
graph structure context and temporal context.
According to the social influence theory [21], a user’s
preference is affected by his/her social relations. Consi-
dering that the effects can be in different degrees,
we adopt attention mechanism to evaluate neighbor
weights in the SIL module. We design a structural-
temporal attention approach, which is constituted of
structural attention and temporal attention. Given a
target user uand a social graph Gt, the structural at-
tention calculates neighbor weights based on the sub-
graph structure with uas the center node, and here how
long each social relation exists is indiscriminate. Mean-
while, the temporal attention differentiates neighbor
influence by considering when the relations are estab-
lished (i.e., time factor). Social relations from different
time contribute differently to the change of user pref-
erence. Therefore, the temporal attention is devised to
capture time effect. The architecture of our structural-
temporal attention is illustrated in Fig.4. Next, we will
discuss our attention strategy in detail.
3.2.1 Structural Attention
In structural attention networks, the importance de-
grees of the neighbors to a user are evaluated based on
their node features. The inputs to structural atten-
tion layer are a graph Gtand the node features Xt=
{x1
t, x2
t,...,x|U|
t},xi
t∈RF, where |U|is the number of
nodes. The outputs are the attention weights. First, all
feature vectors are transformed into intermediate rep-
resentations by a shared weight matrix Wx∈RF×F′:
ˆ
Xt=XtWx,
where ˆ
Xt={ˆx1
t,ˆx1
t,...,ˆx|U|
t}is the linear transforma-
tion of Xt. For a target node uand a neighbor node
u′, the relation between node uand u′is calculated as
follows:
uu′
t=σ(a(ˆ
xu
t⊕ˆ
xu′
t)),
where ais a shared parameter vector and ⊕is con-
catenation operation. The attention is masked, which
means the calculation is only performed among uand
its neighbors at present moment (i.e., Nu
t). After the
relations are obtained, the attention coefficients are the
normalization of {uu′
t,u′∈Nu
t}:
αuu′
t= softmax(uu′
t) = exp(uu′
t)
Σu′∈Nu
texp(uu′
t),
where αuu′
tindicates the importance degree of node u′
to node u.
(xt, xt )
u'
uf(xt, xt ; Wx, a)
u'
u
yt
u
ht
u
zt
u
(xt, xt )f(xt, xt, et ; Wx, Wt, b)
u'u'u'
uu
Fig.4. Illustration of structural-temporal attention architecture.
Vectors in browns are node features and those in grays denote
time factors.
3.2.2 Temporal Attention
In temporal attention networks, the calculation de-
pends on not only node features but also associated
time factors. We use etto denote the embedding vector
of time factor. For each u′(u′∈Nu
t), the correspond-
ing time factor is obtained by embedding look-up. The
calculation of attention weights here is similar to that
of structural attention:
ˆ
et=etWt,
ˆ
uu′t=σ(b(ˆ
xu
t⊕ˆ
xu′
t⊕ˆ
eu′
t)),
βuu′
t= softmax( ˆ
uu′t) = exp( ˆ
uu′t)
Σu′∈Nu
texp( ˆ
uu′t),

286 J. Comput. Sci. & Technol., Mar. 2020, Vol.35, No.2
where Wt∈RF×F′is a weight matrix for linear trans-
formation of time vectors, which is used to map time
vector to a proper vector space. And bis a weight vector
for computing intermediate attention value. The out-
put βuu′
tis the normalized attention coefficient, which
indicates the influence of u′to uwith time effect. Once
obtained, the attention outputs αuu′
tand βuu′
tare used
to aggregate information from two aspects:
yu
t= Σu′∈Nu
tαuu′
tˆ
xu′
t,
zu
t= Σu′∈Nu
tβuu′
tˆ
xu′
t,
˙
hu
t=σ(average(yu
t,zu
t)),
where the meaning of yu
tand zu
tis the social context
vectors from two aspects (i.e., structural context and
temporal context). After the aggregation operation,
˙
hu
tserves as the output of the SIL module, which con-
tains social contextual information with temporal in-
fluence. The above algorithm will be implemented to
calculate dynamic social context vectors for all nodes
in the graph.
3.3 Dynamic Preference Learning
The SIL module is applied to extract dynamic in-
formation in the social domain. To obtain users’ pref-
erences in the consumption domain, we use a sequential
model to explore users’ behavioral patterns. As RNNs
are extensively used to process temporal sequence due
to their ingeniously designed recurrent feedback mecha-
nism, we utilize LSTMs [42], a variant of RNNs, to learn
users’ individual preferences. The architecture of our
sequential behavior modeling part (i.e., the DPL mod-
ule) is showed in Fig.2(b).
3.3.1 Input Pooling
In each time step, the LSTMs take current consump-
tion records as inputs. The corresponding outputs serve
as two purposes, i.e., users’ current representations and
the hidden states of the next step. Note that a user can
interact with multiple items in a time window, and the
LSTM takes a fixed-size vector as input at each time
step. To generate valid model inputs, a pooling ope-
ration will be performed. Before input pooling, we first
use an embedding layer to project the sparse represen-
tations of items to dense vectors. The item embedding
set is denoted as Q={qv1, qv2,...,qv|V|}. For a target
user uand a specific time step t, from the consump-
tion records set Bu
twe can obtain the corresponding
embeddings Qu
t={qvi|vi∈Bu
t}. Then a pooling ope-
ration will be applied to aggregate input vectors. We
choose the commonly used function average pooling as
the aggregation function. Among all the vectors to be
aggregated, average pooling takes the average value of
every dimension. The calculation of input vector is for-
mulated as follows:
xu
t=average(Qu
t) = Σvi∈Bu
tqvi
|Qu
t|.
3.3.2 Dynamic Behavior Modeling
Based on the current input xu
tand the previous hid-
den state hu
t−1, the LSTM unit calculates the updated
hidden state hu
twhich represents the current prefer-
ence of u. The compact forms of the equations for the
forward pass of an LSTM unit are as follows:
ft=σ(Wf×[hu
t−1,xu
t]),
it=σ(Wi×[hu
t−1,xu
t]),
ot=σ(Wo×[hu
t−1,xu
t]),
ct=ft◦ct−1+it◦σ(Wc×[hu
t−1,xu
t]),
hu
t=ot◦σ(ct),
where ft,it,otrepresent the forget gate, the input gate
and the output gate respectively. The initial values c0
and h0are generally set to zero. All W∗are learn-able
parameters and σis the active function. The output
hu
tserves as the preference vector of uin consumption
domain.
3.4 Prediction Layer
The hybrid model DSRS combines the two parts
(i.e., SIL and DPL) to obtain final representation of
user preference. As shown in Fig.2, the SIL module
extracts social contextual information from a temporal
social graph dynamically, and the DPL module adapts
LSTMs to capture the evolving user preference in the
consumption domain. In each time step, the social con-
text vector and the user preference vector are combined
as follows:
¨
hu
t=˙
hu
t+hu
t.
Note that the input features to SIL and DPL are
both calculated from item embedding; thus the out-
puts ˙
hu
tand hu
tcan be combined by vector addition
operation. When predicting the preference that a user
gives to an item, the predicted rating value equals the
dot product of ¨
hu
tand qv:
˙ruv
t=dot(¨
hu
t,qv).

