Graph Neural Networks for Social Recommendation

Wenqi Fan Yao Ma Qing Li

Department of Computer Science Data Science and Engineering Lab Department of Computing
City University of Hong Kong Michigan State University The Hong Kong Polytechnic
[email protected] [email protected] University
[email protected]

Yuan He Eric Zhao Jiliang Tang

JD.com JD.com Data Science and Engineering Lab
[email protected] [email protected] Michigan State University
arXiv:1902.07243v2 [cs.IR] 23 Nov 2019

[email protected]

Dawei Yin
JD.com
[email protected]

ABSTRACT ACM Reference Format:

In recent years, Graph Neural Networks (GNNs), which can Wenqi Fan, Yao Ma, Qing Li, Yuan He, Eric Zhao, Jiliang Tang, and Dawei Yin.
2019. Graph Neural Networks for Social Recommendation. In Proceedings
naturally integrate node information and topological structure,
of the 2019 World Wide Web Conference (WWW ’19), May 13–17, 2019, San
have been demonstrated to be powerful in learning on graph Francisco, CA, USA. ACM, New York, NY, USA, 11 pages. https://doi.org/10.
data. These advantages of GNNs provide great potential to ad- 1145/3308558.3313488
vance social recommendation since data in social recommender
systems can be represented as user-user social graph and user-item
graph; and learning latent factors of users and items is the key.
However, building social recommender systems based on GNNs 1 INTRODUCTION
faces challenges. For example, the user-item graph encodes both
The exploitation of social relations for recommender systems
interactions and their associated opinions; social relations have
has attracted increasing attention in recent years [18, 28, 30].
heterogeneous strengths; users involve in two graphs (e.g., the user-
These social recommender systems have been developed based
user social graph and the user-item graph). To address the three
on the phenomenon that users usually acquire and disseminate
aforementioned challenges simultaneously, in this paper, we present
information through those around them, such as classmates, friends,
a novel graph neural network framework (GraphRec) for social
or colleagues, implying that the underlying social relations of users
recommendations. In particular, we provide a principled approach
can play a significant role in helping them filter information [23].
to jointly capture interactions and opinions in the user-item graph
Hence, social relations have been proven to be helpful in boosting
and propose the framework GraphRec, which coherently models
the recommendation performance [8, 29].
two graphs and heterogeneous strengths. Extensive experiments
Recent years have witnessed great developments in deep neu-
on two real-world datasets demonstrate the effectiveness of the
ral network techniques for graph data [15]. These deep neural
proposed framework GraphRec. Our code is available at https:
network architectures are known as Graph Neural Networks
//github.com/wenqifan03/GraphRec-WWW19
(GNNs) [5, 10, 19], which have been proposed to learn meaningful
representations for graph data. Their main idea is how to iteratively
CCS CONCEPTS aggregate feature information from local graph neighborhoods
• Information systems → Social recommendation; • Com- using neural networks. Meanwhile, node information can be
puting methodologies → Neural networks; Artificial intel- propagated through a graph after transformation and aggregation.
ligence. Hence, GNNs naturally integrate the node information as well
as the topological structure and have been demonstrated to be
KEYWORDS powerful in representation learning [5, 7, 15]. On the other hand,
Social Recommendation; Graph Neural Networks; Recommender data in social recommendation can be represented as graph data
Systems; Social Network; Neural Networks with two graphs. As demonstrated in Figure 1, these two graphs
include a social graph denoting the relationships between users,
This paper is published under the Creative Commons Attribution 4.0 International and a user-item graph denoting interactions between users and
(CC-BY 4.0) license. Authors reserve their rights to disseminate the work on their
personal and corporate Web sites with the appropriate attribution. items. Users are simultaneously involved in both graphs, who can
WWW ’19, May 13–17, 2019, San Francisco, CA, USA bridge them. Moreover, the natural way of social recommendation
© 2019 IW3C2 (International World Wide Web Conference Committee), published is to incorporate the social network information into user and item
under Creative Commons CC-BY 4.0 License.
ACM ISBN 978-1-4503-6674-8/19/05. latent factors learning [37]. Learning representations of items and
https://doi.org/10.1145/3308558.3313488 users is the key to build social recommender systems. Thus, given
their advantages, GNNs provide unprecedented opportunities to framework. Finally, we conclude our work with future directions
advance social recommendation. in Section 5.

User-to-item Interaction
2 THE PROPOSED FRAMEWORK
Social Relations In this section, we will first introduce the definitions and notations
used in this paper, next give an overview about the proposed
3
framework, then detail each model component and finally discuss
how to learn the model parameters.
5

