1. Introduction
Recommender systems, which aim at filtering and suggesting items of potential interests to targeted users, have extensive applications and can bring up a huge amount of commercial benefits, especially when the information overload problem has been pervasive with the rapid development of the World Wide Web. Conventional collaborative filtering (CF) methods for recommendation suffer from data sparsity and cold start problems. To this end, social recommendation, which attempts to leverage the social networks of users to incorporate more useful information, has been proposed to alleviate these two problems (Golbeck, 2006; Jamali and Ester, 2009; Ma et al., 2011, 2008).
Prior Works and Limitations. Nowadays, there are many composite platforms which integrate item recommendation and social network services (SNS). Users on these platforms can click on or rate items as well as interact with their friends. A user’s preference for one item could be impacted by her friends, which motivates us to probe into such social effect to improve recommendation quality. Previous studies for social recommendation attempt to model social effects in various ways, such as by trust propagation (Golbeck, 2006; Jamali and Ester, 2009; Massa and Avesani, 2007a), regularization loss (Ma et al., 2011; Ying et al., 2010), matrix factorization (Ma et al., 2009; Gu et al., 2018; Yang et al., 2013), network embedding (Liu et al., 2018; Wen et al., 2018; Zhang et al., 2018)
, and deep neural network
(Fan et al., 2018; Wang et al., 2017).However, these studies share several common limitations. First, most studies assume that linked users all share similar preferences. This assumption cannot suit well contemporary SNS since there could be various types of online friends, such as close friends, casual friends, and event friends (Zhang et al., 2013). Therefore, directly equating friendship on an SNS with preference similarity may not be sound. Second, most works model friends’ influences statically under the forms of constant weights or fixed constraints. This assumption ignores the dynamic pattern of social effects. In fact, users could be influenced by a specific group of friends when it comes to specific items, which makes the social effect dynamic and dependent of specific contexts. Third, previous methods lack interpretability for social effects, i.e., they cannot explicitly indicate how a user’s preference for one item is influenced by their friends.
Motivations and Rationales. In the real world, people’s decisions could be influenced by various factors, and their behaviors are often the results of multifaceted causes. To this end, we probe into four different social effects in recommender systems, including twofold effects in user domain and the other symmetric twofold ones in item domain, as illustrated in Fig. 1. In user domain, friends could influence each other in two ways. For one thing, users tend to possess similar preferences to their friends, which is called as social homophily (Mcpherson et al., 2001) (Fig. 1.(a)). Social homophily often leads to intrinsic effect for user preference which stays unchanged and independent of external contexts. For another, a user’s friend who has purchased one item may recommend it to that user who then may be more likely to click such item, which is termed as social influence (Marsden and Friedkin, 1994) (Fig. 1.(b)). Actually, social influence contributes to behaviorlevel effect for user preference which may change dynamically with specific contexts. For example, a commercial analyst may be influenced by different communities of friends when it comes to an entertainment article or a commercial news. Therefore, we can distinguish two components that may impact a user’s preference: a static effect from social homophily and a dynamic effect from social influence^{2}^{2}2In some previous studies, there is little difference between social homophily and social influence. Differently, in this paper, we discriminate these two notions in order to better investigate the dynamic social effect., and we call corresponding user’s preference as static and dynamic preference, respectively (left part of Fig. 1).
Moreover, in item domain, there exist analogous ‘social effects’. For one thing, some related items tend to possess similar attractiveness when exposed to users. We term this as itemtoitem homophily, and call the component of an item attribute affected by such effect as static attribute (Fig. 1.(c)). For another, if one item is popular among a certain community, then other related items would become more likely to be clicked by users in this social group. Such phenomenon can be termed as itemtoitem influence, which depends on specific contexts (Fig. 1.(d)). For example, a businessman and a researcher both read one article talking about the development of AI, but the businessman would prefer another article about financial information of AI companies, while the researcher would tend to click on another article about cuttingedge breakthroughs in AI. We call the component of an item attribute affected by such effect as dynamic attribute. Finally, as shown by 1⃝4⃝ in Fig. 1, the twofold social effects in both user domain and item domain may jointly influence a user’s decision on one item. Such phenomenon is pervasive on Epinion, Twitter, Facebook, and Netflix, etc.
Methodologies and Results. To these ends, in this paper, we propose DANSER (Dual graph Attention Networks for modeling multifaceted Social Effects in Recommender systems), a novel architecture that extends graph attention networks (GAT) (Velickovic et al., 2018). DANSER is composed of two dual GATs: one dual GAT for users—including a GAT to capture social homophily and a GAT to capture social influence—and the other dual GAT for items. Specifically, in user domain, one GAT to capture social homophily attentively aggregates adjacent users’ embedding that reflects ones’ preferences, and outputs representations for static user preference. Another GAT to capture social influence aims at convolution on adjacent users’ contextaware preferences, which characterize relevance of users’ rated items to candidate item, and outputs representations for dynamic user preference. In item domain, two analogous GATs as a dual architecture are designed to model items’ static and dynamic attributes under effects of itemtoitem homophily and influence, respectively, among related items. Such dual mechanism possesses two advantages: i) GATs to capture homophily and influence could collaboratively model twofold social effects, taking advantage of both global and local views to investigate useritem interactions; ii) two GATs in item domain incorporate information from related items, which could further alleviate data sparsity problem.
Then, since user static (resp. dynamic) preference as well as item static (resp. dynamic) attribute could jointly affect a user’s decision on one item (shown by 1⃝4⃝ in Fig. 1) and the importance of these four interactions would vary for distinct useritem pairs, we propose a policybased strategy to dynamically weigh the four interactions. Specifically, we model the problem as a contextual multiarmed bandit (Bubeck and CesaBianchi, 2012), and treat the weighing strategy as a policy conditional on the context (the targeted useritem pair). Then our goal is to optimize a reward (w.r.t predicted loss). For model training, we use stochastic policy gradient to update our neural networkbased policy unit, and design a localgraph aware regularization technique to reduce computational cost for regularization. To verify our model, we conduct experiments on one benchmark datasets from Epinions and a commercial dataset from WeChat Top Story. The results show that DANSER outperforms stateoftheart models under both explicit feedback settings ( MAE improvement) and implicit feedback settings ( AUC improvement).