Yang Liu et al.: Exploiting Structural and Temporal Influence for DSRS 287
3.5 Model Training
We adopt Bayesian Personalized Ranking (BPR) [3]
framework for model learning. BPR is a widely used
pairwise ranking framework for implicit feedback. The
basic assumption of BPR is that a user prefers a posi-
tive item more than a negative one. Based on the above,
the following probability needs to be maximized:
p(u, t, v ≻v′) = σ( ˙ru,v
t−˙ru,v′
t),(1)
where vand v′denote a positive sample and a nega-
tive sample respectively, and σ(·) is the sigmoid func-
tion σ(x) = 1/(1 + e−x). With this functional form,
a higher score is expected to be given on the positive
item in comparison with the negative one. Since our
goal is to predict user preference in the future, at time
step t, we treat items that show in t+ 1 as the positive
samples.
In the training step, for each positive sample (v), we
randomly choose an item that the user has not inter-
acted with before as the corresponding negative sample
(v′). The objective function is adding up the log likeli-
hood of (1) and the regularization term:
J= Σln(1 + e−( ˙ru,v
t−˙ru,v ′
t)) + λ
2kΘk2,
where λis a parameter to control the power of regulari-
zation and Θ denotes all the parameters to be esti-
mated. In pr actice, we choose Adam [43]as the opti-
mizer, which has proved to be especially effective for
training neural networks. The model parameters up-
dating procedure is repeated iteratively until the con-
vergence is achieved.
3.6 Model Analysis
3.6.1 Space Complexity
All the model parameters come from two parts: the
parameters Θ1= [Wf,Wi,Wo,Wc] in LSTMs and the
parameters Θ2= [Wx,Wt,a,b] in attentive GCNs. For
Θ1, the space complexity grows linearly with the layer
of LSTMs. In most RNNs-based models the layer num-
ber is set to 1 or 2. The parameter sharing mechanism
in RNNs enables its space complexity to be indepen-
dent of the number of users. As for Θ2, it is lighter
than Θ1, and parameters in Θ2are shared by all users.
Therefore, the total space complexity of DSRS is rea-
sonable.
3.6.2 Time Complexity
Compared with the basic RNN model, the addi-
tional time cost of DSRS mainly lies in the attentive
convolution operation. If the user number is N, the
average neighbor size per time is M, and for Ttime
steps the time complexity of single layer convolution
calculation is O(2N M T ). In the training datasets, T
equals 12 in Epinions and 3 in Gowalla. The average
number of neighbors per time step is 7 in Epinions and
6 in Gowalla. Hence the additional time cost is accept-
able.
4 Experiments
We conduct experiments on the proposed model and
other compared methods. To verify the model perfor-
mance, our experiments are designed mainly to answer
the following questions.
RQ1. Compared with the state-of-the-art methods,
how does the proposed model perform?
RQ2. How do the two modules SIL and DPL per-
form when they are separately used?
RQ3. Is the attention mechanism helpful in our rec-
ommendation task? What roles do the two attention
strategies play?
RQ4. How do the hyper parameters (e.g., embed-
ding size) influence model performance?
4.1 Datasets
We conduct experiments on two real-world datasets,
i.e., Epinions 1
○and Gowalla 2
○. Both the two datasets
contain temporal consumption records and social rela-
tions records. Next, we will briefly introduce the two
datasets, and then we will describe the data processing
in our experimental implementation.
•Epinions [44]. It is a who-trust-whom online social
network of a general consumer review site. Members
of this platform can read new and old reviews about a
variety of items to help them decide on a purchase, and
they can also decide whether to “trust” each other. We
use the public Epinions dataset provided by Richardson
et al. [44]In this dataset, users’ rating and social actions
are timestamped.
•Gowalla [45]. It is a location sharing social net-
working website. Users on this platform are able to
check in at “Spots” in their local vicinity, and they
can also build social relations with each other. We use
1
○http://www.epinions.com/, Jan. 2020.
2
○https://blog.gowalla.com/, Jan. 2020.

