Application Of Graph Neural Network In Session--based Recommendation System

A session-based recommender system(RS)forecasts the item to be clicked next by taking advantage of the item sequence in the current session.The deep learning mod-els,such as the recurrent neural network(RNN)and the graph neural network(GNN),are recently applied in the session-based recommendation.However,to the best of our knowledge,the existing method of session-data modeling ignores the degree of nodes and the fusion of the frequent sub-sequences which refer to those item sub-sequences that ap-pear frequently in different sessions.In addition,intuitively,the more frequently an item sequence appears,the more important it could be.In view of the above problems and phenomena,based on the existing work,this paper studies the role of node degrees in session graphs and the role of frequent subsequences in session recommendation,respectively.The main contributions of this work are as follows:1.Based on the existing work of graph neural network,this paper analyzes the session data,and proposes to use the degree information of nodes in the session graph to improve the effect of session recommendation.This paper constructs in degree matrix and out degree matrix to represent in degree information and out degree information respectively,and models them through gating graph neural network.Through graph visualization,this paper verifies that the degree in the session data has an obvious effect on the model.The validity of degree information is verified in multiple real-world data sets.2.In this paper,we realize the connection between the local and the whole in the session recommendation by the construction of frequent subsequence.In this paper,we propose the frequent sub-sequence modeling(FSM)method.The input of our method is composed of two session graphs.The first graph is constructed by the original item se-quence in the current session.The second one is built by the sub-sequences and the edge weight equals to the occurrence times that each sub-sequence appears in all sessions.Two groups of item feature vectors can be obtained by feeding these two graphs into two GNNs and a gated layer is finally utilized to concatenate these two groups of feature vectors.Our method is tested on four benchmark datasets,namely Tmall,Nowplaying,Diginetica and Yoochoose.It is verified by the results that our method achieves the best improvement on Tmall dataset,in which the distributions of sub-sequences are exhibited in a variety way.