DS-KGAT: A Deep Session GAT with Knowledge Enhancement for CTR Prediction

Le résumé fourni par la source

Click-through rate (CTR) prediction accuracy is an important metric for recommender system strategies, and it always involves dynamic and continuous features. Previous methods based on user historical behavior usually capture the short-term interests of users, while ignoring the complex transitions between various historical behaviors of users and lack of interpretability. In order to model the user behavior sequence and capture the user fine-grained interests, we propose a novel model named Deep Session GAT with Knowledge Enhancement (DS-KGAT). Specifically, the user behavior sequence is divided into sessions of multiple graph data structures. On the basis of the session graph, Graph Attention Network (GAT) can represent each node as an aggregation of neighbor node embeddings. At the same time, the user’s interest in each session is propagated on the knowledge set to form the user’s preference distribution for candidate items. Then combining the long-term and short-term interest attention mechanism, each session is represented as a combination of item node embeddings. Finally, we utilize Bi-LSTM to model the interactions between sessions and the long-term evolution of user interests. We have done extensive experiments on the genuine dataset. The experimental consequences demonstrate that DS-KGAT improves the accuracy of click-through rate prediction, and has made substantial progress.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.