Dynamic Heterogeneous Graph Neural Network And Its Application In Recommendation System

As the information on the Internet growth rapidly,the recommendation system has become an indispensable infrastructure in the Internet.The data in recommendation systems has the characteristics of massive,heterogeneous,unstructured,and time-critical,and there are ubiquitous connections between different data points in the recommendation systems.Therefore,dynamic heterogeneous graph is a very suitable data organization method for recommendation systems.Graph neural network is an emerging method for graph data representation learning,which has been applied in many fields including recommendation systems.Although there are relatively mature representation learning algorithms for various types of graph data,there are relatively few researches focus on representation learning on dynamic heterogeneous graphs.There have been many attempts to apply graph representation learning techniques to recommendation systems,and these attempts also proved the effectiveness of graph representation learning techniques in recommendation systems.However,there is a lack of attempts to apply dynamic heterogeneous graph representation learning methods to recommendation systems.Based on the real recommendation task data sets,this paper studies how to use dynamic heterogeneous graph representation learning method to improve the recommendation systems.When applying the dynamic heterogeneous graph representation learning method to the recommendation systems,we need to solve three core problems:graph data organization method that is capable of accurately represents the structure of dynamic heterogeneous graphs;the effective dynamic heterogeneous graph representation learning method,and the weighted sampling method on dynamic heterogeneous graphs.In response to these three issues,this article has conducted the following investigations:(1)A data representation method that can accurately describe dynamic heterogeneous graphs is proposed.Specifically,we propose a dynamic heterogeneous graph representation method based on event sequences.It also shows how to use dynamic heterogeneous graphs to organize recommendation data.Compared with the existing dynamic graph organization methods based on graph snapshots,our proposed method can avoid the confusion of time sequence and better ensure the timeliness of data.(2)A dynamic heterogeneous graph representation method based on hierarchical attention mechanism is proposed.The timing characteristics of data are learned through the temporal attention mechanism;the edge-level attention mechanism is used to deal with the heterogeneity of heterogeneous graphs.Modeling heterogeneous graph data efficiently without the help of predefined meta-paths solves the problem that predefined meta-paths may introduce artificial bias in graph representation learning.(3)In order to cope with the massive data in the recommendation system,this paper proposes a budget-based neighbor nodes sampling method.It can greatly reduce the computational complexity of graph representation learning while preserving important information as much as possible.(4)This paper conducts rich experiments on three real world recommendation task datasets.Comparing the performance of other existing methods and the method proposed in this paper demonstrated the superiority of the method proposed in this paper.In addition,this paper also verified the effectiveness and necessity of the techniques proposed in this article through ablation experiments.