Research On Session-Based Recommendation Model Of Graph Neural Network Based On Neighborhood Information

Session-based recommender systems emerge to predict the next item that the user is likely to interact with based on the anonymous user’s behavior log in the current session.Compared with traditional recommendation systems,session-based recommender systems can provide more accurate and real-time recommendation services by portraying users’ behavior patterns within a period and capturing their shortterm dynamic preferences only based on session data without user profiles.With the booming development of deep learning,more and more neural network methods are introduced to the session-based recommendation task.Graph neural networks are favored by researchers because of their powerful ability to model complex transition relationships among graph data.However,the current mainstream sessionbased recommendation methods still can be improved in the following aspects: first,most methods focus on capturing item transition patterns within the current session while ignoring potential collaborative behaviors among different users from other sessions;second,existing methods do not consider the various motivational factors that drive users to interact with items when modeling user intent.To address the above issues,this paper proposes two session recommendation methods based on graph neural networks to improve the recommendation performance.The main contributions of this paper are as follows:(1)A Neighbor-Enhanced Graph Transition Network,NEGTN is proposed,which leverages the useful collaborative information from the neighbor sessions to enhance the recommendation performance.Firstly,the model samples the neighbor-session set for each current session and converts them into graph structures.For the current-session graph,it extracts bidirectional transition signals between items based on a gated graph neural network to capture the user’s current preference.For the neighbor-session graph,a weighted graph attention network is employed to learn global collaborative information by considering the popularity of item transition while reducing the influence of irrelevant items in the neighbor sessions.After adaptively combining these two types of features with a gated mechanism,the model further introduces a positional attention mechanism to investigate the impact of items in different positions on the user’s true intention.(2)A Global-Enhanced Disentangled Neural Network,GEDNN is proposed,which applies the idea of disentangled representation to infer the key potential motivations behind the user decisions.The model considers the influence of neighbor node types and their relative distances on the current node when constructing the global graph.In order to capture the key latent factors that affect user behavior in the global context,the initial item embeddings are mapped to multiple subspaces and then fed into a positionaware disentangled graph neural network to encode item transition features under different latent factors.For the current session,a disentangled multi-head self-attentive network is designed to decouple the unary features from pairwise features of items so as to better learn the complex relationships between items.The above two types of representations are deeply fused by a cross-attention network and thus generate the final session representation to predict the next user’s interest item.Extensive experiments and analyses of the proposed models are conducted on multiple public datasets.The experimental results demonstrate that both NEGTN and GEDNN can effectively improve the session recommendation performance with a certain superiority.