Research On Session Recommendation Method Based On Graph Neural Networks

The recommendation system makes personalized recommendation to users through processing the massive interactive data between users and items.Session recommendation is an important branch of recommendation system,which aims to solve the problem of inaccurate recommendation results for anonymous users.Session recommendation can recommend products to users in real time.It can only make relevant recommendations for users based on the user's click history,and bring good use experience to users.At present,there are three kinds of session recommendation methods,which are based on traditional machine learning,deep learning and graph neural network.Through the analysis and summary of these three kinds of methods,it is found that there are still three problems in the current research methods: First,the latent information of the neighborhood nodes in the session graph is not fully utilized,and most of the encoding methods of the session graph is still only transforming it into sequence data for processing;Second,in the process of deep network training,the long distance historical session information will be missed in the transmission process,and the information can not be completely transmitted;Third,in the process of obtaining session information,it is necessary to consider the influence of different click item information on the recommendation results.In view of the above problems,this paper proposes a graph convolutional recurrent neural networks for session-based recommendation and a graph dilated causal convolutional recurrent Attention neural networks for session-based recommendation.The main work of this paper is as follows:First,this paper proposes the graph convolutional recurrent neural networks for sessionbased recommendation model.Considering that the spatial structure of node distribution and the feature information transmitted by neighbor nodes in the session graph can not be obtained in the process of modeling by using recurrent neural network alone,this paper proposes a session recommendation model combining graph convolutional neural network and recurrent neural network.The model first constructs the session data into the session graph,then inputs the session graph data into the graph convolutional neural network layer,and gets the clickitem latent vector which contains the spatial structure information of the node of the session graph and the propagation feature information of the neighbors.Next,this model inputs the click-item latent vector into the recurrent neural network layer to capture the click-item dependency and timing information of the session data and to obtains the click-item hidden layer vector representation.Then,the model combines all the click-item hidden layer vectors to form the global session vector,and splice them with the local session vector to form the session embedding vector.Finally,the model linearly transforms the session embedding vector to the click item vector to get the recommended results.Second,this paper proposes the graph dilated causal convolutional recurrent attention neural networks for session-based recommendation model.Aiming at the problem that long distance historical information will be missed in deep network training,the model uses an dilated causal convolutional neural network layer to encode and infer session graph data.The session graph data is input into the causal convolutional layer with the dilated factor,and the residual blocks are added to the dilated causal convolutional layer to improve the generalization ability of the model,so that the session click-item embedding vector containing more latent information is obtained.By inputting the click-item embedding vector into the recurrent neural network,the click-item hidden layer vector representation containing the dependencies and timing information between the click-item in the session is obtained.The vectors of hidden layer are input into the attention network layer,and different weights are assigned to the vectors of different click-item hidden layer.The vectors of hidden layer of click-item with different weight values are combined into global session vectors and spliced with local session vectors.The final recommendation result is obtained by linearly transformation.Since the session data is causal and has strict time constraints,it can be modeled by causal convolution.The introduction of dilated factors can reduce the number of network layers while obtaining a larger receptive field,thus more input information can be captured.The residual network spreads the shallow network layer information to the deep network through the residual connection,so the entire extended convolutional network layer solves the problem of missing connection of long distance historical information.The attention mechanism assigns more weight to important click item vectors,so that the model can obtain more useful information,and ignores the influence of irrelevant information on the model recommendation results,thus improving the recommendation effect of the model.In this paper,experiments were conducted on two public datasets,Yoochoose and Diginetica.The experimental results show that,on the two data sets,the two models proposed in this paper are improved by 0.74% and 1.11% on P@20,and on MRR@20 Increased by 0.73%and 1.30%.Thus the effectiveness of the models is verified.