Our contributions can be summarized as follows:
i) General Aspects: We distinguish the social homophily and social influence notions in view of static and dynamic effects. Also, we extend the twofold social effects in user domain to item domain, and therefore investigate four social effects in recommender systems. These general aspects enable the model to capture more information and improve its capacity.
ii) Novel Methodologies: We propose DANSER with two dual GATs and a policybased fusion unit. We are one of the first to use GAT for social recommendation task, and the dual GATs can collaboratively model four social effects in both user and item domains. The policy unit, based on a contextual multiarmed bandit, dynamically weighs four interactions of social effects in two domains according to specific contexts.
iii) Multifaceted Experiments: We deploy DANSER on one benchmark dataset and a commercial dataset. The experiment results verify the superiority of DANSER over stateoftheart techniques, the effectiveness of the proposed components, as well as its good interpretability for modeling social effects.
2. Preliminary and Background
In this paper, we consider a useritem interaction matrix , where and are numbers of users and items respectively. For a recommender system with implicit feedback, consists of 0 and 1, where if user has clicked item , and otherwise. For a recommender system with explicit feedback, represents the rating value given by user on item (In Epinions, rating values range from 1 to 5 and user could give a high rating if she likes the item), and if user has not rated item . We use and to respectively denote the set of items rated by user and the set of users who have rated item . Moreover, we assume the existence of a trust or friendship network for users, which is a graph where is the set of users and is the set of edges that connect two users. We use to denote the set of nodes adjacent to in . Also, we consider interaction frequencies between user and user as edge features, denoted by , where each element in denotes the frequency of onetype interaction between and , and is the number of interaction types. The edge features reflect link strengths between two users.
Problem Formulation. The social recommendation problem can be defined as (Tang et al., 2013): given observed interaction records in and trust relationship
, we are to estimate unobserved interactions in
, i.e., the probability of a targeted user
clicking an unobserved candidate item (for implicit feedback) or the rating value user will score item (for explicit feedback).3. Methodologies
We first introduce the model framework. For each part, we start from the motivations and intuitions, and then introduce the technical details. Secondly, we present some training techniques to improve the model performance. Finally, we discuss two key units in our model for justification.
3.1. Model Framework
Fig. 2 provides an overview of DANSER. We will go into the details of each part in order.
Raw Input and Item Implicit Network. The model requires useritem interaction matrix and user social network as input. Social recommendation methods focus on leveraging the user social network to solve the data sparsity problem in recommender systems. Most existing methods treat items independently since there is no prior information that explicitly expresses the relationship between items. One way to calculate the similarity or relevance between two items is by the common users who clicked or rated them (Sarwar et al., 2001). For any item and item , we define their similarity coefficient as the number of users who clicked both items. These coefficients induce an equivalence relation over items as follows: item is related to item if with a fixed threshold. We define the item implicit network as the graph where is the set of items and is the set of edges that connects two related items.
Embedding Layer. The raw input of each user (resp. item) is a onehot vector with high dimension, and the embedding operation is to project each user to a lowdimensional representation. Inspired by (Koren, 2008) and (Cheng et al., 2018), each user can be represented by a userspecific embedding and the items rated by her. The former representation reflects user’s interests while the latter one (called itembased user embedding) captures implicit influence of her rated history on current decision. Such embedding could deal with falsenegative samples and improve model accuracy (Koren, 2008). Similarly, each item can be represented by an itemspecific embedding and the users who rated it (called userbased item embedding).
For the userspecific embedding, we define , where is the embedding dimension and denotes the embedding vector for user . For the itembased user embedding, we define another embedding lookup , where denotes the item latent factor. Then each item in will be mapped via and outputs , which will be used to compute the itembased user embedding, a dimensional vector representation, in the next layer. The user embeddings reflect a user’s inherent preference (the ‘inherent’ notion emphasizes that it contains no social information and only depends on the user oneself). Similarly, for the itemspecific embedding, we define , where denotes the embedding vector for item . For userbased item embedding, we define , where is user latent factor, and each user in will be mapped via and outputs , which will be used to compute userbased item embedding, a dimensional vector representation, in the next layer. Also, the item embeddings reflect item’s inherent attribute.
Dual GCN/GAT Layer. Graph convolution network (GCN) (Kipf and Welling, 2017)
conducts local convolutional operation over neighbor nodes in graph and outputs a new representation for each node, in order to encode the graph structure information as lowdimensional node representations. Such operation can be viewed as an extension of convolution neural network (CNN) from a grid structure to general graphs. GCN equally aggregates the neighbors’ embedding in each convolution and treats each neighbor nodes with equal importance. In contrast, Graph Attention Network (GAT)
(Velickovic et al., 2018) leverages attention mechanism to consider different weights from neighbor nodes, which enables the model to filter out noises and focus on important adjacent nodes.In order to exploit user social network and item implicit network , we propose dual graph attention networks, an extension of GCN/GAT as a means to nonlinearly aggregate the embeddings of the neighbor nodes in both networks, so that the social effects from friends and related items can propagate to adjacent nodes through the graphs. As discussed in Section , we can distinguish two types of social effects for users (resp. items) in social recommender systems: homophily effect and influence effect. They jointly affect user preferences and item attributes in different ways. DANSER uses two dual GATs to collaboratively learn different deep representations for user static/dynamic preference and item static/dynamic attribute.
I. GAT to capture social homophily (marked as blue in Fig. 2). Firstly, via user embedding, we have as representation of inherent user preference factor. Then the GCN/GAT operation could output a new representation, the user static preference factor , by
(1) 
where ,
are activation function, weight matrix, bias vector respectively, and
is the updated representation of users, which incorporates social information by using the attention weights obtained from . Its elements are defined as follows:where is weight matrix, , , and is weight vector. Here and denote the elementwise product and concatenation, respectively. Note that the above GAT weight remains unchanged given user and , which means factor matrix is fixed for users. This design conforms to the intuition that social homophily contributes to static effect for user’s preference, this is why we call it static preference factor.
