Research On Session-based Recommendation Algorithm Based On Graph Neural Network

Recommender systems are becoming a crucial part of several websites.Traditional recommender systems mine users' interests according to their historical behaviors and predict a list of possible favorite products.However,in many modern e-commerce platforms,users' information is anonymous,so recommenders based on historical behavior are no longer feasible,which leads to a new branch of recommendation systems: session-based recommendation systems(SBRSs for short).The task of SBRSs is to make recommendations based on anonymous sessions.Traditional techniques pay more attention to the long-term interests of users,whereas SBRSs focus more on the short-term preferences of a user's current session.At present,with the increasing number of studies on SBRSs,user-based collaborative filtering techniques are not suitable for recommender systems based on anonymous users.Item-based collaborative filtering techniques consider only the last click in a session,which is insufficient when making recommendations with limited information in a session sequence.The successful application of recurrent neural networks,attention mechanisms and graph neural networks in the field of natural language processing provides a new idea for SBRSs.However,the existing session-based recommendation algorithms still have the following shortcomings:(1)Graph neural network uses the features of items between adjacent nodes,and ignores the relationships between non-adjacent items,resulting in information loss.(2)The existing research results regard the same items in the session as a single node,ignore the differences between items in different positions,and do not use other sessions to learn users' short-term preferences.(3)Most models generally pursue high accuracy and design complex models,resulting in large performance overhead,which makes the proposed algorithm difficult to apply in practical production.To solve the above problems,we start from the structure of graph neural networks,innovate on the existing research results,and design three better session-based recommendation models.The main work of this thesis is as follows:(1)A session recommendation model based on attention mechanism and graph neural network is proposed.This model not only captures the dependencies between nodes,but also considers the high-order features of graphs.It first uses the self-attention mechanism to learn the dependencies between items,then uses the soft attention mechanism to fuse the high-order features,and finally designs a simple fully connected layer to update the embeddings of items,making full use of the structural information in the graph,so that the model pays more attention to the useful information in the session and suppresses the unimportant information.(2)A novel position-aware graph neural network model is proposed,which not only retains the position information of items,but also considers the impact of other sessions on the current users' interest.Firstly,it constructs the session sequence into a position-aware graph,and generates the embedding of nodes combined with the position information,then aggregates the neighbor features of the item and the features of the same item in different positions,and finally integrates the users' long-term interests and short-term preferences for prediction.(3)A lightweight graph neural network model is proposed.The model includes a simple graph aggregator and interest encoder,which not only achieves good recommendation results,but also has excellent effective advantages.The model first constructs the items into a graph,then directly updates the embedding of the item by using the average value of neighbor embedding,and finally uses position embedding to learn the users' long-term interest,combined with the last item for prediction.This thesis makes an in-depth and detailed analysis of the problems in SBRSs,proposes three improved graph neural network models on the existing session-based recommendation algorithms,and carries out extensive experiments on the classical e-commerce dataset.Compared with the latest baseline methods,the experimental results consistently show that the proposed models have excellent performance.