| The data shows a blowout growth as the result of rapid development of technology such as internet,big data,cloud computing etc.Therefore,the problem of information overload is becoming more and more serious.Recommendation technology is an effective mechanism to solve this problem.As an important branch in the field of recommendation,session-based recommendation can use anonymous session sequences to generate recommendation results,in the case of insufficient access to users’ historical data and users’ information.Session-based recommendation is facing an important challenge that how to accurately capture user preference characteristics with limited users’ behavior data.The existing methods based on graph neural network generate the item feature representations by aggregating the information between adjacent nodes in the graph,which cannot accurately express the sequential information contained in the session data and neglect the impact of sequential information on users’ interest preferences.In addition,most of the existing methods adopt the strategy of single type graph modeling,which cannot comprehensively consider user interest preferences.This paper proposes two solutions to the above problems.The main contributions of this paper are as follows:(1)Aiming at the problem of loss sequential information between items in graph neural network modeling,this paper proposes a perceiving sequential dependence graph neural network for session-based recommendation model.The model integrates the item embeddings generated by a graph neural network with positional information,thus preserving the sequential information from the original session sequence.In addition,designing a sequence-perceive attention mechanism to adaptively allocates weights for items,while considering the connectivity between adjacent items and the sequential dependencies among non-adjacent items.This method adequately expresses the sequential information between projects and improves the recommendation performance of the model.(2)Aiming at the problem that the method of single type graph modeling unable to consider users’ interest preference comprehensively,this paper proposes a novel joint learning model.This model can not only mine item transition relationships in the current session through the session subgraph,but also mine item transition relationships across sessions in the hypergraph,and fully integrate users’ current interest preference and potential interest preference contained in the two kinds of item transition relationships.The method effectively improves the ability of the model to capture users’ comprehensive interest preferences and achieves better recommendation results.(3)This paper conducts a large number of experiments on two public datasets,Tmall and Nowplaying,and evaluate various models using the P@K and MRR@K evaluation metrics.The experimental results demonstrate that the proposed method effectively improves the performance of session-based recommendation. |