288 J. Comput. Sci. & Technol., Mar. 2020, Vol.35, No.2
the Gowalla dataset provided by [45]. In this dataset,
checking-in actions and social behaviors are both times-
tamped.
In both the two datasets, we treat one month as a
time window. There are total 13 time windows in Epin-
ions and 4 in Gowalla. Since implicit feedback is the
main point of focus, we transform the concrete rating
values to binary. Consumption data is recorded in the
form of (u, v, t) triple ID, which indicates that user u
interacted with item vat time t. Similarly, the social
relation is also recorded as (u, u′, t) triple ID, which in-
dicates that user ubuilt a social relation with user v′
at time t. In our sequential prediction task, the goal
is to predict users’ future preferences. Therefore, the
records in the last time window will be used for model
testing (i.e., T= 13 in Epinions and T= 4 in Gowalla)
and the others for training (i.e., T= 1–12 in Epinions
and T= 1–3 in Gowalla). we randomly select 10%
from the training data as validation data for parame-
ter tuning. We filter out users who have less than two
time windows. After data pruning, there are 3 282 users
and 26 991 items in Epinions. In the Gowalla dataset,
there are 7 035 users and 71 139 items. Table 1lists the
statistics of the two datasets.
4.2 Baselines
We compare our proposed model with the following
baselines.
•BPR [3]. It is a generic optimization criterion and
learning algorithm for personalized ranking. BPR pre-
dicts ratings by calculating the inner product of the
user and item latent vectors. With the assumption
that users prefer positive items to the negative ones,
it adopts a pair-wise loss function for model learning.
•FPMC [29]. It is a traditional sequential model for
next basket recommendation. By combining MC and
MF, FPMC can capture general interest of users and
sequential effects between every two adjacent time in-
tervals.
•SocialMF[6]. This is a classical model for rec-
ommendation with social influence. SocialMF incor-
porates social information into the basic MF model.
Specifically, when updating a user’s latent vector, it
fuses the latent vectors of corresponding neighbors.
•DREAM [13]. We adapt DREAM, an RNNs-based
method, as one of our baselines. For more clearly
comparison, both our basic sequential prediction mod-
ule and DREAM use two-layer LSTMs as the building
block.
•SocialGCN[24 ]. It is a GCNs-based recommenda-
tion algorithm, which captures social influence by an
information diffusion process. The designers of Social-
GCN leverage rich user and item attributes in their
experiment and propose a general version for the sce-
nario where no attributes are available. We employ the
general version SocialGCN which is applicable to our
recommendation scenario.
As we aim to tackle the problem of bridging tempo-
ral social influence and sequential prediction, we choose
methods by considering two aspects: sequence-aware
methods and social-aware methods. Besides, we also
consider three variants of the proposed method, and
the variants are listed as follows.
•DSRS-avg. As a simplified version of DSRS, this
method also leverages social information and it adopts
an average operation rather than attentive aggregation.
•DSRS-s. This is a variant of the proposed method
which uses structural attention only in the attentive
convolution step.
•DSRS-t. This is a variant of the proposed method
where only temporal attention mechanism is used in the
attentive convolution step.
4.3 Metrics and Setups
4.3.1 Metrics
In order to measure the performance, we adopt
two frequently used evaluation metrics hit ratio (HR)
and normalized discounted cumulative gain (NDCG)
for top-krecommendation task, as applied in previous
literature [12,22,24,34].HR@kmeasures the percentage
of the positive samples presented in the top-kranking
list. For a single user, the calculation of HR@kis as
follows:
HR@k=Numberof Hit@k
|GT |,
where |GT |is the length of positive samples in test set.
And NDCG@kis sensitive to the positions of the hit
Table 1. Statistics of the Two Datasets
Dataset Users Items Time Windows Social Links Train Records Test Records Link Density (%) Rating Density (%)
Epinions 3 282 26 991 13 106 076 174 325 8 056 0.980 0.190
Gowalla 7 035 71 139 4 47 864 180 944 18 736 0.096 0.036

Yang Liu et al.: Exploiting Structural and Temporal Influence for DSRS 289
positive samples in the ranking list. The calculation of
NDCG@kis as follows:
DCG@k= Σk
i=1
2reli−1
log2(i+ 1) ,
NDCG@k=DC G@k
IDCG@k,
where iin DCG is the position index and reliis the rele-
vance score, IDCG is the ideal value of DCG, i.e., the
DCG value where items in the predicted top-klist are
all positive samples. When correctly predicted samples
rank highly on the list, NDCG reaches a higher value.
For both HR@kand N DCG@k, the larger the value,
the better the performance.
4.3.2 Setups
In the training step, all methods are optimized with
mini-batch and the batch size is set to 512. The learning
rate and optimizer are searched for each model accord-
ing to their peculiarity. To prevent the neural networks-
based models from overfitting, we employ dropout and
set the ratio to 0.5. For models that are based on latent
factor, the latent vectors are randomly initialized with
Gaussian distribution with mean 0 and standard devia-
tion 0.01. We tune all the parameters to ensure the best
performance of the baselines for fair comparison. In the
testing step, we generate top-kranking list and eva-
luate all benchmarks with H R@kand N DCG@k. For
negative sampling, we randomly select 500 items and
each user has not interacted with as the negative sam-
ples. The negative samples are mixed with the positive
ones for ranking. We compare all methods with diffe-
rent latent factor dimensions d={32,64,128,256}and
set k= 10 for top-k. Experiments are conducted on a
Linux server with four 2.0 GHz IntelrXeonrE5-2620
CPUs and a Tesla K80 GPU.
4.4 Results and Analysis
4.4.1 Overall Comparison (RQ1)
We first compare the performance of the proposed
DSRS with the baselines. All the methods are evaluated
by same metrics (i.e., HR@10 and N DCG@10). Ta-
ble 2and Table 3show the results on datasets Epinions
and Gowalla respectively. We bold the results of DSRS
for better comparing our method with other baselines.
1) From the results we can see that sequential mod-
els give more favorable performance than the static
ones. Among sequence-aware methods, DSRS and
DREAM achieve better performance than FPMC, as
the former can capture multi-step dependency among
users’ preferences and the latter is in a pair-wise way.
Both being neural networks based sequential models,
DSRS performs better than DREAM. As a static latent
factor model, BPR lags behind the sequential methods.
2) By comparing the results of SocialMF and BPR,
the observation highlights the importance of social in-
formation. These two methods are both static la-
Table 2. Overall Performance Comparison on Dataset Epinions
Method Dimension
d= 32 d= 64 d= 128 d= 256
HR@10 N D CG@10 H R@10 N DC G@10 H R@10 N DC G@10 HR@10 N DC G@10
BPR [3]0.158 0.045 1 0.163 0.046 3 0.165 0.046 6 0.159 0.045 4
FPMC [29]0.167 0.046 3 0.175 0.049 0 0.189 0.051 7 0.185 0.051 5
SocialMF [6]0.166 0.045 8 0.168 0.046 4 0.167 0.045 1 0.163 0.049 3
DREAM [13]0.182 0.044 5 0.210 0.051 5 0.221 0.057 0 0.213 0.056 5
SocialGCN [24]0.186 0.046 0 0.184 0.048 8 0.173 0.046 2 0.169 0.046 0
Proposed DSRS 0.201 0.048 3 0.220 0.053 6 0.233 0.060 8 0.230 0.061 1
Table 3. Overall Performance Comparison on Dataset Gowalla
Method Dimension
d= 32 d= 64 d= 128 d= 256
HR@10 N D CG@10 H R@10 N DC G@10 H R@10 N DC G@10 HR@10 N DC G@10
BPR [3]0.480 0.154 0.495 0.160 0.498 0.161 0.488 0.157
FPMC [29]0.511 0.151 0.527 0.167 0.532 0.172 0.536 0.173
SocialMF [6]0.481 0.157 0.516 0.173 0.552 0.182 0.578 0.189
DREAM [13]0.527 0.151 0.557 0.168 0.563 0.175 0.558 0.165
SocialGCN [24]0.484 0.158 0.537 0.177 0.515 0.161 0.493 0.161
Proposed DSRS 0.569 0.170 0.623 0.192 0 .635 0.198 0.663 0.206

