Deep attention dynamic representation learning networks for recommender system review modeling

Abstract

Despite considerable research of utilizing deep learning technology and textual reviews in recommender systems, improving system performance is a contentious matter. This is primarily due to issues faced in learning user-item representations. One issue is the limited ability of networks to model dynamic user-item representations from reviews. Particularly, in sequence-to-sequence learning models, there appears a substantial likelihood of losing semantic knowledge of previous review sequences, as overridden by the next. Another issue lies in effectively integrating global-level and topical-level representations to extract informative content and enhance user-item representations. Existing methods struggle to maintain contextual consistency during this integration process, resulting in suboptimal representation learning, especially attempting to capture finer details. To address these issues, we propose a novel recommendation model called Deep Attention Dynamic Representation Learning (DADRL). Specifically, we employ Latent Dirichlet Allocation and dynamic modulator-based Long Short-Term Memory to extract topical and dynamic global representations. Then, we introduce an attentional fusion methodology to integrate these representations in a contextually consistent manner and construct informative attentional user-item representations. We use these representations into the factorization machines layer to predict the final scores. Experimental results on Amazon categories, Yelp, and LibraryThing show that our model exhibits superior performance compared to several state-of-the-arts. We further examine the DADRL architecture under various conditions to provide insights on the model's employed components.