Multiple interleaving interests modeling of sequential user behaviors in e-commerce platform

An anonymous user-behavior session in an e-commerce platform is a time-stamped series of sequential implicit feedback (e.g., clicks and orders of items) in a short period, without user profiles available (e.g., the non-logged-in users). The accurate modeling of such sessions is crucial for distributed representation learning in user profiling and item embedding. It broadly spans the recommendation scenarios with capabilities, such as next-click item prediction. The statistics of sessions provided by one of the largest e-commerce platforms indicate that a user generally has multiple interests in a session simultaneously, and the interests may be interleaved. Recent advances in recurrent neural networks and attention mechanisms have led to promising approaches for modeling such sessions. However, few of the existing models explicitly consider the effects of multiple interleaving interests. Through specific data analysis, we find there are two characteristics present in those sessions: 1) contiguous items usually share the same user interest(e.g., category), which can measure the intensity of interests in the current window scope (i.e., local features of interests); 2) each interest repeatedly occurs in a session, which shows its importance in the current session (i.e., global features of interests). Based on the observations, we present a novel framework that provides M ultiple I nterleaving I nterests M odeling with the following contributions: 1) a local layer is adopted to extract the local features of interests by the convolution operations; 2) a global layer is utilized to capture the global features of interests in the current sequence by considering the frequency of items; 3) an Interest-GRU layer is adopted to track each interest’s sequential evolution by fusing local and global features. Experimental results of the next-click prediction task on two real-world datasets demonstrate that our proposed method significantly outperforms state-of-the-art models.