Fusion of latent categorical prediction and sequential prediction for session-based recommendation

Recommendation systems have been becoming ubiquitous in most of online e-commerce platforms, news portals and etc. [1-5], which help users to easily discover their interested items in this era of information explosion. In some application scenarios, users' personal information may not be available beforehand, for example, the click behaviors of a user when browsing websites without logging-in. In such cases, we only have many anonymous users' browsing sequences. Recently, sessionbased recommendation (SBR) has been proposed to predict the next behavior of a user (e.g. the next item to click) for an ongoing anonymous session [6]. According to [7-10], the SBR task can be formally defined as follows. An anonymous session is defined as an item sequence s = { v1, v2, ... , vt} ordered by timestamps, where vi E V represents Session-based recommendation is to predict the next item for an anonymous item sequence. Most of recent neural models have focused on how to learn sessions' sequential representations based on the assumption that items can be projected into a single latent embedding space to describe their latent attributes. In this paper, we argue that an item can also be described by some latent categorical abstractions. To examine our argument, we first mine items' latent categorical distributions via random walk on an item graph constructed from sessions. We design a new neural model which consists of two prediction modules: One is to learn a session's latent categorical representation; The other is to learn a session's sequential representation. Each module independently makes a next item prediction, and their predictions are fused as the final recommendation result. Experiments on three public datasets validate that our model achieves performance improvements over the recent state-of-the-art algorithms. (c) 2021 Elsevier Inc. All rights reserved.