Research On Graph Neural Network For Session-based Recommendation Systems

Recommendation System(RS)is one of the most common applications of artificial intelligence,which can help users alleviate the pressure of information overload.However,traditional recommendation systems usually require user profiles and historical interaction records,but with the increasing conflict between personalized recommendation and personal privacy,many users’ identities and information may be unknown.To solve this problem,Session-Based Recommendation System(SBRS)uses anonymous session sequences generated by users during consumption to learn their preferences.The most advanced method currently is the session-based recommendation system based on Graph Neural Network(GNN),but there are some issues:(1)Most of the current research only considers the interaction between items within a single session,but ignores the interaction information between items in different sessions,which is important for recommendation performance regarding different but highly similar sessions.(2)Existing recommendation system models only represent the last clicked item by the user as short-term preference,but ignore the interaction between click amounts and relative position information within the session.Solving these issues is key to improving the performance of SBRS.To address the aforementioned issues,this paper proposes two corresponding improved models,with the following main innovations:(1)To address the issue of extracting cross-session information in existing session-based recommendation models,this paper proposes a Neighbor Enhanced Contextual Graph Neural Network(NECGNN)session-based recommendation system.This model first uses K-nearest neighbor algorithm to select the Top-K similar neighbor sessions to the target session,and models the most similar sessions into the same graph to form a neighbor-enhanced graph,which is used to learn the interaction information between items in different but similar sessions and alleviate the problem of limited item interaction in a single session.Secondly,this model utilizes the Fastformer network to capture the global context information of the session to obtain the user’s global preferences.The proposed NECGNN model is experimentally compared with other advanced recommendation models on Diginetica and Yoochoose1/64 two publicly available datasets,and the experimental results demonstrate that the proposed model has higher recommendation accuracy than other models,verifying the superiority of the proposed model.(2)To address the problems that existing models mostly simply represent the last item of the session as the user’s short-term preference and ignore the relative positional information of items within the session,this paper proposes a Hawkes Process and Graph Neural Network(HPGNN)session-based recommendation system.The model proposes a dual-stream structure consisting of a graph neural position-aware layer and a graph neural Hawkes layer to learn the user’s long-term and short-term preferences,respectively.On the one hand,for the graph neural position-aware layer,GGNN is first used to capture the interaction between nodes for subsequent processing,and a sequentially decreasing residual network is introduced to fuse past encoding information with the current network.Secondly,a position-aware attention network is introduced to capture the location information of the session.Specifically,the location matrix of the session is used to query the position vector embedding of the item,which is added to the final item representation to obtain the position-aware item representation,representing the user’s long-term preferences.On the other hand,inspired by the Hawkes process(HP),for the graph neural Hawkes layer,the Hawkes intensity function is introduced to consider the interactive information between the number of clicks of the items,and the Hawkes process is combined with the graph neural network to capture the interactive influence between the click numbers,which is regarded as the user’s short-term preference,to improve the recommendation performance of the proposed model,making up for the shortcomings of only taking the last item’s representation as the user’s short-term preference.This paper compares the proposed HPGNN model with other recommendation models on Diginetica and Yoochoose1/64 two publicly available datasets,and demonstrates that the proposed model has superior recommendation accuracy compared to other models,thereby verifying the superiority of the proposed model.