AMBR: Boosting the Performance of Personalized Recommendation via Learning from Multi-behavior Data

The performance of personalized recommendation can be further improved by exploiting multiple user behaviors (e.g., browsing, adding-to-cart, product purchasing) to predict items of user interests. However, the challenge lies in how to accurately model the relations among multiple user behaviors. The commonly adopted cascade relation over-simplifies the problem and cannot model the real user behavior patterns. In this paper, we propose a novel multi-behavior recommendation algorithm called AMBR (Attentive Multi-Behavior Recommendation), which can well capture the complicated relations among multiple behaviors. AMBR integrates the representation learning module and the matching function learning module into one framework. By utilizing the modern neural network techniques, AMBR is more flexible in modeling the relations of multiple behaviors without presuming a fixed cascade relation. Finally, we also conduct a set of experiments based on two real-world datasets, and the results show that our AMBR algorithm significantly outperforms other state-of-the-art algorithms by over 8.6%, 9.3% in terms of HR and NDCG.