II. GAT to capture social influence (marked as yellow in Fig. 2). In contrast to social homophily, social influence effect is often contextaware, and the model needs to output different attention weights for friends w.r.t different candidate items. Through the embedding layer, we have as the embedding of the items clicked by user . Then we let each item clicked by user interact with candidate item ,
This product operation can help to focus on the candidate item and model dynamic social influence under a specific context. We define the itembased user embedding , which depends on candidate item
with max pooling to select the most dominating features for
dimensions:where , , and are the th feature of , , respectively. The max pooling operation can help to focus on the most important value and alleviate noises in users’ clicked history. The itembased embedding includes two information: i) user ’s contextaware preference (w.r.t candidate item ), and ii) inherent representation for (independent of social information). In order to define the user dynamic preference factor , we proceed to incorporate the social information from friend users,
for . Note that the above attention weight depends on the user’s history of rated items as well as specific candidate item , which indicates that factor matrix would change dynamically with different contexts. This design conforms to the intuition that social influence contributes to contextaware effect for a user’s preference, so we term it as dynamic preference factor.
III. GAT to capture itemtoitem homophily (marked as green in Fig. 2). Similarly, we use the item embedding as representation of inherent item attribute factor, and then leverage GAT to incorporate social information,
where , and is a weight vector. The GAT weight remains unchanged given item and . Correspondingly, itemtoitem homophily contributes to static effect for item’s attribute, and we call it static attribute factor.
IV. GAT to capture itemtoitem influence (marked as red in Fig. 2). We now model itemtoitem influence, which is contextaware and depends on the specific targeted user. Hence, our model needs to output different attention weights for distinct related items w.r.t different targeted users. Similarly to social influence, the userbased item representation for a given targeted user can be defined as follows:
where , , and are the th feature of , , respectively. Then the representation for item dynamic attribute can be computed as follows:
Pairwise Neural Interaction Layer. Since user’s decision on one item often depends on both user preference and item attribute, we can let two preference factors and two attribute factors pairwisely interact with each other. Then we borrow the idea from (He et al., 2017), feeding the four results into different neural networks indexed by :
Here we employ a tower structure for each network, where higher layers have smaller number of neurons
(He et al., 2017).PolicyBased Fusion Layer. We further fuse the four interacted features into a synthetic one. Note that homophily effect and influence effect could jointly affect user preference and item attribute, but for distinct users and items, the importance of the twofold social effects could be different. To this end, we propose a new policybased fusion strategy to dynamically allocate weights to the four interacted features according to specific useritem pairs.
We model the weight allocation as a contextual multiarmed bandit problem(Bubeck and CesaBianchi, 2012),where an action, denoted by , indicates which feature to choose, a context is a useritem pair and the reward after playing an action represents a recommendation loss. In this problem, a stochastic policy can be written as the conditional probability , which could characterize the selecting probability or importance weights for different combinations of social effects given a specific pair .
To make the problem solvable, we approximate by a neural network (called policy network):
Then the synthetic representation can be expressed as
We call the above strategy singleheaded policybased fusion. Inspired by (Vaswani et al., 2017), we can extend it to a multiheaded version. We harness different independent policy networks and the final weights are given by the averaged weights given by each policy net. The training of policy networks is by stochastic policy gradient, which will be discussed in Section 3.2.
Output Layer. Then probability of user clicking item can be predicted by
If a clicking probability is required (for implicit feedback), can be a fullyconnected layer with a sigmoid activation function. If a rating value is needed (for explicit feedback), can be a fullyconnected layer without activation function.
Loss Function
. The loss function measures the discrepancy between predicted value and groundtruth value. For implicit feedback, the most widely adopted loss function is the crossentropy defined as
For explicit feedback, we adopt the mean square loss
Here, we concentrate on pointwise loss. Also, other pairwise loss functions can be used and we leave them for future investigation.
3.2. Training
In this subsection, we introduce some training techniques to improve the model performance.
MiniBatch Training. We also leverage minibatch training to calculate the gradient. For each iteration, we consider useritem pairs. For each pair , the neighbor nodes , in two networks as well as the rated history , , and ,
will be together input into the model to compute the gradient. Once an epoch is finished, the batch partition will be randomly reset to incorporate enough noise during optimization. In each minibatch, since numbers of friends are different for users, we need to pad zeros and input a matrix with dimension of max friend number in one batch. We observe that number of friends tends to obey a longtail distribution, so the padding would lead to both high space and time complexity. Therefore, we use a sample technique: 1) for friend number more than
, we uniformly sample friends as input data; 2) for friend number less than , we pad zeros to obtain a dimensional vector. We also experiment other sample methods like random walk with restart (RWR) (Tong et al., 2006) on a subgraph near central node, but we find that uniformly sampling over adjacent nodes can achieve competitive performance. This observation could be justified by one argument in sociology research that influence from users always propagates to adjacent users in social networks (Goyal et al., 2010). Thus we adopt the simpler method to reduce time complexity. In experiment part, we will investigate the impact of sample size on the performance.LocalGraph Aware Regularization. The other issue is regularization to avoid overfitting problem. In our model, we adopt L1 regularization to constrain the embedding parameters to sparse forms. The regularization loss can be expressed as
However, the above loss introduces high computational complexity, since for each batch training we need to calculate the gradient of regularization over all users and items. Such operation can not generalize to largescale data set. A recent work (Zhou et al., 2018b) proposes a minibatch aware regularization trick to solve this problem. In each batch training, only the embedding parameters for useritem pair in the batch will be used to calculate the regularization loss. Here we extend this idea to graph structure and call it as localgraph aware regularization. In each batch training, we constrain the embedding parameters of users, items as well as their neighbor nodes in social network and item implicit network. The new regularization loss can be written as
In fact, the localgraph aware regularization is to split the regularization loss into small parts for each minibatch and guarantees that once an epoch is finished, each user and each item can be regularized equally. To sum up, the final loss function is
(2) 
where is a tradeoff parameter between accuracy and complexity.
