Research On Session Recommendation Algorithm Based On Multi-Feature Fusion

With the development and popularization of the Internet,people are generally faced with the problem of information overload.To alleviate the information overload,recommendation system came into being.The existence of recommendation system greatly reduces the cost of obtaining effective information and improves the user's online experience.Generally speaking,recommendation system could mine users' preferences according to their historical behaviors,and then discovers things that may be of interest to the user from a large amount of information and resources and generate recommendations.However,in practical applications,for many unknown users or new users,their historical behavior is unknown.the recommendation can only be generated according to the user's current behavior.To solve this problem,session-based recommendation appears.At present,session-based recommendation has become a hot issue in industry and academia.Therefore,based on advanced deep learning technology,this paper studies the problem of session-based recommendation.(1)To solve the problem that the session-based recommendation algorithm mainly relies on the limited information in the target session,but does not make full use of the global information between items,a multi-graph neural networks-based session perception recommendation model(MGNN-SPRM)is proposed.First,according to the target session and all sessions in the training set,the item transfer graph and the collaborative association graph are constructed.Based on the two graphs,the graph neural network is applied to gather the information of the nodes,and two kinds of nodes are obtained;Then,the two types of nodes are modeled by the two-layer attention module to obtain the session level representation;Finally,the attention mechanism is used for information fusion to get the final conversation representation and predict the next interactive item.(2)To make up for the limited information in the target session,the item type information can be introduced to expand the information.Therefore,on the basis of MGNN-SPRM,this paper further proposes a category message enhancing-based session perception recommendation model(CME-SPRM).When extracting subgraphs from the collaborative association graph based on the current session,the same kind of nodes can be extracted by using the item type information,which is equivalent to using the item type information to further introduce the global information.In addition,considering that the nodes other than the introduced session have different importance to the target session,the sub-graph is modeled through the graph attention network,and important global information is adaptively selected.(3)Experiments are carried out on two public data sets in the field of e-commerce.The experimental results show that the two proposed models are superior to the frontier benchmark models in all indicators.The CME-SPRM model has better performance than MGNN-SPRM model because it makes use of the information of item types effectively.