Research On Session-based Recommendation Methods For Complex Transfer Relarionship

Session-based recommendation is a branch of recommendation system.The essential task is to predict the item that the user may interact with at the next moment based on the session sequence consisted of interactive items in the session data.Session-based recommendation methods based on Graph Neural Network convert session data into the form of session sub-graph,then take advantage of rich neighbor node information and the topology of sub-graph to obtain more complex transfer relationships between items(that is spatial structure information).At present,this type of method has the following problems: first,Graph Neural Network mines the local dependencies of items in the session data,but cannot fully describe the global temporal dependencies of session data;secondly,when the Graph Neural Network is used to update the node information in the session sub-graph,the time that item appears in the session data is not taken into consideration.By capturing the above-mentioned complex transfer relationship,the dynamic change of user's interest and the focus of user's interest can be mined,which are of practical significance for session-based recommendation task.This paper proposes a framework(STSR)based on temporal information and spatial structure information for session-based recommendation.In the hybrid framework,Graph Neural Network,which is used to model session sub-graph,mines more complex item transfer relationships in the session data.At the same time,the hybrid framework uses Gated Recurrent Unit to model the session sequence to obtain the temporal information of session data,and on the basis of the original Gated Graph Neural Network to update the node information,the hybrid framework adds a frequency gating mechanism to control the update of node information in the session sub-graph with reset gate and update gate.The randomness of user behavior would lead to the uncertainty of obtained item transfer relationship.On the basis of the hybrid framework,this paper solves this problem from two perspectives.The first point of view is to start fro m the part of the session data,the first interacted item and the last interacted item in the session data are regarded as important factors that affect the user's initial will and main intention in the current session data.Using attention mechanism to calculate the similarity between these two factor with each item in session data and assign different weight value for each item in the session data,and a session-based recommendation method based on the user's real purpose analysis(STPSR)is proposed.The second perspective is to start from the overall situation of the session data and draw s on the idea of average pooling from temporal perspective.From spatial perspective,the graph representation method is used and the idea of clustering is borrowed to compress node information,which analyzes the current entire session data to obtain the representation of the user's behavior habit,then calculates the similarity between the user behavior habit and each item in the current session data as corresponding weight,so this paper proposes a method named STHSR to analyze user's behavior habit.Experimental results show that our proposed framework STSR,as well as methods STPSR ? STHSR all have greatly improvement on the accuracy of recommendation results compared with baseline methods in this paper.Besides,the improvement of STHSR is greater than that of STPSR.