Besides, for parameters in neural networks (weight matrix and bias vectors), we adopt dropout strategy to replace the traditional regularization. Our experiments will further discuss the influence of different dropout probabilities .
Policy Gradient. In training stage, we train the policy networks in a stochastic way. Assume to be the active probability given by
th policy network, and random variable
obeys a multinomial distribution. For each policy network, we draw , and feed into output layer, which gives prediction and returns loss. We leverage the loss to train the policy network. This design can be formulated as an actorcritic terminology: i) given the environment (user and item ), the th agent (policy net) needs to execute an action according to policy (); ii) every action would generate a reward (recommendation loss) that could lead the training of policy network. We define the reward as subtraction of loss generated by selected interacted feature, i.e., (where denotes the loss when we input into the output layer). We aim at maximizing the expected reward . The policy gradient method REINFORCE (Williams, 1992) can be used to update parameters in one policy network (denoted as ), and the gradient can be derived as follows:(3) 
Training Algorithm. The training of DANSER updates policy networks and feedforward networks (including embedding, dual GAT, pairwise neural interaction and output layers) iteratively, which is presented in Alg. 1. First, we update feedforward networks for steps (one minibatch for one step) by optimizing (2). Then we train policy networks independently by conducting corresponding policy gradient (3).
Complexity Analysis. Since complexity and scalability are two important concerns for graph algorithm, we investigate complexity of DANSER for model training and inference respectively. For training, we need to update embedding of all friend users and related items for useritem pairs in one batch, which requires time complexity , where . We can control to a limited value by truncating recent rated items for users. The friend number is controlled by our sample method. Also, for inference, we need to compute GAT representations. Hence, DANSER can scale linearly w.r.t number of useritem interactions.
3.3. Discussions
In this subsection, we discuss on effectiveness of our key units, dual GAT layer and policybased fusion layer, in order to shed more insights on DANSER.
Justification of Dual GATs. During the training process, the GAT to capture the homophily effect would give larger weights to friends whose preferences for items are globally similar to the targeted user (resp. related items whose attractiveness to users are globally similar to candidate item), so the static preference (resp. attribute) factor captures static effect of friend users (resp. related items) from a global view. Differently, the weights given by the GAT to capture the influence effect depend on the targeted user (resp. item), so the corresponding dynamic factor captures contextaware effects from a local view. In each domain, a dual GAT collaboratively models the twofold effects and takes both global and local views to conduct a more accurate investigations into distinct users and items.
To put it more symbolically, we illustrate its effectiveness by considering the following two cases for the recommendation of candidate item to targeted user : item is similar to item and is similar to (see left and right columns of Fig. 3). Here, we assume that user has only clicked on and (and no other items), and users to are ’s friends. Previous works for social recommendation consider static effect of friends in an average manner. As illustrated in Fig. 3.(a), for friend who clicked on both items, the attention weight would be large, and the opposite for friend . For and , their weights are medium since they are both partially similar to . Therefore static weights cannot discriminate between the influence of user and when it comes to item even if it is known that it is similar to or . In contrast, DANSER contains two different attention weights, where one captures userspecific static effect (for homophily) and the other can dynamically change under different contexts (i.e., for influence). When candidate item is similar to (resp. ), the social influence GAT would focus more on and (resp. and ). Then the resultant force of the two GATs would make pay more attention to when is similar to and to in the other case. The global and local views of DANSER work collaboratively for selecting the ‘right’ neighbors that are similar to the targeted user under a specific context, which improves model expressiveness. In Section 4.5, we will verify the above argument via a realcase study.
Justification of PolicyBased Fusion. The policybased fusion layer acts like a stochastic network while training, and the REINFORCE algorithm can help to explore the optimal weights distribution over four interacted features. Since the interacted feature with larger active probability is more likely to be selected, it would be more frequent for the interaction neural layer that gives the ‘best’ feature to be trained. Such mechanism brings up two major benefits. Firstly, the policy networks can filter ‘right’ useritem samples for training of corresponding interaction neural layer as well as previous GCN/GAT and embedding layers. Second, the active probability introduces random pattern to the training process, which can help to escape from local optimums. Our ablation study in experiment part also validate the superiority of the policybased fusion compared with other straightforward strategies.
4. Experiments
To comprehensively evaluate proposed model DANSER^{3}^{3}3The codes are released at https://github.com/echo740/DANSERWWW19., we conduct experiments to answer the following research questions:
RQ1 How does DANSER perform compared with stateoftheart models for recommendation and social recommendation?
RQ2 Are the key components in DANSER, such as dual GATs and policybased fusion, necessary for improving performance?
RQ3 How do hyperparameters in DANSER impact recommendation performance?
RQ4 How can DANSER interpret the four different social effects from realworld data?
4.1. Experiment Setup
4.1.1. Data sets
We apply our model to one public benchmark datasets for social recommendation Epinions and a commercial dataset WeChat Top Story. We provide some statistics of these two datasets in Table 1. The basic information about the two datasets is summarized as follows:
i) Epinions: Epinions is a consumer review websites, where users can rate some items and add other users in their trust lists. Epinions dataset (Massa and Avesani, 2007b) contains two parts of information: useritem interaction pairs, where items are rated from 1 to 5 (explicit feedback), as well as the directed trust relationships between users. The dataset has been widely used as benchmarks for social recommendation. Following other works (Guo et al., 2015; Yang et al., 2013; Wang et al., 2017; Liu et al., 2018), we randomly choose useritem interactions as training set and the remaining as test set for each dataset.
ii) WeChat: We also deploy our model on a realworld article recommender system, WeChat Top Story. This dataset is constructed by userarticle clicking records on this platform. Different from Epinions, this dataset only contains implicit feedback, i.e., we only know whether a user clicked an article or not. Besides, the friend relationships are bidirectional. We chronologically order the userarticle clicking records, and use the first records to train our model and the remaining records to evaluate the prediction performance. Since the positive samples and negative samples are fairly unbalanced in this dataset, we uniformly sample the negative ones such that the numbers of positive and negative samples are the same for one user.