… Table 1: Notation
1

Symbols Definitions and Descriptions

ri j The rating value of item v j by user ui
Figure 1: Graph Data in Social Recommendation. It contains qj The embedding of item v j
two graphs including the user-item graph (left part) and the pi The embedding of user ui
user-user social graph (right part). Note that the number on The opinion embedding for the rating level r ,
er
the edges of the user-item graph denotes the opinions (or such as 5-star rating, r ∈ {1, 2, 3, 4, 5}
rating score) of users on the items via the interactions. d The length of embedding vector
C(i) The set of items which user ui interacted with
The set of social friends who user ui
Meanwhile, building social recommender systems based on N (i)
directly connected with
GNNs faces challenges. The social graph and the user-item graph B(j) The set of users who have interacted the item v j
in a social recommender system provide information about users The item-space user latent factor from
from different perspectives. It is important to aggregate information hiI
item set C(i) of user ui
from both graphs to learn better user representations. Thus, the first The social-space user latent factor from
challenge is how to inherently combine these two graphs. Moreover, hiS
the social friends N (i) of user ui
the user-item graph not only contains interactions between users The user latent factor of user ui , combining
and items but also includes users’ opinions on items. For example, hi
from item space hiI and social space hiS
as shown in Figure 1, the user interacts with the items of “trousers"
The opinion-aware interaction representation
and “laptop"; and the user likes “trousers" while disliking “laptop". xia
of item va for user ui
Therefore, the second challenge is how to capture interactions
The opinion-aware interaction representation
and opinions between users and items jointly. In addition, the low fjt
of user ut for item v j
cost of link formation in online worlds can result in networks
zj The item latent factor of item v j
with varied tie strengths (e.g., strong and weak ties are mixed
The item attention of item va in
together) [36]. Users are likely to share more similar tastes with α ia
strong ties than weak ties. Considering social relations equally contributing to hiI
The social attention of neighboring user uo in
could lead to degradation in recommendation performance. Hence, βio
the third challenge is how to distinguish social relations with contributing to hiS
heterogeneous strengths. The user attention of user ut in
µ jt
In this paper, we aim to build social recommender systems based contributing to zj
on graph neural networks. Specially, we propose a novel graph r i′j The predicted rating value of item v j by user ui
neural network GraphRec for social recommendations, which can ⊕ The concatenation operator of two vectors
address three aforementioned challenges simultaneously. Our major T The user-user social graph
contributions are summarized as follows: R The user-item rating matrix (user-item graph)
• We propose a novel graph neural network GraphRec, which W, b The weight and bias in neural network
can model graph data in social recommendations coherently;
• We provide a principled approach to jointly capture interac-
tions and opinions in the user-item graph; 2.1 Definitions and Notations
• We introduce a method to consider heterogeneous strengths Let U = {u 1 , u 2 , ..., un } and V = {v 1 , v 2 , ..., vm } be the sets of
of social relations mathematically; and users and items respectively, where n is the number of users, and
• We demonstrate the effectiveness of the proposed framework m is the number of items. We assume that R ∈ Rn×m is the user-
on various real-world datasets. item rating matrix, which is also called the user-item graph. If ui
The remainder of this paper is organized as follows. We introduce gives a rating to v j , r i j is the rating score, otherwise we employ
the proposed framework in Section 2. In Section 3, we conduct ex- 0 to represent the unknown rating from ui to v j , i.e., r i j = 0. The
periments on two real-world datasets to illustrate the effectiveness observed rating score  r i j can be seen as user ui ’s opinion on the
of the proposed method. In Section 4, we review work related to our item v j . Let O = ui , v j |r i j , 0 be the set of known ratings
 Rating Prediction

r'

Concatenation

User Modeling Item Modeling

User Latent Factor

… Item Latent Factor

μ3
μ2 Attention Network

μ1
Concatenation
… …
Item-space Social-space

α1 β1
α2
Attention Network β2
Attention Network
α3
β3
… … … …
Item-space Item-space Item-space

User Aggregation

Item-space: Item-space User Latent Factor

Social-space: Social-space User Latent Factor

Item Embedding

User Embedding
Item Aggregation Social Aggregation
Opinion Embedding

Figure 2: The overall architecture of the proposed model. It contains three major components: user modeling, item modeling,
and rating prediction.

and T = ui , v j |r i j = 0 be the set of unknown ratings. Let N (i)


understand users via interactions between users and items in the
be the set of users whom ui directly connected with, C(i) be the user-item graph (or item-space). The other is social aggregation,
set of items which ui have interacted with, and B(j) be the set of the relationship between users in the social graph, which can help
users who have interacted with v j . In addition, users can establish model users from the social perspective (or social-space). Then, it
social relations to each other. We use T ∈ Rn×n to denote the user- is intuitive to obtain user latent factors by combining information
user social graph, where Ti j = 1 if u j has a relation to ui and zero from both item space and social space. The second component is
otherwise. Given the user-item graph R and social graph T, we aim item modeling, which is to learn latent factors of items. In order to
to predict the missing rating value in R. Following [11], we use an consider both interactions and opinions in the user-item graph, we
embedding vector pi ∈ Rd to denote a user ui and an embedding introduce user aggregation, which is to aggregate users’ opinions in
vector qj ∈ Rd to represent an item v j , where d is the length item modeling. The third component is to learn model parameters
of embedding vector. More details will be provided about these via prediction by integrating user and item modeling components.
embedding vectors in the following subsections. The mathematical Next, we will detail each model component.
notations used in this paper are summarized in Table 1.
2.3 User Modeling
2.2 An Overview of the Proposed Framework User modeling aims to learn user latent factors, denoted as hi ∈ Rd
The architecture of the proposed model is shown in Figure 2. The for user ui . The challenge is how to inherently combine the user-
model consists of three components: user modeling, item modeling, item graph and social graph. To address this challenge, we first
and rating prediction. The first component is user modeling, which use two types of aggregation to learn factors from two graphs, as
is to learn latent factors of users. As data in social recommender shown in the left part in Figure 2. The first aggregation, denoted as
systems includes two different graphs, i.e., a social graph and a item aggregation, is utilized to learn item-space user latent factor
user-item graph, we are provided with a great opportunity to hiI ∈ Rd from the user-item graph. The second aggregation is
learn user representations from different perspectives. Therefore, social aggregation where social-space user latent factor hiS ∈ Rd is
two aggregations are introduced to respectively process these two learned from the social graph. Then, these two factors are combined
different graphs. One is item aggregation, which can be utilized to together to form the final user latent factors hi . Next, we will
introduce item aggregation, social aggregation and how to combine To alleviate the limitation of mean-based aggregator, inspired by
user latent factors from both item-space and social-space. attention mechanisms [3, 38], an intuitive solution is to tweak α i
Item Aggregation. As user-item graph contains not only to be aware of the target user ui , i.e., assigning an individualized
interactions between users and items but also users’ opinions weight for each (va , ui ) pair,
(or rating scores) on items, we provide a principled approach to  
 
jointly capture interactions and opinions in the user-item graph for I
 Õ
hi = σ (W · α ia xia + b)
 

