Research On Cross Sessions-based Graph Neural Network Recommdation Algorithm

With the development of the Internet,more and more information is generated by the network.In e-shopping and videos web sites,recommendation systems is the first choice to search information for users.The core of the recommendation system is the recommendation algorithm.The traditional recommendation algorithm generate the recommendation result with the user’s long-term historical behavior,item characteristics,and the user’s short-term behavior.With the increase of anonymous websites and the emphasis on private data,it is more and more difficult to collect user behavior records.Relying only on item characteristics and short-term user behavior,the performance of traditional recommendation algorithms will drop a lot without long-term historical behavior of users.SessionBased recommendation is based on the user’s short-term behavior records and item characteristics,and does not reply on the user’s long-term historical behavior.For this reason,session-Based recommendation has received more and more attention.The graph is good at expressing the transfer relationship of sessions items,and the graph neural network is the mainstream direction of sessionbased recommendation.Session-based recommendation algorithms which based on graph neural networks usually establish a global session and expand the information available in the current session.However,the global session expands sessions with different interest trends as contexts,which cannot accurately capture the session interest trends,resulting in a decrease in the accuracy of the recommendation results.Aiming at the problem that session-based recommendation cannot capture the interest trend of session interest,this paper proposes a cross session-based graph neural network algorithm(CSGNN).CSGNN establishes cross sessions which are based on collaborative mode,and divides the sessions into cross session group with similar interest trends through collaborative mode.Session-based recommendation cannot capture the interest trend of session interest,which improves the accuracy of recommendation results.Comparative experiments results show that CSGNN has a 3%improvement in Recall@20 and MRR@20 compared to mainstream graph neural network session-based recommendation algorithms.CSGNN converts the sequence into directed edges of the graph when modeling a session.The graph structure of a long session is usually huge and sparse,and there is little information shared by other nodes,and the accuracy of the recommendation results decreases.The self-attention mechanism treats any item distance as 1 when processing sequences,and it is good at processing long sessions.Session-based recommendation algorithms which are based on attention mechanism usually use Gap-filling pre-trained item embeddings,but mask-label makes the probability distribution of items in the train data set and test data set inconsistent.In addition,the position encoding of the attention mechanism model is a constant that does not change with the session,and cannot be generalized to the session content.Aiming at the problem that CSGNN is difficult to deal with long sessions,this paper uses attention mechanism to help CSGNN deal with long sessions,proposes a CSGNN recommendation algorithm fused with attention mechanism,and uses FSG(forward sequence graph)position encoding to solve attention mechanism model cannot generalize the content of the session.Instead of Gap-filling,the item embedding of CSGNN solves the problem of inconsistent probability distribution of items in the train data set and test data set,and improves the accuracy of the recommendation results.The comparative experiments results show that the CSGNN fused with the attention mechanism has a 1%improvement in Recall@20 and MRR@20 compared to the ordinary CSGNN.In order to verify the effectiveness of the algorithm proposed in this paper,this paper compares the proposed algorithm with other mainstream session-based recommendation algorithms on the datasets MovieLens and Yoochoose.The experimental results show that the algorithm proposed in this paper is better than mainstream session-based recommendation algorithms in multiple evaluation indicators.