290 J. Comput. Sci. & Technol., Mar. 2020, Vol.35, No.2
tent factor based, and the difference is that SocialMF
leverages social information for recommendation. As
a social-aware deep model, SocialGCN does not give
significant upgrades as expected. The ordinary per-
formance of SocialGCN lies in two aspects. First, the
temporal information is neglected in SocialGCN. Sec-
ond, it does not differentiate the important degrees of
all neighbors, and the diffusion process in SocialGCN
brings more information meanwhile it causes more dis-
turbance. DSRS, by contrast, processes sequential so-
cial graphs and uses attention mechanism to differenti-
ate neighbor weights, and it outperforms SocialGCN.
3) Generally, the proposed model outperforms all
other methods in most cases. Compared with SocialMF
and SocialGCN, DSRS extracts social information dy-
namically by constructing a graph sequence. The bene-
fit of such an approach is that we do not have to process
a large social graph at one time. Moreover, the designed
structural-temporal attention mechanism can eliminate
redundancy and maintain quality social information ef-
fectively. Compared with all the baselines, on the Epin-
ions dataset, DSRS achieves 8.06%–44.65% increase in
HR@10 and 4.07%–34.81% in N DCG@10, and on the
Gowalla dataset, the improvements are 7.97%–35.86%
in HR@10 and 7.59%–31.21% in N DCG@10.
4.4.2 Module Performance (RQ2)
To test the performance of the two modules (i.e., SIL
and DPL) in DSRS, we conduct comparative experi-
ments and show the results in Table 4and Table 5,
where DSRS-SIL utilizes only social context informa-
tion to make predictions and DSRS-DPL only learns
user personal preference. By comparing model per-
formance on the two datasets, we can see that social
information plays different roles in different scenarios.
On the Epinions dataset, DPL performs better than
SIL. On the Gowalla dataset, SIL is superior to DPL.
The above results indicate that in dataset Gowalla so-
cial context gives more information than consumption
records, and it is the reverse in dataset Epinions. By
combining SIL and DPL, DSRS reaches the best per-
formance in all cases.
4.4.3 Attention Effect (RQ3)
To investigate what role the structural-temporal at-
tention mechanism plays, we conduct experiments on
three variants of DSRS. In this subsection, we will first
compare performance of all variant methods and then
we will analyze the working mechanism of the two at-
tention strategies with examples.
1) To investigate the impact of the designed atten-
tion networks, we test the performance of three variants
of DSRS on the two datasets. In this part of experi-
ments, we fix the vector size and set it d= 32. Table 6
shows the results and the improvements of all the vari-
ant methods. The values in “Improv.” column indicate
the lifting percentages with DSRS-avg as benchmark.
Among all the variants, DSRS-s and DSRS-t are both
single layer attention networks and they are designed
for learning structural weights and temporal weights
respectively. As shown in Table 6, the attentive meth-
ods achieve better performance than DSRS-avg, which
aggregates social information by a simple average sum
operation. Besides, we can observe that the tempo-
ral attention (DSRS-t) attains a slight increase over
the structural attention (DSRS-s) in both H R@10 and
NDCG@10. It indicates that time factor could con-
tribute to the extraction of dynamic social information.
Table 4. Module Performance Comparison on Dataset Epinions
Method Dimension
d= 32 d= 64 d= 128 d= 256
HR@10 N D CG@10 H R@10 N DC G@10 H R@10 N DC G@10 HR@10 N DC G@10
DSRS-SIL 0.181 0.044 4 0.186 0.047 1 0.209 0.054 5 0.194 0.051 8
DSRS-DPL 0.189 0.044 6 0.215 0.051 0 0.225 0.059 0 0.220 0.057 3
DSRS 0.201 0.048 3 0.220 0.053 6 0.233 0.060 8 0.230 0.061 1
Table 5. Module Performance Comparison on Dataset Gowalla
Method Dimension
d= 32 d= 64 d= 128 d= 256
HR@10 N D CG@10 H R@10 N DC G@10 H R@10 N DC G@10 HR@10 N DC G@10
DSRS-SIL 0.560 0.167 0.600 0.181 0.616 0.188 0.623 0.190
DSRS-DPL 0.553 0.161 0.583 0.174 0.603 0.182 0.625 0.195
DSRS 0.569 0.170 0.623 0.192 0.63 5 0.198 0.663 0.206