(4)
learning item-space user latent factors hiI , which is used to model  a ∈C(i)
 

user latent factor via interactions in the user-item graph.  
The purpose of item aggregation is to learn item-space user where α ia denotes the attention weight of the interaction with
latent factor hiI by considering items a user ui has interacted with va in contributing to user ui ’s item-space latent factor when
and users’ opinions on these items. To mathematically represent characterizing user ui ’s preference from the interaction history C(i).
this aggregation, we use the following function as: Specially, we parameterize the item attention α ia with a two-layer
neural network, which we call as the attention network. The input
hiI = σ (W · Aддreit ems ({xia , ∀a ∈ C(i)}) + b) (1) to the attention network is the opinion-aware representation xia
of the interaction and the target user ui ’s embedding pi . Formally,
where C(i) is the set of items user ui has interacted with (or ui ’s the attention network is defined as,
neighbors in the user-item graph), xia is a representation vector
to denote opinion-aware interaction between ui and an item va ,
∗
α ia = wT2 · σ (W1 · [xia ⊕ pi ] + b1 ) + b2 (5)
and Aддreit ems is the items aggregation function. In addition, σ The final attention weights are obtained by normalizing the
denotes non-linear activation function (i.e., a rectified linear unit), above attentive scores using Softmax function, which can be
and W and b are the weight and bias of a neural network. Next we interpreted as the contribution of the interaction to the item-space
will discuss how to define opinion-aware interaction representation user latent factor of user ui as:
xia and the aggregation function Aддreit ems .
exp(α ia
∗ )
A user can express his/her opinions (or rating scores), denoted as α ia = Í (6)
a ∈C(i) exp(α ia )
∗
r , to items during user-item interactions. These opinions on items
can capture users’ preferences on items, which can help model Social Aggregation. Due to the social correlation theories [20,
item-space user latent factors. To model opinions, for each type of 21], a user’s preference is similar to or influenced by his/her
opinions r , we introduce an opinion embedding vector er ∈ Rd directly connected social friends. We should incorporate social
that denotes each opinion r as a dense vector representation. For information to further model user latent factors. Meanwhile, tie
example, in a 5-star rating system, for each r ∈ {1, 2, 3, 4, 5}, we strengths between users can further influence users’ behaviors from
introduce an embedding vector er . For an interaction between user the social graph. In other words, the learning of social-space user
ui and item va with opinion r , we model opinion-aware interaction latent factors should consider heterogeneous strengths of social
representation xia as a combination of item embedding qa and relations. Therefore, we introduce an attention mechanism to select
opinion embedding er via a Multi-Layer Perceptron (MLP). It can social friends that are representative to characterize users social
be denoted as дv to fuse the interaction information with the information and then aggregate their information.
opinion information as shown in Figure 2. The MLP takes the In order to represent user latent factors from this social per-
concatenation of item embedding qa and its opinion embedding er spective, we propose social-space user latent factors, which is to
as input. The output of MLP is the opinion-aware representation of aggregate the item-space user latent factors of neighboring users
the interaction between ui and va , xia , as follows: from the social graph. Specially, the social-space user latent factor
of ui , hiS , is to aggregate the item-space user latent factors of users
xia = дv ([qa ⊕ er ]) (2)
in ui ’s neighbors N (i), as the follows:
where ⊕ denotes the concatenation operation between two vectors.
hiS = σ (W · Aддreneiдbhor s ( hoI , ∀o ∈ N (i) ) + b)
n o
(7)
One popular aggregation function for Aддreit ems is the mean
operator where we take the element-wise mean of the vectors in where Aддreneiдbhor s denotes the aggregation function on user’s
{xia , ∀a ∈ C(i)}. This mean-based aggregator is a linear approxi- neighbors.
mation of a localized spectral convolution [15], as the following One natural aggregation function for Aддreneiдbhor s is also the
function: mean
 I operator which take the element-wise mean of the vectors in

 
 ho , ∀o ∈ N (i) , as the following function:
I
 Õ
hi = σ (W · α i xia + b)
 

(3) 
 

hiS = σ (W · βi hoI + b)
 a ∈C(i)
   Õ
 

(8)

 
o ∈N (i)
 

where α i is fixed to 1 for all items in the mean-based aggregator.  
|C(i)|
It assumes that all interactions contribute equally to understand the where βi is fixed to |N1(i)| for all neighbors for the mean-based
user ui . However, this may not be optimal, due to the fact that the aggregator. It assumes that all neighbors contribute equally to the
influence of interactions on users may vary dramatically. Hence, representation of user ui . However, as mentioned before, strong
we should allow interactions to contribute differently to a user’s and weak ties are mixed together in a social network, and users are
latent factor by assigning each interaction a weight. likely to share more similar tastes with strong ties than weak ties.
 in fjt , ∀t ∈ B(j) as:

Thus, we perform an attention mechanism with a two-layer neural
network to extract these users that are important to influence ui ,
zj = σ (W · Aддreuser s ( fjt , ∀t ∈ B(j) ) + b)

(16)
and model their tie strengths, by relating social attention βio with
hoI and the target user embedding pi , as below, In addition, we introduce an attention mechanism to differentiate
the importance weight µ jt of users with a two-layer neural attention
network, taking fjt and qj as the input,

 

hiS = σ (W · βio hoI + b)
 Õ
 

(9)  
o ∈N (i)
 
 
 Õ 
zj = σ (W · µ jt fjt + b)
 

  (17)
∗
βio = wT2 · σ (W1 · [hoI ⊕ pi ] + b1 ) + b2 (10)  t ∈B(j)
 

 
exp(βio
∗ )
µ ∗jt = wT2 · σ (W1 · [fjt ⊕ qj ] + b1 ) + b2 (18)
βio = Í (11)
o ∈N (i) exp(βio )
∗
exp(µ ∗jt )
where the βio can be seen as the strengths between users. µ jt = Í (19)
t ∈B (j) exp(µ jt )
∗
Learning User Latent Factor. In order to learn better user
latent factors, item-space user latent factors and social-space user This user attention µ jt is to capture heterogeneous influence from
latent factors are needed to be considered together, since the social user-item interactions on learning item latent factor.
graph and the user-item graph provide information about users
from different perspectives. We propose to combine these two latent
2.5 Rating Prediction
factors to the final user latent factor via a standard MLP where the In this subsection, we will design recommendation tasks to learn
item-space user latent factor hiI and the social-space user latent model parameters. There are various recommendation tasks such
factor hiS are concatenated before feeding into MLP. Formally, the as item ranking and rating prediction. In this work, we apply the
user latent factor hi is defined as, proposed GraphRec model for the recommendation task of rating
 users and items (i.e., hi and
prediction. With the latent factors of
c1 = hiI ⊕ hiS
h i
(12) zj ), we can first concatenate them hi ⊕ zj and then feed it into
MLP for rating prediction as:
c2 = σ (W2 · c1 + b2 ) (13)
g1 = hi ⊕ zj
 
... (20)

hi = σ (Wl · cl −1 + bl ) (14) g2 = σ (W2 · g1 + b2 ) (21)

