Research And Implementation Of Session Recommendation Algorithm Based On Graph Neural Network And Self-supervised Learning

The objective of the session-based recommendation algorithm is how to make accurate recommendation for the scenarios where the user’s historical behavior information cannot be utilized.Obtaining effective representation of anonymous users is the key to solve this problem.Early session-based recommendation algorithms adopt Markov chain method for modeling.Since this method independently combines past operations,the assumption of such independence is too strong,thus limiting the accuracy of prediction.At present,the session-based recommendation algorithm uses the graph neural network for modeling,directly constructs the session data of anonymous users into graph structure data,uses the graph neural network to capture the transformation relationship of items in the session sequence,and then obtains the representation of the whole session as the representation of anonymous users for personalized recommendation.However,some time sequence information is lost in the process of converting the session sequence data into graph structured data,and the recommendation based on a single session has the problem of data sparsity,which leads to the low accuracy of recommendation.The session-based recommendation system needs to solve the following key problems:firstly,how to solve the problem of time sequence information loss in the process of converting session data into graph structured data;Second,how to solve the problem of insufficient node representation information obtained from single session learning;Finally,how to solve the problem of data sparsity in session recommendation?In view of the above key scientific issues,the main research content and innovation points of this paper are as follows:(1)Aiming at the problem of information loss in the process of converting session data into graph structured data in graph neural network modeling,a session graph coding method for adaptive calibration of time sequence information for edges is proposed.This method assigns the timing information of transitions between items in a session to the edges of the session diagram throughout the session.In the learning process of graph neural network nodes,different fusion modes are adopted according to the representation of edges.And the representation of the edge is constantly updated adaptively during the learning process.(2)A session recommendation algorithm based on unsupervised clustering and self-supervised learning is proposed to solve the problem of insufficient node representation information and data sparsity obtained from single session learning.Firstly,all sessions in the current training are constructed into graph structure data,which is called session global graph.The session global graph represents the session data of all anonymous users in the current system and contains general information about transitions between projects.The global representation of a node learned by a graph neural network contains more information than that learned by a single session.In order to better reflect the preferences of current anonymous users,this paper adopts the unsupervised clustering algorithm without parameters to get the cluster center of the project and finally get the category representation of the project.The category information of all items in the current session is fused to obtain the session-level category information that represents the preferences of the current anonymous user.On this basis,the comparative learning paradigm in self-supervised learning is used to construct auxiliary tasks and improve the accuracy of the final recommendation results.