Researches On Session-based Recommendation Based On Graph Neural Network And Attention Mechanism

With the continuous development of information technology,recommendation systems have been recognized as an effective and powerful tool to solve the problem of information overload for a long time.Compared with traditional recommendation algorithms,algorithms for session-based recommendation systems(SBRS)use the temporal information of user behaviors,which is more effective for modeling users' preferences.The current SBRS algorithms mainly suffer from two problems.Firstly,they usually lack temporal information,especially temporal information of items.Secondly,most of them pay attention to the linear relationship while ignoring the transition relationship of items within the session.Aiming at the above two problems,this dissertation proposes two session-based recommendation algorithms,based on graph neural network and attention mechanism.The main contents are as follows:(1)Aiming at the problem of insufficient feature extraction for temporal information,this dissertation proposes an attention-based algorithm for SBRS,which fully extracts the temporal features of users and items.For users,the algorithm uses the attention network to extract shortterm and long-term features from the current session and past sequential interactions respectively.For items,the algorithm also uses the attention network to extract short-term features,while treating pre-trained features as long-term features.Finally,the prediction layer combines all the features and predicts the result.(2)Aiming at the problem of disregarding transition relationships for items within a session,this dissertation proposes an algorithm based on graph neural network and attention mechanism.After converting sequences of the current session and historical sessions to graph data,the algorithm firstly employs a graph neural network to learn representations for nodes of each graph.Then,an attention mechanism is used for generating graph-level features.Next,a recurrent neural network is applied for extracting dynamic features among historical sessions using their graph-level representations.Finally,the prediction layer combines graph representations of all levels and predicts the result.(3)This dissertation conducts comprehensive experiments and analyses for two proposed algorithms on public datasets.Through parameter experiments,the values of parameters in our models are adjusted to the best.The effectiveness of modules in our models has been verified through ablation experiments.In the comparison experiments,the two algorithms have been compared with classic and advanced algorithms.The experimental results show that the two algorithms we propose have better performance,reflecting their superiority in the task of sessionbased recommendation.