...
where l is the index of a hidden layer.
gl −1 = σ (Wl · gl −1 + bl ) (22)
T
2.4 Item Modeling r i′j = w · gl −1 (23)
As shown in the right part of Figure 2, item modeling is used to learn where l is the index of a hidden layer, and r i′j is the predicted rating
item latent factor, denoted as zj , for the item v j by user aggregation. from ui to v j .
Items are associated with the user-item graph, which contains
interactions as well as user’s opinions. Therefore, interactions and 2.6 Model Training
opinions in the user-item graph should be jointly captured to further To estimate model parameters of GraphRec, we need to specify an
learn item latent factors. objective function to optimize. Since the task we focus on in this
User Aggregation. Likewise, we use a similar method as work is rating prediction, a commonly used objective function is
learning item-space user latent factors via item aggregation. For formulated as,
each item v j , we need to aggregate information from the set of
users who have interacted with v j , denoted as B(j). 1 Õ ′
Even for the same item, users might express different opinions Loss = (r i j − r i j )2 (24)
2 |O|
during user-item interactions. These opinions from different users i, j ∈ O
can capture the characteristics of the same item in different ways where |O| is the number of observed ratings , and r i j is the ground
provided by users, which can help model item latent factors. For an truth rating assigned by the user i on the item j.
interaction from ut to v j with opinion r , we introduce an opinion- To optimize the objective function, we adopt the RMSprop [31]
aware interaction user representation fjt , which is obtained from as the optimizer in our implementation, rather than the vanilla
the basic user embedding pt and opinion embedding er via a MLP, SGD. At each time, it randomly selects a training instance and
denoted as дu . дu is to fuse the interaction information with the updates each model parameter towards the direction of its negative
opinion information, as shown in Figure 2: gradient. There are three embedding in our model, including
item embedding qj , user embedding pi , and opinion embedding
fjt = дu ([pt ⊕ er ]) (15)
er . They are randomly initialized and jointly learned during the
Then, to learn item latent factor zj , we also propose to aggregate training stage. We do not use one-hot vectors to represent each
opinion-aware interaction representation of users in B(j) for item user and item, since the raw features are very large and highly
v j . The users aggregation function is denoted as Aддreuser s , which sparse. By embedding high-dimensional sparse features into a
is to aggregate opinion-aware interaction representation of users low-dimensional latent space, the model can be easy to train [11].
Opinion embedding matrix e depends on the rating scale of the • SoReg [18]: Social Regularization models social network
system. For example, for a 5-star rating system, opinion embedding information as regularization terms to constrain the matrix
matrix e contains 5 different embedding vectors to denote scores factorization framework.
in {1, 2, 3, 4, 5}. Overfitting is a perpetual problem in optimizing • SocialMF [13]: It considers the trust information and prop-
deep neural network models. To alleviate this issue, the dropout agation of trust information into the matrix factorization
strategy [26] has been applied to our model. The idea of dropout is model for recommender systems.
to randomly drop some neurons during the training process. When • TrustMF [37]: This method adopts matrix factorization
updating parameters, only part of them will be updated. Moreover, technique that maps users into two low-dimensional spaces:
as dropout is disabled during testing, the whole network is used for truster space and trustee space, by factorizing trust networks
prediction. according to the directional property of trust.
• NeuMF [11]: This method is a state-of-the-art matrix factor-
3 EXPERIMENT ization model with neural network architecture. The original
implementation is for recommendation ranking task and we
3.1 Experimental Settings
adjust its loss to the squared loss for rating prediction.
3.1.1 Datasets. In our experiments, we choose two representative • DeepSoR [8]: This model employs a deep neural network
datasets Ciao and Epinions1 , which are taken from popular social to learn representations of each user from social relations,
networking websites Ciao (http://www.ciao.co.uk) and Epinions and to integrate into probabilistic matrix factorization for
(www.epinions.com). Each social networking service allows users rating prediction.
to rate items, browse/write reviews, and add friends to their ‘Circle • GCMC+SN [1]: This model is a state-of-the-art recom-
of Trust’. Hence, they provide a large amount of rating information mender system with graph neural network architecture. In
and social information. The ratings scale is from 1 to 5. We randomly order to incorporate social network information into GCMC,
initialize opinion embedding with 5 different embedding vectors we utilize the node2vec [9] to generate user embedding as
based on 5 scores in {1, 2, 3, 4, 5}. The statistics of these two datasets user side information, instead of using the raw feature social
are presented in Table 2. connections (T ∈ Rn×n ) directly. The reason is that the raw
feature input vectors is highly sparse and high-dimensional.
Table 2: Statistics of the datasets Using the network embedding techniques can help compress
the raw input feature vector to a low-dimensional and dense
Dataset Ciao Epinions vector, then the model can be easy to train.
# of Users 7,317 18,088
PMF and NeuMF are pure collaborative filtering model without
# of Items 10,4975 261,649
social network information for rating prediction, while the others
# of Ratings 283,319 764,352
are social recommendation. Besides, we compared GraphRec with
# of Density (Ratings) 0.0368% 0.0161%
two state-of-the-art neural network based social recommender
# of Social Connections 111,781 355,813 systems, i.e., DeepSoR and GCMC+SN.
# of Density (Social Relations) 0.2087% 0.1087%
3.1.4 Parameter Settings. We implemented our proposed method
3.1.2 Evaluation Metrics. In order to evaluate the quality of the on the basis of Pytorch2 , a well-known Python library for neural
recommendation algorithms, two popular metrics are adopted to networks. For each dataset, we used x% as a training set to learning
evaluate the predictive accuracy, namely Mean Absolute Error parameters, (1 − x%)/2 as a validation set to tune hyper-parameters,
(MAE) and Root Mean Square Error (RMSE) [34]. Smaller values and (1 −x%)/2 as a testing set for the final performance comparison,
of MAE and RMSE indicate better predictive accuracy. Note that where x was varied as {80%, 60%}. For the embedding size d, we
small improvement in RMSE or MAE terms can have a significant tested the value of [ 8, 16, 32, 64, 128, 256 ]. The batch size and
impact on the quality of the top-few recommendations [16]. learning rate was searched in [ 32, 64, 128, 512 ] and [ 0.0005,
0.001, 0.005, 0.01, 0.05, 0.1 ], respectively. Moreover, we empirically
3.1.3 Baselines. To evaluate the performance, we compared our set the size of the hidden layer the same as the embedding size
GraphRec with three groups of methods including traditional and the activation function as ReLU. Without special mention,
recommender systems, traditional social recommender systems, we employed three hidden layers for all the neural components.
and deep neural network based recommender systems. For each The early stopping strategy was performed, where we stopped
group, we select representative baselines and below we will detail training if the RMSE on validation set increased for 5 successive
them. epochs. For all neural network methods, we randomly initialized
• PMF [24]: Probabilistic Matrix Factorization utilizes user- model parameters with a Gaussian distribution, where the mean and
item rating matrix only and models latent factors of users standard deviation is 0 and 0.1, respectively. The parameters for the
and items by Gaussian distributions. baseline algorithms were initialized as in the corresponding papers
• SoRec [17]: Social Recommendation performs co-factorization and were then carefully tuned to achieve optimal performance.
on the user-item rating matrix and user-user social relations
matrix.
1 http://www.cse.msu.edu/∼tangjili/index.html 2 https://pytorch.org/
 Table 3: Performance comparison of different recommender systems