Yang Liu et al.: Exploiting Structural and Temporal Influence for DSRS 291
Table 6. Effect of Structural and Temporal Attention Mechanism
Method Dataset
Epinions Gowalla
HR@10 Improv. (%) NDC G@10 Improv. (%) HR@10 Improv. (%) NDC G@10 Improv. (%)
DSRS-avg 0.188 – 0.044 7 – 0.554 – 0.163 –
DSRS-s 0.191 +1.60 0.045 4 +1.57 0.560 +1.08 0.167 2.45
DSRS-t 0.195 +3.72 0.045 7 +2.24 0.565 +1.99 0.168 +3.07
DSRS 0.201 +6.91 0.048 3 +8.05 0.569 +2.71 0.170 +4.29
As a hybrid approach, DSRS reaches the best results.
2) With the learned attention weights, we can track
back what and when the social information dominates
the prediction, which helps us to understand the rec-
ommendation strategy. The structural attention esti-
mates which neighbors among current social relations
are more important. The temporal attention measures
the effects of time when relations are established. Final
neighbor weights come from comprehensive results of
the two attention strategies. For better demonstrating
the working mechanism of our attention networks, we
show the tendency of structural and temporal atten-
tion weights during three time steps in Fig.5. We ran-
domly choose five center users (the real IDs of user 1–
user 5 are 4 003, 8 003, 12 003, 16 003, 21 003 respec-
tively) from dataset Gowalla, and draw the learned at-
tention weights of their certain neighbors in three time
steps. As shown in Fig.5, when getting closer to current
time step (t= 3 in this example), the overall values of
temporal attention weights are on a decreasing trend.
The explanation behind this is: as time progresses, the
influence of old neighbors will decrease and their posi-
tion could be replaced by newly-added neighbors.
4.4.4 Hyper Parameter Investigation (RQ4)
The common hyper parameter of all methods is the
embedding size d. As shown in Table 2and Table 3,
different methods reach their best results at different d
due to their unique property. A larger vector size does
not always bring a significant improvement, and the
time consumption will become more expensive when
the vector size increases. Hence the determination of
a proper dcomes from the trade off. In DSRS, there
is another hyper parameter, i.e., the embedding size
dtfor time factor. In the overall comparison stage,
we set the default value of dtequal to d. To explore
the impact of time factor size, we test model perfor-
mance by setting dto a fixed value and changing the
value of dt. Fig.6illustrates the results on Epinions and
Gowalla, where in each sub-figure, the curve in lighter
ttt
Time Steps
User 5 User 4 User 3 User 2 User 1
ttt
Time Steps
1.0
0.8
0.6
0.4
0.2
0.0
1.0
0.8
0.6
0.4
0.2
0.0
User 5 User 4 User 3 User 2 User 1
Structural Attention Weights
0.13 0.15 0.24
0.29 0.14 0.16
0.34 0.05 0.01
0.64 0.23 0.60
0.02 0.18 0.12
0.16 0.09 0.01
0.27 0.05 0.02
0.07 0.02 0.00
0.59 0.32 0.12
0.04 0.02 0.00
Temporal Attention Weights
Fig.5. Visualization of the learned neighbor weights by structural and temporal attention during three time steps.

292 J. Comput. Sci. & Technol., Mar. 2020, Vol.35, No.2
0.22
0.21
0.20
0.19
0.054
0.052
0.050
0.048
0.046
0.62
0.60
0.58
0.56
0.54
0.19
0.18
0.17
0.16
16 32 64 128
dt
16 32 64 128
dt
16 32 64 128
dt
16 32 64 128
dt
HR∂
NDCG∂
NDCG∂
HR∂
d/32
d/64
d/32
d/64
d/32
d/64
d/32
d/64
(b)(a)
(c) (d)
Fig.6. Changing tendency of DSRS performance when changing the dimension of time factor. (a) Epinions-HR. (b) Epinions-NDCG.
(c) Gowalla-HR. (d) Gowalla-NDCG.
color indicates the results with d= 32, dt={16,32,64}
and the curve in darker color shows the results with
d= 64, dt={32,64,128}. Generally, the variations on
the two metrics are not obvious, but we can get the
observation that keeping dtno larger than dis a more
appropriate choice. Time factor is a kind of auxiliary
information, for which a larger size is superfluous.
5 Conclusions
In this paper, we aimed at exploring user social
context to enhance the performance of recommender
systems. To that end, we conducted a comprehen-
sive study to reveal the dynamic social influence on
users’ preferences, and then we proposed a deep model
called Dynamic Social-Aware Recommender System
(DSRS) to address the dynamic social-ware recommen-
dation task. DSRS contains two main components, i.e.,
Social Influence Learning (SIL) and Dynamic Prefer-
ence Learning (DPL). Specifically, we firstly arranged
user social status in a sequential order and developed
graph convolution networks to learn social context
of the target users in SIL. Moreover, we designed a
structural-temporal attention mechanism to discrimi-
natively model the social influence on structural and
temporal aspects. Then, we modeled the users’ indi-
vidual dynamic preferences by DPL. Finally, with a
prediction layer, we integrated the users’ social context
and dynamic preferences to generate personalized rec-
ommendations. We conducted quantitative and quali-
tative experiments on two real-world datasets and com-
pared the proposed DSRS with several state-of-the-art
methods. Experimental results clearly demonstrated
the rationality and effectiveness of DSRS.
References
[1] Mnih A, Salakhutdinov R R. Probabilistic matrix factoriza-
tion. In Proc. the 21st Annual Conference on Neural Infor-
mation Processing Systems, December 2007, pp.1257-1264.
[2] Koren Y. Factorization meets the neighborhood: A multi-
faceted collaborative filtering model. In Proc. the 14th ACM
SIGKDD International Conference on Knowledge Discov-
ery and Data Mining, August 2008, pp.426-434.
[3] Rendle S, Freudenthaler C, Gantner Z et al. BPR: Bayesian
personalized ranking from implicit feedback. In Proc. the
25th Conference on Uncertainty in Artificial Intelligence,
June 2009, pp.452-461.
[4] Ma H, Yang H, Lyu M R et al. SoRec: Social recom-
mendation using probabilistic matrix factorization. In Proc.

