Research On Session-based Recommendation Algorithm With Graph Neural Network

In the era of big data,the amount of information in the Internet has increased significantly and the problem of information overload has become increasingly serious.Recommendation system is an important technology to alleviate information overload.It uses data and recommendation algorithms to mine user interests and has great commercial values.When user information is unavailable,the traditional recommendation methods lose their effects.For addressing this issue,a session-based recommendation method has been proposed recently,which predicts users’ behavior based on anonymous session data and has high practical values.We study the session recommendation algorithms based on graph neural network.Compared with the recurrent neural network used in previous session recommendation models,the graph neural network based on session graph can better model complex transitions among items in the sessions,generate accurate representations of items,and then obtain the representation of the entire session.However,the existing graph neural network session recommendation models still have some problems:(1)When constructing the session graph and generating the session representation,the models loss the position information of the items in the session sequences;(2)In the information aggregation stage,the models using gated graph neural network to generate the item representation did not consider the features of the item nodes in the session graph;(3)Most models only focused on modeling the current single session,ignoring the session context information.In order to solve the above problems,we carry out the following research work:Firstly,regarding the first two problems,we propose a position-aware gated graph attention network model.In this model,a reverse position coding mechanism is designed to express the position information of the items in the session sequences,and then obtain the position-aware session graph and session representation.Meanwhile,a gated graph attention network used to generate item representations is constructed.By introducing a self-attention mechanism in the information aggregation stage,the gated graph neural network used in the session recommendation model is improved,so that the model can adaptively aggregate the information on the session graph according to the features of the item nodes.Secondly,on the basis of the aforementioned model,we propose a graph neural network model with session context fusion.In order to capture the session context information and filter out the item noises,an intentaware graph network is constructed.Based on the global session graph,a global item representation is generated,which contains session context information.At the same time,by combining both the global item representation and the item representation of current session,the session context information is introduced at the item-level.We evaluate the effectiveness of the two proposed session recommendation algorithms on two e-commerce datasets,YOOCHOOSE and DIGINETICA.Through the comparisons with multiple popular session recommendation algorithms,it is demonstrated that the solutions we proposed for the aforementioned three problems can effectively improve the recommendation effects.Moreover,the experimental results show that the proposed algorithms can improve the recommendation performance for sessions with different length.