Research On Session-based Recommendation With Graph Convolutional Networks

As an important technology in the field of computer,recommendation system has been applied more and more widely in various fields in the era of big data.People increasingly rely on recommendation system to screen useful information for themselves in massive data.In the scenario where the user information is transparent to the system,the recommendation algorithm based on session has attracted extensive attention.It does not need the user identity information,but only recommends the interested goods for the user according to the user's historical behavior information.Session-based recommendation methods and models based on deep learning have achieved good recommendation effects after being proposed.From the application of recurrent neural network to graph neural network,the performance of the model has been greatly improved.However,there are still many defects in the use of graph neural network,such as the inability to transfer the features of some nodes to the whole graph,and the excessive smoothing problem of deep neural network,which can not well distinguish the features of nodes.This essay makes an in-depth study of the above problems and puts forward corresponding improvement strategies.Because existing models cannot extract deeper node features,this essay proposes a session-based recommendation model HAMA-GCNS based on the graph convolutional network model of high-order aggregation.The model superimposed graph convolution layer to form a deep neural network to aggregate the neighbor node information of higher order and mine the deeper characteristic information.At the same time,the design of message passing mechanism of graph convolution layer alleviates the over-smooth phenomenon of the propagation process.Then,the multi-head attention mechanism is integrated into the conversational expression to improve the conversational expression ability.Based on the HAMA-GCNS model,this essay proposes a session recommendation model,EAPAE-GCNS,which is based on the extended adaptive propagation and attention enhancement graph convolutional network model.In this model,an adaptive propagation process is added into the multi-layer neural network of HAMA-GCNS,and the number of communication steps of each node is calculated adaptively in the training process,so as to determine the number of neural network layers.At the same time,the message passing mechanism of the neural network is changed,and the excessive smoothing phenomenon of the deep neural network is eliminated.In conversation expression,besides using multi-head attention mechanism,we also consider the influence of target item,and add the attention mechanism of target perception to the conversation vector coding.In order to prove the validity of the model proposed in this essay,a comparative experiment is conducted on the open data set.The results show that the two models proposed in this essay are superior to the existing models under the two indexes of Recall@20 and MRR@20.At the same time,a large number of experiments are carried out to demonstrate the rationality of the model proposed in this essay.For example,the recommended accuracy of HAMA-GCNS is 3.01%higher than that of the existing model under the MRR@20 index of Diginetica dataset,and that of EAPAE-GCNS is 4.88%higher than that of the existing model under the zero index of Yoochoosel/4 dataset.