| In the information age,recommender systems have become an indispensable part of life as an effective method to handle information overload.Graph data structures widely exist in various recommendation scenarios,such as bipartite graph representing the interaction between users and items,social networks between users and attribute networks of items.These data express information about users and items from different aspects,and can effectively alleviate the connection sparse problem on the graph.In recent years,with the development of graph neural network and its superior performance in graph representation learning,a variety of recommendation methods based on graph neural network have emerged.In this paper,we consider the complex heterogeneous user-item graph that introduces the auxiliary connection between users and items and propose recommendation methods based on heterogeneous graph neural networks.The specific research contents are as follows:(1)We propose a recommendation method based on heterogeneous graph convolutional networks.Graph convolutional networks(GCNs)aim to learn meaningful representations on graphs and have been widely used in recommender systems.However,GCN encounters over-smoothing problem when the number of layers increases,which indicates that long paths between users and items on the user-item graph are not very useful for embedding learning in recommendation.Therefore,auxiliary connections between users and items are introduced to shorten the average path length between users and items,which also makes the user-item graph become a heterogeneous user-item graph containing multiple edge types.The method first introduces user-user and item-item connections to construct a heterogeneous user-item graph,and then decomposes it into three homogenous subgraphs of single edge type according to edge types,while preserving the heterogeneous user-item graph to keep the original global graph structure information.Apply GCN to learn representations of users and items from subgraphs and global graph.Then the attention mechanism is used to integrate the representations from the subgraphs and the global graph to obtain the final user and item representations,and finally the likelihood of the user liking the item is predicted by the dot product of the user and item representations.Extensive experimental results on three public datasets show that the proposed method significantly improves the recommendation performance compared to the state-of-the-art methods.(2)A recommendation method based on self-supervised heterogeneous graph convolutional network is proposed.Due to the sparseness of the introduced auxiliary connection data itself,and the existing modeling methods based on heterogeneous graphs usually decompose the heterogeneous graph into single edge type subgraphs for modeling separately,which violates the original intention of connection sparsity that leads to the suboptimal results of heterogeneous data.This method constructs two views,a local view and a global view,on the heterogeneous user-item graph,uses the representation learned on the homogeneous subgraph and the representation learned on the original heterogeneous user-item graph to construct an auxiliary self-supervised heterogeneous contrast learning task to capture the semantic information represented by different edge types,and better node vector representations are generated by joint learning the recommendation task and the auxiliary contrast learning task.The model not only takes full advantage of the rich information of heterogeneous graphs,but also effectively learns the more robust users and items representations.Extensive experiments on three public datasets show that the proposed model can effectively enhance the performance of recommender systems. |