Yang Liu et al.: Exploiting Structural and Temporal Influence for DSRS 293
the 17th ACM Conference on Information and Knowledge
Management, October 2008, pp.931-940.
[5] Ma H, Zhou D, Liu C et al. Recommender systems with
social regularization. In Proc. the 4th ACM International
Conference on Web Search and Data Mining, February
2011, pp.287-296.
[6] Jamali M, Ester M. A matrix factorization technique with
trust propagation for recommendation in social networks. In
Proc. the 4th ACM Conference on Recommender Systems,
September 2010, pp.135-142.
[7] Guo G, Zhang J, Yorke-Smith N. TrustSVD: Collaborative
filtering with both the explicit and implicit influence of user
trust and of item ratings. In Proc. the 29th AAAI Confe-
rence on Artificial Intelligence, January 2015, pp.123-129.
[8] Jiang M, Cui P, Wang F et al. Scalable recommendation
with social contextual information. IEEE Transactions on
Knowledge and Data Engineering, 2014, 26(11): 2789-2802.
[9] Biao C, Tong X, Qi L et al. Study on information diffusion
analysis in social networks and its applications. Interna-
tional Journal of Automation and Computing, 2018, 15(4):
377-401.
[10] Huang Z, Pan Z, Liu Q et al. An Ad CTR prediction method
based on feature learning of deep and shallow layers. In
Proc. the 2017 ACM on Conference on Information and
Knowledge Management, November 2017, pp.2119-2122.
[11] Guo H, Tang R, Ye Y et al. DeepFM: A factorization-
machine based neural network for CTR prediction.
arXiv:1703.04247, 2017. https://arxiv.org/abs/1703.04247,
Dec. 2019.
[12] He X, Liao L, Zhang H et al. Neural collaborative filtering.
In Proc. the 26th International Conference on World Wide
Web, April 2017, pp.173-182.
[13] Yu F, Liu Q, Wu S et al. A dynamic recurrent model for
next basket recommendation. In Proc. the 39th Interna-
tional ACM SIGIR Conference on Research and Develop-
ment in Information Retrieval, July 2016, pp.729-732.
[14] Hidasi B, Karatzoglou A, Baltrunas L et al. Session-
based recommendations with recurrent neural networks.
arXiv:1511.06939, 2015. https://arxiv.org/abs/1511.06939,
Dec. 2019.
[15] Li J, Ren P, Chen Z et al. Neural attentive session-based
recommendation. In Proc. the 2017 ACM Conference on
Information and Knowledge Management, November 2017,
pp.1419-1428.
[16] Hidasi B, Karatzoglou A. Recurrent neural networks with
top-kgains for session-based recommendations. In Proc. the
27th ACM International Conference on Information and
Knowledge Management, October 2018, pp.843-852.
[17] Lei C, Liu D, Li W et al. Comparative deep learning of hy-
brid representations for image recommendations. In Proc.
the 2016 IEEE Conference on Computer Vision and Pat-
tern Recognition, June 2016, pp.2545-2553.
[18] Hou M, Wu L, Chen E et al. Explainable fashion
recommendation: A semantic attribute region guided
approach. arXiv:1905.12862, 2019. https://arxiv.org/p-
df/1905.12862.pdf, Dec. 2019.
[19] Kim D, Park C, Oh J et al. Convolutional matrix fac-
torization for document context-aware recommendation. In
Proc. the 10th ACM Conference on Recommender Systems,
September 2016, pp.233-240.
[20] Wang Q, Li S, Chen G. Word-driven and context-aware re-
view modeling for recommendation. In Proc. the 27th ACM
International Conference on Information and Knowledge
Management, October 2018, pp.1859-1862.
[21] Aral S, Muchnik L, Sundararajan A. Distinguishing
influence-based contagion from homophily-driven diffusion
in dynamic networks. Proceedings of the National Academy
of Sciences, 2009, 106(51): 21544-21549.
[22] Sun P, Wu L, Wang M. Attentive recurrent social recom-
mendation. In Proc. the 41st International ACM SIGIR
Conference on Research and Development in Information
Retrieval, July 2018, pp.185-194.
[23] Wu L, Sun P, Hong R et al. Collaborative neural social rec-
ommendation. IEEE Transactions on Systems, Man, and
Cybernetics: Systems. doi:10.1109/TSMC.2018.2872842.
[24] Wu L, Sun P, Hong R et al. SocialGCN: An efficient
graph convolutional network based model for social recom-
mendation. arXiv:1811.02815, 2018. https://arxiv.org/a-
bs/1811.02815, Dec. 2019.
[25] Fan W, Ma Y, Li Q et al. Graph neural networks for social
recommendation. arXiv:1902.07243, 2019. https://arxiv.or-
g/pdf/1902.07243.pdf, Dec. 2019.
[26] Qiu J, Tang J, Ma H et al. DeepInf: Social influence
prediction with deep learning. In Proc. the 24th ACM
SIGKDD International Conference on Knowledge Discov-
ery and Data Mining, August 2018, pp.2110-2119.
[27] Wang H, Xu T, Liu Q et al. MCNE: An end-to-end frame-
work for learning multiple conditional network representa-
tions of social network. In Proc. the 25th ACM SIGKDD
Conference on Knowledge Discovery and Data Mining, Au-
gust 2019, pp.1064-1072.
[28] Xu T, Zhu H S, Zhong H et al. Exploiting the dynamic
mutual influence for predicting social event participation.
IEEE Transactions on Knowledge and Data Engineering,
2019, 31(6): 1122-1135.
[29] Rendle S, Freudenthaler C, Schmidt-Thieme L. Factorizing
personalized Markov chains for next-basket recommenda-
tion. In Proc. the 19th International Conference on World
Wide Web, April 2010, pp.811-820.
[30] Wang P, Guo J, Lan Y et al. Learning hierarchical represen-
tation model for next basket recommendation. In Proc. the
38th International ACM SIGIR Conference on Research
and Development in Information Retrieval, August 2015,
pp.403-412.
[31] Wu C Y, Ahmed A, Beutel A et al. Recurrent recommender
networks. In Proc. the 10th ACM International Conference
on Web Search and Data Mining, February 2017, pp.495-
503.
[32] Li Z, Zhao H, Liu Q et al. Learning from history and
present: Next-item recommendation via discriminatively
exploiting user behaviors. In Proc. the 24th ACM SIGKDD
International Conference on Knowledge Discovery and
Data Mining, August 2018, pp.1734-1743.
[33] Sordoni A, Bengio Y, Vahabi H et al. A hierarchical re-
current encoder-decoder for generative context-aware query
suggestion. In Proc. the 24th ACM International on Confe-
rence on Information and Knowledge Management, Octo-
ber 2015, pp.553-562.

