Modeling User Micro-Behavior via Adaptive Multi-Attention Network for Session-based Recommendation

Abstract

Session-based recommendation (SR) aims to recommend the next item that matches the user’s preferences based on their short-term session information.It does not need user’s profile and long-term historical information and has a broad application prospect.Existing SR models usually focus on user click behavior or only use a single type of behavior data,ignoring the specific semantics of user click behavior,such as browsing,collecting,carting,purchasing,and so on.These behaviors with different semantics are called micro-behavior,which reflects the transformation of user intention and decision-making process in the shopping process from the micro level and will provide valuable information for improving the recommendation effect.Therefore, an adaptive multi-attention network (AMAN) based on user micro-behavior is proposed for session-based recommendation.First,we model the sequence of micro-behavior as heterogeneous directed graph,and then build three components for session-based recommendation.Directed graph attention network (DGAT) learns item representation from the item level and adaptively captures the associations between items with the same micro-operation.Operation-level heterogeneous graph attention network (OHGAT) learns item representation at the operation-level and adaptively captures the associations between items with different micro-ope-ration.Micro-behavior co-attention network (MBCAT) learns the representation of micro-behavior sequences and adaptively captures the dependencies between different micro-behavior sequences.Experimental results on Yoochoose,Taobao14 and Taobao15 datasets show that AMAN outperforms the state-of-the-art baselines.

Keywords

• session-based recommendation|micro-behavior|heterogeneous directed graph|attention network