Algorithms
Training Metrics
PMF SoRec SoReg SocialMF TrustMF NeuMF DeepSoR GCMC+SN GraphRec
Ciao MAE 0.952 0.8489 0.8987 0.8353 0.7681 0.8251 0.7813 0.7697 0.7540
(60%) RMSE 1.1967 1.0738 1.0947 1.0592 1.0543 1.0824 1.0437 1.0221 1.0093
Ciao MAE 0.9021 0.8410 0.8611 0.8270 0.7690 0.8062 0.7739 0.7526 0.7387
(80%) RMSE 1.1238 1.0652 1.0848 1.0501 1.0479 1.0617 1.0316 0.9931 0.9794
Epinions MAE 1.0211 0.9086 0.9412 0.8965 0.8550 0.9097 0.8520 0.8602 0.8441
(60%) RMSE 1.2739 1.1563 1.1936 1.1410 1.1505 1.1645 1.1135 1.1004 1.0878
Epinions MAE 0.9952 0.8961 0.9119 0.8837 0.8410 0.9072 0.8383 0.8590 0.8168
(80%) RMSE 1.2128 1.1437 1.1703 1.1328 1.1395 1.1476 1.0972 1.0711 1.0631

3.2 Performance Comparison of Recommender 3.3 Model Analysis

Systems In this subsection, we study the impact of model components and
We first compare the recommendation performance of all methods. model hyper-parameters.
Table 3 shows the overall rating prediction error w.r .t . RMSE and 3.3.1 Effect of Social Network and User Opinions. In the last
MAE among the recommendation methods on Ciao and Epinions subsection, we have demonstrated the effectiveness of the proposed
datasets. We have the following main findings: framework. The proposed framework provides model components
to (1) integrate social network information and (2) incorporate
users’ opinions about the interactions with items. To understand
• SoRec, SoReg, SocialMF, and TrustMF always outperform
the working of GraphRec, we compare GraphRec with its two
PMF. All of these methods are based on matrix factorization.
variants: GraphRec-SN, and GraphRec-Opinion. These two variants
SoRec, SoReg, SocialMF, and TrustMF leverage both the
are defined in the following:
rating and social network information; while PMF only uses
the rating information. These results support that social • GraphRec-SN: The social network information of GraphRec
network information is complementary to rating information is removed . This variant only uses the item-space user latent
for recommendations. factor hiI to represent user latent factors hi ; while ignoring
• NeuMF obtains much better performance than PMF. Both the social-space user latent factors hiS .
methods only utilize the rating information. However, NeuMF • GraphRec-Opinion: For learning item-space user latent
is based on neural network architecture, which suggests the factor and item latent factor, the opinion embedding is
power of neural network models in recommender systems. removed during learning xia and fjt . This variant ignores
• DeepSoR and GCMC+SN perform better than SoRec, SoReg, the users’ opinion on the user-item interactions.
SocialMF, and TrustMF. All of them take advantage of both The performance of GraphRec and its variants on Ciao and
rating and social network information. However, DeepSoR Epinions are given in Figure 3. From the results, we have the
and GCMC+SN are based on neural network architectures, following findings:
which further indicate the power of neural network models • Social Network Information. We now focus on analyzing
in recommendations. the effectiveness of social network information. GraphRec-
• Among baselines, GCMC+SN shows quite strong perfor- SN performs worse than GraphRec. It verifies that social
mance. It implies that the GNNs are powerful in representa- network information is important to learn user latent factors
tion learning for graph data, since it naturally integrates the and boost the recommendation performance.
node information as well as topological structure. • Opinions in Interaction. We can see that without opin-
• Our method GraphRec consistently outperforms all the ion information, the performance of rating prediction is
baseline methods. Compared to DeepSoR and GCMC+SN, deteriorated significantly. For example, on average, the
our model provides advanced model components to integrate relative reduction on Ciao and Epinions is 3.50% and 2.64%
rating and social network information. In addition, our model on RMSE metric, and 5.84% and 5.02% on MAE metric,
provides a way to consider both interactions and opinions in respectively. It justifies our assumption that opinions on
the user-item graph. We will provide further investigations user-item interactions have informative information that
to better understand the contributions of model components can help to learn user or item latent factors and improve the
to the proposed framework in the following subsection. performance of recommendation.

3.3.2 Effect of Attention Mechanisms. To get a better understand-

To sum up, the comparison results suggest (1) social network ing of the proposed GraphRec model, we further evaluate the
information is helpful for recommendations; (2) neural network key components of GraphRec - Attention mechanisms. There are
models can boost recommendation performance and (3) the pro- three different attention mechanisms during aggregation, including
posed framework outperforms representative baselines. item attention α, social attention β, and user attention µ. We
 0.88
1.02 0.80 1.09 0.87
1.01 0.86
0.78 1.08 0.85
1.00
1.07 0.84
0.99
RMSE

RMSE
MAE

MAE
0.76 0.83
0.98
1.06 0.82
0.97 0.74 0.81
1.05
0.96 0.80
0.95 0.72 1.04 0.79
Rec-SN Rec-Opi
nion Graph
Rec Rec-SN Rec-Opi
nion Graph
Rec Rec-SN Rec-Opi
nion Graph
Rec Rec-SN Rec-Opi
nion Graph
Rec
Graph Graph Graph Graph Graph Graph Graph Graph

(a) Ciao-RMSE (b) Ciao-MAE (c) Epinions-RMSE (d) Epinions-MAE

Figure 3: Effect of social network and user opinions on Ciao and Epinions datasets.