294 J. Comput. Sci. & Technol., Mar. 2020, Vol.35, No.2
[34] Manotumruksa J, Macdonald C, Ounis I. A contextual at-
tention recurrent architecture for context-aware venue rec-
ommendation. In Proc. the 41st International ACM SIGIR
Conference on Research and Development in Information
Retrieval, July 2018, pp.555-564.
[35] Smirnova E, Vasile F. Contextual sequence modeling for
recommendation with recurrent neural networks. In Proc.
the 2nd Workshop on Deep Learning for Recommender Sys-
tems, August 2017, pp.2-9.
[36] Manotumruksa J, Macdonald C, Ounis I. A contextual at-
tention recurrent architecture for context-aware venue rec-
ommendation. In Proc. the 41st International ACM SIGIR
Conference on Research and Development in Information
Retrieval, July 2018, pp.555-564.
[37] Bruna J, Zaremba W, Szlam A et al. Spectral networks
and locally connected networks on graphs. arXiv:1312.6203,
2013. https://arxiv.org/abs/1312.6203, Dec. 2013.
[38] Feng C, Liu Z, Lin S et al. Attention-based graph convo-
lutional network for recommendation system. In Proc. the
2019 IEEE International Conference on Acoustics, Speech
and Signal Processing, May 2019, pp.7560-7564.
[39] Zhang J, Shi X, Xie J et al. GaAN: Gated atten-
tion networks for learning on large and spatiotempo-
ral graphs. arXiv:1803.07294, 2018. https://arxiv.org/p-
df/1803.07294.pdf, Dec. 2019.
[40] Veli˘ckovi´c P, Cucurull G, Casanova A et al. Graph atten-
tion networks. arXiv:1710.10903, 2017. https://arxiv.org/a-
bs/1710.10903, Dec. 2019.
[41] Wu L K, Li Z, Zhao H K et al. Estimating early fundraising
performance of innovations via graph-based market envi-
ronment model. arXiv:1912.06767, 2019. https://arxiv.or-
g/abs/1912.06767v1, Dec. 2019.
[42] Hochreiter S, Schmidhuber J. Long short-term memory.
Neural Computation, 1997, 9(8): 1735-1780.
[43] Kingma D P, Ba J. Adam: A method for stochastic
optimization. arXiv:1412.6980, 2014. https://arxiv.org/ab-
s/1412.6980, Dec. 2019.
[44] Richardson M, Agrawal R, Domingos P. Trust management
for the semantic Web. In Proc. the 2nd International Se-
mantic Web Conference, October 2003, pp.351-368.
[45] Scellato S, Noulas A, Mascolo C. Exploiting place fea-
tures in link prediction on location-based social networks.
In Proc. the 17th ACM SIGKDD International Conference
on Knowledge Discovery and Data Mining, August 2011,
pp.1046-1054.
Yang Liu received her B.E. degree
in information security from University
of Science and Technology of China
(USTC), Hefei, in 2016. She is currently
working toward her Master’s degree in
the Anhui Province Key Laboratory
of Big Data Analysis and Application,
USTC, Hefei. Her research interests
include deep learning and its application in recommender
systems.
Zhi Li received his B.E. degree in
software engineering from the Xi’an
Jiaotong University, Xi’an, in 2015.
He is currently pursuing his Ph.D.
degree with the School of Data Science,
University of Science and Technology of
China, Hefei. His current research inte-
rest includes data mining, recommender
systems and industrial intelligence. He has published
papers in in refereed journals and conference proceedings,
such as Journal of Computer Science and Technology,
ACM SIGKDD, AAAI and IJCAI.
Wei Huang received his B.E. degree
in software engineering from Sichuan
University (SCU), Chengdu, in 2017.
He is currently working toward his
Ph.D. degree in the School of Data
Science, University of Science and
Technology of China (USTC), Hefei.
His research interests include data
mining, deep learning, natural language processing and
applications in text classification, such as patent annota-
tion.
Tong Xu received his Ph.D. degree
in computer science in University of Sci-
ence and Technology of China (USTC),
Hefei, in 2016. He is currently working
as an associate researcher of the Anhui
Province Key Laboratory of Big Data
Analysis and Application, University
of Science and Technology of China
(USTC), Hefei. He has authored more than 40 journal and
conference papers in the fields of social network and social
media analysis, including TKDE, TMC, KDD, AAAI,
ICDM, SDM, etc.
En-Hong Chen is a professor and
vice dean of the School of Computer Sci-
ence, University of Science and Techno-
logy of China (USTC), Hefei. His gene-
ral area of research includes data min-
ing and machine learning, social network
analysis, and recommender systems. He
has published more than 100 papers in
refereed conferences and journals, including Nature Com-
munications, IEEE/ACM Transactions, KDD, NIPS, IJ-
CAI and AAAI, etc. He was on program committees of
numerous conferences including KDD, ICDM, and SDM.
He received the Best Application Paper Award on KDD-
2008, the Best Research Paper Award on ICDM-2011, and
the Best of SDM-2015. His research is supported by the Na-
tional Science Foundation for Distinguished Young Scholars
of China.