Dataset  #users  #items  #interactions  #relationships 

Epinions  49,290  139,738  664,824  487,181 
200,000  100,000  4,000,000  2,000,000 
4.1.2. Implementation Details
We use Tensorflow to implement our model and deploy it on GTX 1080 GPU with 8G memory. The hyperparameter settings for
Epinions are as follows: batch size , embedding dimension , dropout ratio , regularization coefficient , sample size , truncation length , multihead number , policy gradient period . The learning rates ,. The Leaky ReLu slope is 0.2. In pairwise interaction neural layer, we adopt threelayer neural networks and neuron numbers are
.4.1.3. Competitors
We choose several comparative methods, including some stateoftheart models for recommendation and social recommendation, to evaluate the performance of our DANSER. The optimal hyperparameter settings for each method are determined either by our experiments or suggested by previous works.
The recommendation models include SVD++(Koren, 2008) and DELF(Cheng et al., 2018). These two methods only leverage useritem interactions as model input. They can be used to evaluate the effectiveness of other models that use social information.

SVD++ (Koren, 2008) is a basic modelbased recommendation method which use both userspecific and itembased user embedding to represent user’s preferences.

DELF (Cheng et al., 2018) is a stateoftheart CF method which proposes dual embedding for users and items, and adopts deep neural networks to capture complex nonlinear information.
The social recommendation models include TrustPro, TrustMF, TrustSVD, NSCR, SREPS. These methods all leverage both useritem interactions and user friendships as input information.

TrustPro (Ye et al., 2011) is a trust propagation method, using rating of friends to deduce rating of targeted user.

TrustMF (Yang et al., 2013), as one matrix factorization method, optimizes user embedding to retrieve the trust matrix.

TrustSVD (Guo et al., 2015) is another matrix factorizationbased method, incorporating friends’ embedding vectors into targeted user’s predicted rating.

NSCR (Wang et al., 2017), as a strong baseline for social recommendation, adopts deep neural networks to learn latent representations of users and items, and leverages graph regularization to constrain the embedding of adjacent users to be similar.

SREPS (Liu et al., 2018) is another strong baseline, using a network embedding approach to encode social network.
We also construct several variants of DANSER as ablation study. To evaluate the necessity of dual GAT layer, we simplify our model as DualEMB, DualGCN, userGAT, and itemGAT. DualEMB removes the convolution over adjacent nodes, while DualGCN replace GAT by GCN. userGAT removes two GATs in item domain, and itemGAT removes two GATs in user domain. Moreover, to evaluate the necessity of policybased fusion layer, we consider the following variants. DANSERm and DANSERa leverage max and average pooling to fuse four interacted features, respectively. DANSERc concatenates the four interacted features as the input of output layer. DANSERw removes the policy net, set four sharing fusion weights, and update the weights together with training of feedforward networks.
4.1.4. Evaluation Protocol
We adopt different metrics to evaluate recommendation performance. Since two datasets possess different feedbacks, we consider different metrics for them. For Epinions with explicit feedback, we use MAE and RMSE
, which are widely adopted by other works, as evaluation metrics. For
WeChat with implicit feedback, we use Precision@k (short as P@k) and AUC, two universally acknowledged metrics for 01 classification, to evaluate the performance. We repeated each experiment ten times and report the average performance.4.2. Comparative Results: RQ1
We report experiment results of DANSER and other comparative methods in Table 2. As we can see, our model DANSER outperforms other comparative methods and achieves great improvements, i.e., MAE impv. for Epinions and AUC impv. for WeChat. The results verify the superiority of proposed model for social recommendation. Also, there are some findings in these comparative experiments. First, matrix factorization methods like TrustMF and TrustSVD can achieve good accuracy in Epinions with explicit feedback, but perform poorly in WeChat with implicit feedback. Second, NSCR performs quite better than other linear methods, which indicates that neural network plays an important role in capturing complex nonlinear information. Third, even though DELF has not leveraged any social information, it outperforms some social recommendation methods like TrustSVD and NSCR. This is possibly due to the facts: i) the dual embedding and neural interaction layer in DELF can better exploit information in useritem interactions; ii) other social recommendation methods like TrustSVD and NSCR do not make good use of social information. Indeed, TrustSVD assumes equal importance of each friend user while NSCR uses graph regularization that constrains representations for adjacent users in a static way. These settings resort to a prior assumption that friend users’ preferences are all similar, which may not be true for contemporary SNS (also discussed in Section 1).
Epinions  
MAE  RMSE  P@10  AUC  
SVD++ (Koren, 2008)  0.8321  1.0772  0.0653  0.7304 
DELF (Cheng et al., 2018)  0.8115  1.0561  0.0752  0.7818 
TrustPro (Ye et al., 2011)  0.9130  1.1124  0.0561  0.6482 
TrustMF (Yang et al., 2013)  0.8214  1.0715  0.0625  0.7005 
TrustSVD (Guo et al., 2015)  0.8144  1.0492  0.0664  0.7325 
NSCR (Wang et al., 2017)  0.8044  1.0425  0.0736  0.7727 
SREPS (Liu et al., 2018)  0.8014  1.0393  0.0725  0.7745 
DANSER  0.7781  1.0268  0.0823  0.8165 
Impv.^{1}  2.87%  1.25%  9.33%  4.48% 

The improvement compares DANSER with the best competitor (underlined).
4.3. Ablation Study: RQ2
In order to verify the effectiveness of some components in our model, we do some ablation studies and the results are shown in Table 3. We first compare DANSER with DualEMB and DualGCN. As we can see, GCNs can capture enough social information hidden in graph structure and improve recommendation accuracy (about impv. for MAE over DualEMB), while GATs can further improve performance by differentiate importance weights for friend users (about extra impv. for MAE over DualGCN). We secondly compare DANSER with userGAT and itemGAT. The results show that incorporation of user and item networks can both bring significant performance improvement, and the benefits brought by user social network are greater.
