Relational Metric Learning With Graph Attention Networks For Social Recommendation

In the era of mobile Internet,people are often troubled by the information overload problem caused by the explosively growing online contents.As a good tool to alleviate such problem,recommender system(RS)could mine user's personalized preferences from the user-item history interactions and then filter out the contents they dislike.As we know,traditional recommender models typically face issues like data sparsity and coldstart.In recent years,more and more researches focus on modeling side information like social network to extract rich user features so as to improve performance of current models.However,through relevant researches,we find that existing social recommenders typically have following limitations,including unreasonable assumption about user's interest similarities in social domain which may introduce too much noise,learning a unified representation for each user when modeling its features and unable to capture high-order neighborhood information encoded in both item and social domains.To handle these problems,we conduct research from the relation modeling persperctive and aim at combining social network with user-items interactions to build new social recommenders based on graph attention networks and metric learning.The main contributions are summarized as below:(1)Current social recommenders typically assume that connected users in the social network have similar preferences to items,which may introduce too much noise.In addition,a unified representation learned by them is insufficient for finer-grained user modeling.To address these two issues,we consider combing user-item interactions and social relations under relation modeling framework and then propose a new social-based model called RML-DGATs(Relational Metric Learning with Dual Graph Attention Networks).It adaptively aggregates user's or item's neighborhood information by two carefully designed dual graph attention networks(DGATs)structures in respective domain and then adopts two MLPs to model the complex interactions between two neighborhoods as relation vectors.With these two kinds of relation vectors,each user could be translated to both multiple item-aware and social-aware representations to achieve finer-grained user modeling.In addition,we propose distance and neighborhood regularizations to better modeling the relation vectors.Finally,the above modules are jointly trained under a unified metric learning framework.(2)Furthermore,considering that both high-order neighborhood information in item and social domains are beneficial for mining deeper user-item or user-user relations,we propose another social recommender So HRML(Social Recommendation with Highorder Relational Metric Learning)based on RML-DGATs.To fully encode the high-order neighborhood features,the proposed model builds two layer-wise aggregation structures in two domains.By stacking this layer-wise structure,high-order neighborhood information could be adaptively encoded into the learned vectors.We also formulate the matrix form of the layer-wise aggregation process in two domains,which could accelerate the training process and also avoid the node sampling procedure.In addition,to better model the relations between positive and negative samples,we propose an adaptive margin learning method to build the loss function.Finally,the two relation modeling parts are also trained under the metric learning framework.(3)Finally,we conduct extensive experiments along with detailed model analysis on Ciao,Epinions and Flixster datasets.The results show that compared to multiple competitive baselines,our proposed RML-DGATs model outperforms the best baseline by 1.91% to 4.74% on several ranking-based metrics.By modeling high-order neighborhood information in both domains,So HRML can further improve the recommendation performance.