SessNet: A Deep Hybrid-state Session-based Recommender System

Abstract

The present article studies static and dynamic signals in Session-based recommender systems in order to track and estimate users' actions. Session-based recommendation is the task of predicting user actions during short online sessions. Previous work considers the user to be anonymous in this setting, with no past behavior history available. In reality, this is not often the case. A seamless integration of the user history when available has not been offered prior to the present work in the session-based recommendation context. In this paper, we propose a novel deep hybrid-state session-based recommender system, called SessNet. SessNet performs next-click prediction and takes advantage of historical user preferences when accessible. To that end, SessNet's architecture is designed to be a hybrid of two states, namely, the dynamic and static states. First, the dynamic state, adopted from CyberBERT, employs a bidirectional transformer network to model short-term and long-term session intent. Second, the novel static state provides a deep user profile, drawing on rich item and session embeddings obtained from the dynamic state. These user representations along with the current session are processed to predict the next click. We evaluate the efficacy of the proposed method using a benchmark dataset, namely, DIGINETICA. Experiments show that our architecture achieves state-of-the-art session-based recommendation for P@20 and MRR @ 20 on this dataset.