We also evaluate the necessity of policybased fusion layer. The results show that our policybased fusion and policy gradient method can improve the accuracy a lot. Also, for Epinions, the improvement is quite significant. One possible reason is that the social effects in Epinions are more heterogeneous than those in WeChat, so to dynamically weigh different social effects are more effective for Epinions. More importantly, since DualEMB can be seen as the extension of DELF (Cheng et al., 2018), where we replace the concatenation of four embedding by our policy unit. Comparing DualEMB in Table 3 with DELF in Table 2, we find that the policy unit do help to improve its performance and is superior to straightforward concatenation.
In Section 1, we emphasized that social recommendation is proposed to solve the data sparsity and coldstart problem in recommender systems. Here we would like to investigate the performance of DANSER for coldstart users. Fig. 4 shows the results on Epinions and WeChat for users with different number of clicked items and friends. We present the median, the and percentile MAE/AUC. As we can see, with number of clicked items increasing, the gap between upper and lower bounds becomes narrow, which indicates that performance becomes better. Generally, DANSER can achieve better overall performance as well as lower variation of MAE/AUC. For users with different friend numbers, as depicted in the figure, when friend number goes up, the variation trends for three methods are different. For DualEMB, the performance improves a little. For DualGCN, there is an obvious performance decline, while performance of DANSER becomes better. This is possibly because for users with large number of friends, the GCN operation may incorporate more misleading information, and instead, GAT can attentively filter the noise and concentrate more on important neighbor nodes.
Epinions  
MAE  RMSE  P@10  AUC  
DualEMB  .7920(1.7%)^{1}  1.0363(0.7%)  .0794(3.6%)  .7992(2.2%) 
DualGCN  .7840(0.7%)  1.0335(0.4%)  .0814(1.1%)  .8102(0.8%) 
userGAT  .7858(0.9%)  1.0364(0.7%)  .0813(1.2%)  .8136(0.4%) 
itemGAT  .7919(1.7%)  1.0335(0.4%)  .0813 (1.2%)  .8138(0.3%) 
DANSERw  .8191(4.9%)  1.0659(3.4%)  .0820(0.4%)  .8151(0.2%) 
DANSERm  .8211(5.2%)  1.0681(3.6%)  .0815(1.0%)  .8144(0.3%) 
DANSERa  .8232(5.4%)  1.0710(3.9%)  .0814 (1.1%)  .8140(0.3%) 
DANSERc  .8091(3.8%)  1.0659(3.4%)  .0809(1.7%)  .8118(0.6%) 
DANSER  0.7787  1.0292  0.0823  0.8165 

The ratio indicates the impv. comparing DANSER with corresponding variant.
4.4. Parameter Sensitivity: RQ3
We study the performance variation for our model w.r.t some hyperparameters including regularization parameter , dropout rate , embedding dimension , sample size . (The other hyperparameters have little impact on model performance, so we skip discussions on them for space limit.) The results are shown in Fig. 5 and enlighten us that. i) The regularization parameter plays an important role, and if it is too large, the model cannot focus on minimizing the recommendation loss. ii) If the dropout ratio is too large, the random pattern would hinder the training. If is too small, it would lose the regularization efficacy. iii) A proper embedding size is needed. If it is too small, the model lacks expressiveness, while if it is too large, the representation vector would become so sparse, which leads to performance decline. iv) With the sample size increasing, the accuracy improves but the computational cost increases as well. Hence an appropriate sample size can keep a good balance between complexity and accuracy.
4.5. Case Study: RQ4
We conduct a case study of four useritem pairs and show the weights given by four GATs by node size in Fig. 6. As we can see, the weights of two homophily GATs stay unchanged for the same user or item, while the weights of two influence GATs vary for different useritem pairs. Moreover, influence GATs incline to allocate larger weights to nodes with the same labels, which echoes the argument in Section 3.3 that the influence GATs tend to concentrate more on the similar nodes under a specific context. Also, the policy weights play a part in selecting the most dominant interacted features from four GATs representations.
More importantly, we can decompose the model interpretability for social effects in recommender systems by a twolevel illustration: i) for a specific useritem pair , the policy weights first indicate the dominant combinations of social effects in user domain and item domain, ii) the corresponding GATs for four social effects further give out the influential degree of each friend given item as well as the influential degree of each related item given user .
5. Related Works
Social Recommendation. In online social networks, friends tend to have a large impact on users’ decisions and behaviors (Guille and Hacid, 2012; Sharara et al., 2011). There are plenty of studies for social effects, such as homophily(Mcpherson et al., 2001) and influence(Marsden and Friedkin, 1994), over graph structures in social network analysis. For recommender systems, (Kautz et al., 1997) is one of the pioneer who takes advantage of social networks to alleviate the data sparsity problem for item recommendation. Then many following studies begin to model social effects in recommender systems from various perspectives. (Golbeck, 2006; Jamali and Ester, 2009; Massa and Avesani, 2007a)
propose trust propagation methods by leveraging rating of friends to deduce rating of targeted user. This kind of approach is quite heuristic and is called
similaritybased method. Then a more modern approach, modelbased method, is proposed by other studies. (Ma et al., 2011; Ying et al., 2010) use regularization to constrain the neighbor users’ embedding vectors to be similar and (Ma et al., 2009; Gu et al., 2018; Yang et al., 2013) adopt matrix factorization to incorporate the adjacent matrix into user embedding. Since the above methods can only capture linear information, some recent studies like (Fan et al., 2018; Wang et al., 2017) and (Liu et al., 2018; Wen et al., 2018; Zhang et al., 2018) leverage deep neural networks and network embedding approaches, respectively, to learn a more complex representation for graph structures. However, comparing with dual GATs used in this paper, the common limitations of existing studies lie in: i) they assume neighbors’ influences to be equally important or statically constrained, ii) they ignore the social effects from related items, iii) modeling of social effects lacks interpretability.Graph Convolution/Attention Network. GCN and GAT, as two powerful techniques to encode a complex graph into lowdimensional representations, are extensively applied into various problems involved with graph data since their proposals (Kipf and Welling, 2017; Defferrard et al., 2016; Velickovic et al., 2018). Using GCN and GAT to solve semisupervised classification problem in graph could achieve statetheofart performance (GAT can be seen as an extension of GCN and provide better performance according to (Velickovic et al., 2018)), and its good scalability enables it to tackle largescale dataset. Existing studies leverage GAT to tackle social influence analysis(Qiu et al., 2018), graph node classification(Gong and Cheng, 2018), conversation generation(Zhou et al., 2018a), relevence matching(Zhang et al., 2018). For recommendation, some recent studies like (Ying et al., 2018a) adopt GCN to convolve on useritem network (a bipartite graph) to obtain better representations for items, and there are also some works using GCN to capture social information for recommendation (Ying et al., 2018b)(Wu et al., 2018). We are the first to use GAT for social recommendation task, and our new architecture, dual GATs, can capture social information in both user and item networks.
