Session Enhanced GNN Algorithm For Recommendation

With the rapid development of the Internet and related technologies,people have to face the annoyance of excessive information,which requires recommendation systems to help people filter out redundant and uninteresting content.In today’s popular e-commerce and short video fields,there is a lot of session data,and the user’s interests and preferences it represents are short-term and changeable,and traditional recommendation algorithms are no longer suitable for session-based recommendation scenarios.Session-based recommendation algorithms are designed to process session data generated by anonymous users,so they have received much attention in recent years.This paper proposes an improved Session Enhanced GNN Algorithm for Recommendation,SEGR.The main research work of the paper is reflected in the following aspects:First of all,the emerging graph neural network technology can well capture the conversion relationship of items in a session,but some sessions contain meaningless interactive behaviors,and the dependencies these actions represent affect the item embeddings learned by graph neural networks.In the stage of learning item embeddings,SEGR uses a self-attention mechanism and graph neural network.Complement each other and learn more representative item embeddings.Second,in order for the session embeddings to more fully represent dynamic user preferences,in the learning session embeddings stage,SEGR combines the reverse position information with the item embeddings to enhance the sequential relationship of the item embeddings,and abstract item embeddings into session embeddings by using a multi-target soft attention mechanism.Existing session-based recommendation algorithms do not fully consider the impact of similar sessions on recommendation performance,to enhance the expressive power of a single session embeddings,SEGR introduces the connections between similar sessions by constructing a global session graph,and designs a graph attention layer that fuses co-occurrence parameters to capture the collaborative information between sessions.Finally,extensive experiments are carried out on three public datasets,Yoochoose1/64,Diginetica and Tmall.The comparative experimental results show that SEGR wins in the performance comparison with 10 existing baseline algorithms,which verifies the effectiveness of the SEGR proposed in this paper.In addition,the ablation experiments,session embeddings experiments and hyperparameter experiments designed in this paper verify the rationality of SEGR from multiple perspectives.