Research On Deep Recommendation System Based On Graph Representation Learning

With the explosive growth of Internet and data,recommendation system becomes indispensable to alleviate the information overload problem.As a prominent research topic,recommendation systems strive to model previous interactions between users and items,investigate user preferences and recommend users their personalized interested items.In actuality,recommendation is capable of mining the historical interactions,and a variety of side information,such as item features,user attributes,and context information.The side information may help alleviate the cold-start and data sparsity issues,and improve the recommendation quality.Moreover,the recommendation system and side information normally form a graph structure,with the majority of items being explicitly or indirectly related to one another.For instance,the user and item interactions constitute bipartite graphs,social networks among users,knowledge graph representing the features of items.The graph structure gives another view of the data connection from high-order dimension and semantical dimension.Thus,graph learning excels in increasing the accuracy of recommendations and developing explainable recommendations.Thus,the implementation of graph representation learning dealing with different types of graph structures becomes important in developing recommendation systems.Currently,graph representation learning has considerable potential.However,there are still obstacles to overcome:(1)The majority of current graph representation learning neglect the effect of high-order and implicit relations,which are important to model the user preference.(2)The node's neighbors in the graph influence this node differently.The node embedding should take into consideration the different semantics of its neighbors.(3)Because the recommendation system incorporates many graph structures,it is necessary to use multiple graph representation learning approaches.The integration of different graph representation learning is also a challenge.To address these issues,we propose the research on deep recommendation system based on graph representation learning,namely different strategies analyzing different graph structures.Our main contributions are listed as follows:(1)First,we study the graph representation learning applied on the bipartite graph.In the recommendation system,the interaction between users and items from the bipartite graph,in which users and items are two distinct kinds of nodes,and their interactions become edges linked to each other.Those edges represent the explicit relations,however,implicit relations also exist among the same type of nodes.Thus,we propose AIRC,a framework that explores the explicit and the implicit relations in the bipartite graph.We first construct the implicit relation graphs and embed the side information into node features,then adopt graph attention mechanism to mine the implicit relation graphs,separating the influence of neighbors.For explicit relations,we use a graph autoencoder.Finally,we train the parameters using these two graphs.The findings indicate that implicit relationships may help increase recommendation accuracy.(2)Second,we study the graph representation learning integrating social networks and recommendation systems.In social networks,members' social connections and proximity influence their choice.Thus social relations are effective to learn user preference.We propose SR-AIR,a framework that leverages social networks to assist recommendation systems.We model the users and items separately.The user side contains the social relations,the interacted items,and the implicit relations among users;the item side contains the users they interact with and the implicit relations among items,exploring the high-order transitive relations among users and items.Then we use a graph multi-attention mechanism for each side and learn them together.Thus we capture the social relations and the implicit relations in the graph and improve the recommendation.(3)Finally,we study the graph representation leaning integrating knowledge graphs and recommendation systems.The knowledge graphs include useful details about the objects and serve to reinforce their semantic relationships.We propose two approaches: 1)an approach based on propagation.We present AKUPP,a framework that applies dual propagation mechanism.One is user preference propagation that explores the user-item interactions' latent features,the other one is attentive knowledge propagation that explores the high-order relations in the knowledge graph.We use sequential learning to integrate the two propagation mechanisms,improving the recommendation quality.2)an approach based on neighbors.We present MNI,a framework that transforms the interactions between users-items,items-features into users' neighbors-items' neighbors.We consider the knowledge graph and recommendation system as one graph and reconstruct the neighbors' graph,thus with a multi-head attention mechanism we can explore the high-order neighbors.Besides,by multi-task learning,we enhance the neighbors' graph with the semantical information in the knowledge graph,allowing us to discern the effect of various semantic relations and therefore improve recommendation accuracy.