Dual Mechanism. There are many dual phenomenons in reallife which inspire several dual structures in model design. (Xia et al., 2018) proposes a dualmodel paradigm to merge the training of two dual tasks. (Zhuang and Ma, 2018) extends GCN to dual structures, combining global and local consistencies in graph. (Cheng et al., 2018) proposes DELF conducting dual embedding for users and items in recommender systems. Compared to (Cheng et al., 2018), our model possesses several key differences: i) for the itembased user embedding (resp. the userbased item embedding), DANSER considers dynamic embedding under specific candidate items (resp. targeted users), while DELF treats embedding statically given a user (resp. item); ii) in dual GCN/GAT layer, DANSER harnesses two dual GATs to capture social information in user social network and item implicit network, while DELF considers no usertouser and itemtoitem mutual influences; iii) in the policybased fusion layer, DANSER dynamically weighs the importances of the four GAT representations, while DELF straightforwardly concatenates four embedding vectors.
6. Conclusion
In this paper, we proposed DANSER, which includes two dual graph attention networks, to learn deep representations for social effects in recommender systems. Our method is equipped with good expressiveness because: i) it can collaboratively model homophily and influence effects from both friend users and related items; ii) it can investigate heterogeneous importance distributions among different interactions of the two twofold social effects for distinct useritem pairs. Our comparative experiments and ablation studies on a benchmark dataset and a commercial dataset showed that DANSER can learn effective representations of multifaceted social effects and leverage this information to significantly improve recommendation accuracy.
As extension of our work, we believe that the construction of item implicit network could leverage extra attribute features or distill more highlevel connections through knowledge graphs if the information is available. More generally, one could extend DANSER to other information retrieval tasks, like question answering, where DANSER could model ’social effects’ from related questions and similar answers given the corresponding networks of them.
References
 (1)

Bubeck and
CesaBianchi (2012)
Sébastien Bubeck and
Nicolò CesaBianchi.
2012.
Regret Analysis of Stochastic and Nonstochastic
Multiarmed Bandit Problems.
Foundations and Trends in Machine Learning
5, 1 (2012), 1–122.  Cheng et al. (2018) Weiyu Cheng, Yanyan Shen, Yanmin Zhu, and Linpeng Huang. 2018. DELF: A DualEmbedding based Deep Latent Factor Model for Recommendation. In IJCAI. 3329–3335.
 Defferrard et al. (2016) Michaël Defferrard, Xavier Bresson, and Pierre Vandergheynst. 2016. Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering. In NIPS. 3837–3845.
 Fan et al. (2018) Wenqi Fan, Qing Li, and Min Cheng. 2018. Deep Modeling of Social Relations for Recommendation. In AAAI. 8075–8076.
 Golbeck (2006) Jennifer Golbeck. 2006. Generating Predictive Movie Recommendations from Trust in Social Networks. In iTrust. 93–104.
 Gong and Cheng (2018) Liyu Gong and Qiang Cheng. 2018. Adaptive Edge Features Guided Graph Attention Networks. CoRR abs/1809.02709 (2018).
 Goyal et al. (2010) Amit Goyal, Francesco Bonchi, and Laks V. S. Lakshmanan. 2010. Learning influence probabilities in social networks. In WSDM. 241–250.
 Gu et al. (2018) Yulong Gu, Jiaxing Song, Weidong Liu, Lixin Zou, and Yuan Yao. 2018. CAMF: Context Aware Matrix Factorization for Social Recommendation. Web Intelligence 16, 1 (2018), 53–71.
 Guille and Hacid (2012) Adrien Guille and Hakim Hacid. 2012. A predictive model for the temporal dynamics of information diffusion in online social networks. In WWW. 1145–1152.
 Guo et al. (2015) Guibing Guo, Jie Zhang, and Neil YorkeSmith. 2015. TrustSVD: Collaborative Filtering with Both the Explicit and Implicit Influence of User Trust and of Item Ratings. In AAAI. 123–129.
 He et al. (2017) Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and TatSeng Chua. 2017. Neural Collaborative Filtering. In WWW. 173–182.
 Jamali and Ester (2009) Mohsen Jamali and Martin Ester. 2009. TrustWalker: a Random Walk Model for Combining Trustbased and Itembased Recommendation. In KDD. 397–406.
 Kautz et al. (1997) Henry A. Kautz, Bart Selman, and Mehul A. Shah. 1997. Referral Web: Combining Social Networks and Collaborative Filtering. Communication of ACM 40, 3 (1997), 63–65.
 Kipf and Welling (2017) Thomas N. Kipf and Max Welling. 2017. SemiSupervised Classification with Graph Convolutional Networks. In ICLR.
 Koren (2008) Yehuda Koren. 2008. Factorization Meets the Neighborhood: a Multifaceted Collaborative Filtering Model. In KDD. 426–434.
 Liu et al. (2018) ChunYi Liu, Chuan Zhou, Jia Wu, Yue Hu, and Li Guo. 2018. Social Recommendation with an Essential Preference Space. In AAAI. 346–353.
 Ma et al. (2009) Hao Ma, Irwin King, and Michael R. Lyu. 2009. Learning to Recommend with Social Trust Ensemble. In SIGIR. 203–210.