0.995 0.750 0.840

1.080 0.835
0.990 0.745
0.830
0.740 1.075
0.985 0.825

RMSE
RMSE

0.735
MAE

MAE
0.820
1.070
0.980
0.730 0.815

0.975 1.065 0.810

0.725
0.805
0.970 0.720 1.060
Rec- Rec- Rec- & GraphRec
-
Graph
Rec Rec- Rec- Rec- & GraphRec
-
Graph
Rec Rec- Rec- Rec- & GraphRec
-
Graph
Rec Rec- Rec- Rec- & GraphRec
-
Graph
Rec
Graph Graph Graph Graph Graph Graph Graph Graph Graph Graph Graph Graph

(a) Ciao-RMSE (b) Ciao-MAE (c) Epinions-RMSE (d) Epinions-MAE

Figure 4: Effect of attention mechanisms on Ciao and Epinions datasets.

1.005 0.78 1.090 0.850

1.000 1.085 0.845

0.77
1.080 0.840
0.995
0.76 1.075 0.835
0.990
RMSE

RMSE

0.830
MAE

MAE
1.070
0.985 0.75
1.065 0.825
0.980
1.060 0.820
0.74
0.975 0.815
1.055
0.970 8 16 32 64 128 256
0.73 8 16 32 64 128 256 8 16 32 64 128 256
0.810 8 16 32 64 128 256

(a) Ciao-RMSE (b) Ciao-MAE (c) Epinions-RMSE (d) Epinions-MAE

Figure 5: Effect of embedding size on Ciao and Epinions datasets.