JOURNAL OF COMPUTER SCIENCE AND TECHNOLOGY
Volume 35, Number 2, March 2020
Special Section on Learning and Mining in Dynamic Environments
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Min-Ling Zhang, Yu-Feng Li, and Qi Liu ( 231 )
Incremental Multi-Label Learning with Active Queries .... .............................................................. .
............................................. Shen-Jun Huang, Guo-Xiang Li, Wen-Yu Huang, and Shao-Yuan Li ( 234 )
Joint Lab el-Specific Features and Correlation Information for Multi-Lab el Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Xiu-Yi Jia, Sai-Sai Zhu, and Wei-Wei Li ( 247 )
Discrimination-Aware Domain Adversarial Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
...................................... Yun-Yun Wang, Jian-Min Gu, Chao Wang, Song-Can Chen, and Hui Xue ( 259 )
Efficient Multiagent Policy O ptimization Based on Weighted Estimators in Stochastic Co operative Environments . . . . . . . . .
............................ Yan Zheng, Jian-Ye Hao, Zong-Zhang Zhang, Zhao-Peng Meng, and Xiao-Tian Hao ( 268 )
Exploiting Structural and Temporal Influence for Dynamic So cial-Aware Recommendation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..
. ................................................... Yang Liu, Zhi Li, Wei Huang, Tong Xu, and En-Hong Chen ( 281 )
Semi-Supervised Classification of Data Streams by BIRCH Ensemble and Local Structure Mapping . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yi-Min Wen and Shuai Liu ( 295 )
Sequential Recommendation via Cross-Domain Novelty Seeking Trait Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.Fu-Zhen Zhuang, Ying-Min Zhou, Hao-Chao Ying, Fu-Zheng Zhang, Xiang Ao, Xing Xie, Qing He, and Hui Xiong ( 305 )
Finding Communities by Decomposing and Embedding Heterogeneous Information Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
................................................... Yue Kou, De-Rong Shen, Dong Li, Tie-Zheng Nie, and Ge Yu ( 320 )
Exploiting Multiple Correlations Among Urban Regions for Crowd Flow Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
................................. Qiang Zhou, Jing-Jing Gu, Chao Ling, Wen-Bo Li, Yi Zhuang, and Jian Wang ( 338 )
You Are How You Behave – Spatiotemporal Representation Learning f or College Student Academic Achievement . . . . . . . . .
............................................................ Xiao-Lin Li, Li Ma, Xiang-Dong He, and Hui Xiong ( 353 )
PetroKG: Construction and Application of Knowledge Graph in Upstream Area of PetroChina . . . . . . . . . . . . . . . . . . . . . . . . . . .
. ......................................... Xiang-Guang Zhou, Ren-Bin Gong, Fu-Geng Shi, and Zhe-Feng Wang ( 368 )
Special Section of ChinaSys 2019
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Wen-Guang Chen, Ying-Wei Luo, and Guang-Yu Sun ( 379 )
Optimistic Transaction Processing in Deterministic Database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . Zhi-Yuan Dong, Chu-Zhe Tang, Jia-Chen Wang, Zhao-Guo Wang, Hai-Bo Chen, and Bin-Yu Zang ( 382 )
MPI-RCDD: A Framework for MPI R untime Communication Deadlock Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
................................. Hong-Mei Wei, Jian Gao, Peng Qing, Kang Yu, Yan-Fei Fang, and Ming-Lu Li ( 395 )
Interference Analysis of Co-Located Container Workloads: A Perspective from Hardware Performance Counters. . . . . . . . . . .
............................. Wen-Yan Chen, Ke-Jiang Ye, Cheng-Zhi Lu, Dong-Dai Zhou, and Cheng-Zhong Xu ( 412 )
IMPULP: A Hardware Approach for In-Process Memory Protection via User-Level Partitioning. . . . . . . . . . . . . . . . . . . . . . . . . . .
.Yang-Yang Zhao, Ming-Yu Chen, Yu-Hang Liu, Zong-Hao Yang, Xiao-Jing Zhu, Zong-Hui Hong, and Yun-Ge Guo ( 418 )
Huge Page Friendly Virtualized Memory Management........................................................ ............
..................................... Sai Sha, Jing-Yuan Hu, Ying-Wei Luo, Xiao-Lin Wang, and Zhenlin Wang ( 433 )
Bigflow: A General Optimization Layer for Distributed Computing Fr ameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
....................Yun-Cong Zhang, Xiao-Yang Wang, Cong Wang, Yao Xu, Jian-Wei Zhang, Xiao-Dong Lin,
Guang-Yu Sun, Gong-Lin Zheng, Shan-Hui Yin, Xian-Jin Ye, Li Li, Zhan Song, and Dong-Dong Miao ( 453 )
A Machine Learning Framework with Feature Selection for Floorplan Acceleration in IC Physical Design . . . . . . . . . . . . . . . . .
.............................................. Shu-Zheng Zhang, Zhen-Yu Zhao, Chao-Chao Feng, and Lei Wang ( 468 )
SIES: A Novel Implementation of Spiking Convolutional Neural Network Inference Engine on Field-Programmable Gate
Array ..................................................................... ...................................... .
.Shu-Quan Wang, Lei Wang, Yu Deng, Zhi-Jie Yang, Sha-Sha Guo, Zi-Yang Kang, Yu-Feng Guo, and Wei-Xia Xu ( 475 )
JOURNAL OF COMPUTER SCIENCE AND TECHNOLOGY
5
O

Å

Æ
E
â
Æ

6
Volume 35 Number 2 2020 (Bimonthly, Started in 1986)
Indexed in: SCIE, Ei, INSPEC, JST, AJ, MR, CA, DBLP
Edited by:
THE EDITORIAL BOARD OF JOURNAL OF COMPUTER SCIENCE AND TECHNOLOGY
Guo-Jie Li, Editor-in-Chief, P.O. Box 2704, Beijing 100190, P.R. China
Managing Editor: Feng-Di Shu E-mail: jcst@ict.ac.cn http://jcst.ict.ac.cn Tel.: 86-10-62610746
Copyright c
Institute of Computing Technology, Chinese Academy of Sciences 2020
Sponsored by: Institute of Computing Technology, CAS & China Computer Federation
Supervised by: Chinese Academy of Sciences
Undertaken by: Institute of Computing Technology, CAS
Published by: Science Press, Beijing, China
Printed by: Beijing Kexin Printing House
Distributed by:
China: All Local Post Offices
Other Countries: Springer Nature Customer Service Center GmbH, Tiergartenstr. 15, 69121 Heidelberg, Germany
Available Online: https://link.springer.com/journal/11390
I
S
Ú

r
Ò
: CN11-2296/TP
I
S
e
u

Ò
: 2-578 RMB U150.00