 Ma et al. (2008) Hao Ma, Haixuan Yang, Michael R. Lyu, and Irwin King. 2008. SoRec: Social Recommendation using Probabilistic Matrix Factorization. In CIKM. 931–940.
 Ma et al. (2011) Hao Ma, Dengyong Zhou, Chao Liu, Michael R. Lyu, and Irwin King. 2011. Recommender Systems with Social Regularization in Social Networks. In WSDM. 287–296.
 Marsden and Friedkin (1994) Peter V Marsden and Noah E Friedkin. 1994. Network studies of social influence. Sociological Methods & Research 22, 1 (1994), 127–151.
 Massa and Avesani (2007a) Paolo Massa and Paolo Avesani. 2007a. TrustAware Recommender Systems. In KDD. 17–24.
 Massa and Avesani (2007b) Paolo Massa and Paolo Avesani. 2007b. Trustaware recommender systems. In RecSys. 17–24.
 Mcpherson et al. (2001) Miller Mcpherson, Lynn SmithLovin, and James M. Cook. 2001. Birds of a Feather: Homophily in Social Networks. Annual Review of Sociology 27, 1 (2001), 415–444.

Qiu
et al. (2018)
Jiezhong Qiu, Jian Tang,
Hao Ma, Yuxiao Dong,
Kuansan Wang, and Jie Tang.
2018.
DeepInf: Social Influence Prediction with Deep Learning. In
KDD. 2110–2119.  Sarwar et al. (2001) Badrul Munir Sarwar, George Karypis, Joseph A. Konstan, and John Riedl. 2001. Itembased collaborative filtering recommendation algorithms. In WWW. 285–295.
 Sharara et al. (2011) Hossam Sharara, William Rand, and Lise Getoor. 2011. Differential Adaptive Diffusion: Understanding Diversity and Learning whom to Trust in Viral Marketing. In ICWSM.
 Tang et al. (2013) Jiliang Tang, Xia Hu, and Huan Liu. 2013. Social recommendation: a review. Social Netw. Analys. Mining 3, 4 (2013), 1113–1133.
 Tong et al. (2006) Hanghang Tong, Christos Faloutsos, and JiaYu Pan. 2006. Fast Random Walk with Restart and Its Applications. In ICDM. 613–622.
 Vaswani et al. (2017) Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is All you Need. In NIPS. 6000–6010.
 Velickovic et al. (2018) Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. 2018. Graph Attention Networks. In ICLR.
 Wang et al. (2017) Xiang Wang, Xiangnan He, Liqiang Nie, and TatSeng Chua. 2017. Item Silk Road: Recommending Items from Information Domains to Social Users. In SIGIR. 185–194.
 Wen et al. (2018) Yufei Wen, Lei Guo, Zhumin Chen, and Jun Ma. 2018. Network Embedding Based Recommendation Method in Social Networks. In WWW. 11–12.

Williams (1992)
Ronald J. Williams.
1992.
Simple Statistical GradientFollowing Algorithms for Connectionist Reinforcement Learning.
Machine Learning 8 (1992), 229–256.  Wu et al. (2018) Le Wu, Peijie Sun, Richang Hong, Yanjie Fu, Xiting Wang, and Meng Wang. 2018. SocialGCN: An Efficient Graph Convolutional Network based Model for Social Recommendation. CoRR abs/1811.02815 (2018).
 Xia et al. (2018) Yingce Xia, Xu Tan, Fei Tian, Tao Qin, Nenghai Yu, and TieYan Liu. 2018. ModelLevel Dual Learning. In ICML. 5379–5388.
 Yang et al. (2013) Bo Yang, Yu Lei, Dayou Liu, and Jiming Liu. 2013. Social Collaborative Filtering by Trust. In IJCAI. 2747–2753.
 Ye et al. (2011) Mao Ye, Peifeng Yin, WangChien Lee, and Dik Lun Lee. 2011. Exploiting geographical influence for collaborative pointofinterest recommendation. In SIGIR. 325–334.
 Ying et al. (2010) Rex Ying, Ruining He, Kaifeng Chen, Pong Eksombatchai, William L. Hamilton, and Jure Leskovec. 2010. A Matrix Factorization Technique with Trust Propagation for Recommendation in Social Networks. In RecSys. 135–142.
 Ying et al. (2018a) Rex Ying, Ruining He, Kaifeng Chen, Pong Eksombatchai, William L. Hamilton, and Jure Leskovec. 2018a. Graph Convolutional Neural Networks for WebScale Recommender Systems. In KDD. 974–983.
 Ying et al. (2018b) Rex Ying, Ruining He, Kaifeng Chen, Pong Eksombatchai, William L. Hamilton, and Jure Leskovec. 2018b. Graph Convolutional Neural Networks for WebScale Recommender Systems. In KDD. 974–983.
 Zhang et al. (2018) Ting Zhang, Bang Liu, Di Niu, Kunfeng Lai, and Yu Xu. 2018. Multiresolution Graph Attention Networks for Relevance Matching. In CIKM. 933–942.
 Zhang et al. (2013) Xingang Zhang, Qijie Gao, Christopher S.G. Khoo, and Amos Wu. 2013. Categories of Friends on Social Networking Sites: An Exploratory Study. In ALIEP. 244–259.
 Zhou et al. (2018b) Guorui Zhou, Xiaoqiang Zhu, Chengru Song, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai. 2018b. Deep Interest Network for ClickThrough Rate Prediction. In KDD. 1059–1068.
 Zhou et al. (2018a) Hao Zhou, Tom Young, Minlie Huang, Haizhou Zhao, Jingfang Xu, and Xiaoyan Zhu. 2018a. Commonsense Knowledge Aware Conversation Generation with Graph Attention. In IJCAI. 4623–4629.
 Zhuang and Ma (2018) Chenyi Zhuang and Qiang Ma. 2018. Dual Graph Convolutional Networks for GraphBased SemiSupervised Classification. In WWW. 499–508.
Comments
There are no comments yet.