compare GraphRec with its four variants: GraphRec-α, GraphRec-β, • GraphRec-α&β: This variant eliminates two attention mech-
GraphRec-α&β, and GraphRec-µ. These four variants are defined anisms (item attention α and social attention β) on item
in the following: aggregation and social aggregation for modeling user latent
factors.
• GraphRec-µ: The user attention µ of GraphRec is eliminated
• GraphRec-α: The item attention α of GraphRec is eliminated during aggregating opinion-aware interaction user represen-
during aggregating the opinion-aware interaction repre- tation. This variant employs the mean-based aggregation
sentation of items. This variant employs the mean-based function on user aggregation for modeling item latent
aggregation function on item aggregation for modeling item- factors.
space user latent factors.
• GraphRec-β: The social attention α is to model users’ tie The results of different attention mechanisms on GraphRec are
strengths. The social attention α of GraphRec in this variant shown in Figure 4. From the results, we have the following findings,
is eliminated during aggregating user’s neighbors. This • Not all interacted items (purchased history) of one user
variant employs the mean-based aggregation function on contribute equally to the item-space user latent factor, and
social aggregation for modeling social-space user latent not all interacted users (buyers) have the same importance
factors. to learning item latent factor. Based on these assumptions,
 our model considers these difference among users and items such as speech recognition [12], Computer Vision (CV) [14] and
by using two different attention mechanisms (α and µ). From Natural Language Processing (NLP) [4]. Some recent efforts have
the results, we can observe that GraphRec-α and GraphRec- applied deep neural networks to recommendation tasks and shown
µ obtain worse performance than GraphRec. These results promising results [41], but most of them used deep neural networks
demonstrate the benefits of the attention mechanisms on to model audio features of music [32], textual description of
item aggregation and user aggregation. items [3, 33], and visual content of images [40]. Besides, NeuMF [11]
• As mentioned before, users are likely to share more sim- presented a Neural Collaborative Filtering framework to learn the
ilar tastes with strong ties than weak ties. The attention non-linear interactions between users and items.
mechanism β at social aggregation considers heterogeneous However, the application of deep neural network in social
strengths of social relations. When the attention mechanism recommender systems is rare until very recently. In particular,
β is removed, the performance of GraphRec-β is dropped NSCR [35] extended the NeuMF [11] model to cross-domain
significantly. It justifies our assumption that during social social recommendations, i.e., recommending items of information
aggregation, different social friends should have different domains to potential users of social networks, and presented a
influence for learning social-space user latent factor. It’s neural social collaborative ranking recommender system. However,
important to distinguish social relations with heterogeneous the limitation is NSCR requires users with one or more social
strengths. networks accounts (e.g., Facebook, Twitter, Instagram), which
To sum up, GraphRec can capture the heterogeneity in ag- limits the data collections and its applications in practice. SMR-
gregation operations of the proposed framework via attention MNRL [42] developed social-aware movie recommendation in social
mechanisms, which can boost the recommendation performance. media from the viewpoint of learning a multimodal heterogeneous
network representation for ranking. They exploited the recurrent
3.3.3 Effect of Embedding Size. In this subsection, to analyze the neural network and convolutional neural network to learn the
effect of embedding size of user embedding p , item embedding q, representation of movies’ textual description and poster image, and
and opinion embedding e, on the performance of our model. adopted a random-walk based learning method into multimodal
Figure 5 presents the performance comparison w.r .t . the length neural networks. In all these works [35] [42], they addressed the
of embedding of our proposed model on Ciao and Epinions datasets. task of cross-domain social recommendations for ranking metric,
In general, with the increase of the embedding size, the performance which is different from traditional social recommender systems.
first increases and then decreases. When increasing the embedding Most related to our task with neural networks includes DLMF [6]
size from 8 to 64 can improve the performance significantly. and DeepSoR [8]. DLMF [6] used auto-encoder on ratings to learn
However, with the embedding size of 256, GraphRec degrades the representation for initializing an existing matrix factorization. A
performance. It demonstrates that using a large number of the two-phase trust-aware recommendation process is proposed to
embedding size has powerful representation. Nevertheless, if the utilize deep neural networks in matrix factorization’s initialization
length of embedding is too large, the complexity of our model and to synthesize the user’s interests and their trust friends’
will significantly increase. Therefore, we need to find a proper interests together with the impact of community effect based on
length of embedding in order to balance the trade-off between the matrix factorization for recommendations. DeepSoR [8] integrated
performance and the complexity. neural networks for user’s social relations into probabilistic matrix
factorization. They first represented users using pre-trained node
4 RELATED WORK embedding technique, and further exploited k-nearest neighbors to
In this section, we briefly review some related work about social bridge user embedding features and neural network.
recommendation, deep neural network techniques employed for More recently, Graph Neural Networks (GNNs) have been proven
recommendation, and the advanced graph neural networks. to be capable of learning on graph structure data [2, 5, 7, 15, 25].
Exploiting social relations for recommendations has attracted In the task of recommender systems, the user-item interaction
significant attention in recent years [27, 28, 37]. One common contains the ratings on items by users, which is a typical graph data.
assumption about these models is that a user’s preference is similar Therefore, GNNs have been proposed to solve the recommendation
to or influenced by the people around him/her (nearest neighbours), problem [1, 22, 39]. sRMGCNN [22] adopted GNNs to extract
which can be proven by social correlation theories [20, 21]. Along graph embeddings for users and items, and then combined with
with this line, SoRec [17] proposed a co-factorization method, which recurrent neural network to perform a diffusion process. GCMC [1]
shares a common latent user-feature matrix factorized by ratings proposed a graph auto-encoder framework, which produced latent
and by social relations. TrustMF [37] modeled mutual influence features of users and items through a form of differentiable message
between users, and mapped users into two low-dimensional spaces: passing on the user-item graph. PinSage [39] proposed a random-
truster space and trustee space, by factorizing social trust networks. walk graph neural network to learn embedding for nodes in web-
SoDimRec [30] first adopted a community detection algorithm scale graphs. Despite the compelling success achieved by previous
to partition users into several clusters, and then exploited the work, little attention has been paid to social recommendation with
heterogeneity of social relations and weak dependency connec- GNNs. In this paper, we propose a graph neural network for social
tions for recommendation. Comprehensive overviews on social recommendation to fill this gap.
recommender systems can be found in surveys [29].
In recent years, deep neural network models had a great impact
on learning effective feature representations in various fields,
5 CONCLUSION AND FUTURE WORK [13] Mohsen Jamali and Martin Ester. 2010. A matrix factorization technique with
trust propagation for recommendation in social networks. In Proceedings of the
We have presented a Graph Network model (GraphRec) to model fourth ACM conference on Recommender systems. ACM, 135–142.
social recommendation for rating prediction. Particularly, we [14] Hamid Karimi, Jiliang Tang, and Yanen Li. 2018. Toward End-to-End Deception
Detection in Videos. In 2018 IEEE Big Data. IEEE, 1278–1283.
provide a principled approach to jointly capture interactions and [15] Thomas N. Kipf and Max Welling. 2017. Semi-Supervised Classification
opinions in the user-item graph. Our experiments reveal that the with Graph Convolutional Networks. In International Conference on Learning
opinion information plays a crucial role in the improvement of our Representations (ICLR).
[16] Yehuda Koren. 2008. Factorization meets the neighborhood: a multifaceted
model performance. In addition, our GraphRec can differentiate collaborative filtering model. In Proceedings of the 14th ACM SIGKDD international
the ties strengths by considering heterogeneous strengths of social conference on Knowledge discovery and data mining. ACM, 426–434.
relations. Experimental results on two real-world datasets show [17] Hao Ma, Haixuan Yang, Michael R Lyu, and Irwin King. 2008. Sorec: social
recommendation using probabilistic matrix factorization. In Proceedings of the
that GraphRec can outperform state-of-the-art baselines. 17th ACM conference on Information and Knowledge Management. ACM, 931–940.
Currently we only incorporate the social graph into recommen- [18] Hao Ma, Dengyong Zhou, Chao Liu, Michael R Lyu, and Irwin King. 2011.
Recommender systems with social regularization. In Proceedings of the fourth
dation, while many real-world industries are associated rich other ACM international conference on Web Search and Data Mining. ACM, 287–296.
side information on users as well as items. For example, users [19] Yao Ma, Suhang Wang, Charu C. Aggarwal, Dawei Yin, and Jiliang Tang. 2019.
and items are associated with rich attributes. Therefore, exploring Multi-dimensional Graph Convolutional Networks. In Proceedings of the 2019
SIAM International Conference on Data Mining(SDM).
graph neural networks for recommendation with attributes would [20] Peter V Marsden and Noah E Friedkin. 1993. Network studies of social influence.
be an interesting future direction. Beyond that, now we consider Sociological Methods & Research 22, 1 (1993), 127–151.
both rating and social information static. However, rating and social [21] Miller McPherson, Lynn Smith-Lovin, and James M Cook. 2001. Birds of a feather:
Homophily in social networks. Annual review of sociology 27, 1 (2001), 415–444.
information are naturally dynamic. Hence, we will consider building [22] Federico Monti, Michael Bronstein, and Xavier Bresson. 2017. Geometric matrix
dynamic graph neural networks for social recommendations with completion with recurrent multi-graph neural networks. In Advances in Neural
Information Processing Systems. 3700–3710.
dynamic. [23] Paul Resnick and Hal R Varian. 1997. Recommender systems. Commun. ACM 40,
3 (1997), 56–58.
ACKNOWLEDGMENTS [24] Ruslan Salakhutdinov and Andriy Mnih. 2007. Probabilistic Matrix Factorization.
In 21th Conference on Neural Information Processing Systems, Vol. 1. 2–1.
The work described in this paper has been supported, in part, by a [25] David I Shuman, Sunil K Narang, Pascal Frossard, Antonio Ortega, and Pierre
general research fund from the Hong Kong Research Grants Council Vandergheynst. 2013. The emerging field of signal processing on graphs:
Extending high-dimensional data analysis to networks and other irregular
(project PolyU 1121417/17E), and an internal research grant from domains. IEEE Signal Processing Magazine 30, 3 (2013), 83–98.
the Hong Kong Polytechnic University (project 1.9B0V). Yao Ma and [26] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan
Jiliang Tang are supported by the National Science Foundation (NSF) Salakhutdinov. 2014. Dropout: a simple way to prevent neural networks from
overfitting. The Journal of Machine Learning Research 15, 1 (2014), 1929–1958.
under grant numbers IIS-1714741, IIS-1715940 and CNS-1815636, [27] Jiliang Tang, Charu Aggarwal, and Huan Liu. 2016. Recommendations in signed
and a grant from Criteo Faculty Research Award. social networks. In Proceedings of the 25th International Conference on World Wide
Web. International World Wide Web Conferences Steering Committee, 31–40.
[28] Jiliang Tang, Xia Hu, Huiji Gao, and Huan Liu. 2013. Exploiting local and global
REFERENCES social context for recommendation.. In IJCAI, Vol. 13. 2712–2718.
[1] Rianne van den Berg, Thomas N Kipf, and Max Welling. 2017. Graph [29] Jiliang Tang, Xia Hu, and Huan Liu. 2013. Social recommendation: a review.
convolutional matrix completion. arXiv preprint arXiv:1706.02263 (2017). Social Network Analysis and Mining 3, 4 (2013), 1113–1133.
[2] Michael M Bronstein, Joan Bruna, Yann LeCun, Arthur Szlam, and Pierre [30] Jiliang Tang, Suhang Wang, Xia Hu, Dawei Yin, Yingzhou Bi, Yi Chang, and Huan
Vandergheynst. 2017. Geometric deep learning: going beyond euclidean data. Liu. 2016. Recommendation with Social Dimensions. In AAAI. 251–257.
IEEE Signal Processing Magazine 34, 4 (2017), 18–42. [31] Tijmen Tieleman and Geoffrey Hinton. 2012. Lecture 6.5-RMSProp, COURSERA:
[3] Chong Chen, Min Zhang, Yiqun Liu, and Shaoping Ma. 2018. Neural Attentional Neural networks for machine learning. University of Toronto, Technical Report
Rating Regression with Review-level Explanations. In Proceedings of the 27th (2012).
International Conference on World Wide Web. 1583–1592. [32] Aaron Van den Oord, Sander Dieleman, and Benjamin Schrauwen. 2013.
[4] Hongshen Chen, Xiaorui Liu, Dawei Yin, and Jiliang Tang. 2017. A survey on Deep content-based music recommendation. In Advances in neural Information
dialogue systems: Recent advances and new frontiers. ACM SIGKDD Explorations Processing Systems. 2643–2651.
Newsletter 19, 2 (2017), 25–35. [33] Hao Wang, Naiyan Wang, and Dit-Yan Yeung. 2015. Collaborative deep learning
[5] Michaël Defferrard, Xavier Bresson, and Pierre Vandergheynst. 2016. Convolu- for recommender systems. In Proceedings of the 21th ACM SIGKDD International
tional neural networks on graphs with fast localized spectral filtering. In Advances Conference on Knowledge Discovery and Data Mining. ACM, 1235–1244.
in Neural Information Processing Systems. 3844–3852. [34] Suhang Wang, Jiliang Tang, Yilin Wang, and Huan Liu. 2018. Exploring
[6] Shuiguang Deng, Longtao Huang, Guandong Xu, Xindong Wu, and Zhaohui Wu. Hierarchical Structures for Recommender Systems. IEEE Transactions on
2017. On deep learning for trust-aware recommendations in social networks. Knowledge and Data Engineering 30, 6 (2018), 1022–1035.
IEEE transactions on neural networks and learning systems 28, 5 (2017), 1164–1177. [35] Xiang Wang, Xiangnan He, Liqiang Nie, and Tat-Seng Chua. 2017. Item silk road:
[7] Tyler Derr, Yao Ma, and Jiliang Tang. 2018. Signed Graph Convolutional Networks. Recommending items from information domains to social users. In Proceedings
In 2018 IEEE International Conference on Data Mining (ICDM). IEEE, 929–934. of the 40th International ACM SIGIR conference on Research and Development in
[8] Wenqi Fan, Qing Li, and Min Cheng. 2018. Deep Modeling of Social Relations Information Retrieval. ACM, 185–194.
for Recommendation. In AAAI. [36] Rongjing Xiang, Jennifer Neville, and Monica Rogati. 2010. Modeling relationship
[9] Aditya Grover and Jure Leskovec. 2016. node2vec: Scalable feature learning for strength in online social networks. In Proceedings of the 19th international
networks. In Proceedings of the 22nd ACM SIGKDD International Conference on conference on World wide web. ACM, 981–990.
Knowledge Discovery and Data Mining. ACM, 855–864. [37] Bo Yang, Yu Lei, Jiming Liu, and Wenjie Li. 2017. Social collaborative filtering by
[10] Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation trust. IEEE transactions on pattern analysis and machine intelligence 39, 8 (2017),
learning on large graphs. In Advances in Neural Information Processing Systems. 1633–1647.
1024–1034. [38] Zichao Yang, Diyi Yang, Chris Dyer, Xiaodong He, Alex Smola, and Eduard
[11] Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Hovy. 2016. Hierarchical attention networks for document classification. In
Chua. 2017. Neural Collaborative Filtering. In Proceedings of the 26th International Proceedings of the 2016 Conference of the North American Chapter of the Association
Conference on World Wide Web, WWW. 173–182. for Computational Linguistics: Human Language Technologies. 1480–1489.
[12] Geoffrey Hinton, Li Deng, Dong Yu, George E Dahl, Abdel-rahman Mohamed, [39] Rex Ying, Ruining He, Kaifeng Chen, Pong Eksombatchai, William L. Hamilton,
Navdeep Jaitly, Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara N and Jure Leskovec. 2018. Graph Convolutional Neural Networks for Web-Scale
Sainath, et al. 2012. Deep neural networks for acoustic modeling in speech Recommender Systems. In KDD ’18. ACM, 974–983.
recognition: The shared views of four research groups. IEEE Signal Processing [40] Lili Zhao, Zhongqi Lu, Sinno Jialin Pan, and Qiang Yang. 2016. Matrix
Magazine 29, 6 (2012), 82–97. Factorization+ for Movie Recommendation. In IJCAI. 3945–3951.
[41] Xiangyu Zhao, Liang Zhang, Zhuoye Ding, Long Xia, Jiliang Tang, and Dawei [42] Zhou Zhao, Qifan Yang, Hanqing Lu, Tim Weninger, Deng Cai, Xiaofei He, and
Yin. 2018. Recommendations with Negative Feedback via Pairwise Deep Yueting Zhuang. 2018. Social-aware movie recommendation via multimodal
Reinforcement Learning. In KDD’18. ACM, 1040–1048. network learning. IEEE Transactions on Multimedia 20, 2 (2018), 430–440.