Review-Based Recommendation Under Preference Uncertainty: An Asymmetric Deep Learning Framework

Abstract

Online reviews are one of the most trusted resources for inferring customer needs and understanding consumer decision-making behavior. This study attempts to integrate textual reviews and user-item ratings to improve recommendation performance. To achieve this goal, we propose a deep neural network model. Specifically, the proposed model applies a review-level aggregation strategy to learn user preferences while using an aspect-based document-level aggregation strategy to learn item representation. In this process, we introduce two attention modules at the review and aspect levels, respectively. The review-level attention is used to learn the user preferences that are most related to the target item. The aspect-level attention attempts to learn the item's aspect features that users are most concerned about. In addition, we design a latent stochastic attention mechanism based on the probabilistic generative mechanism, to model the user preference uncertainty. For evaluation, we conduct extensive experiments on several real-world datasets. Using several state-of-the-art methods as comparisons, we find that the proposed model can significantly improve rating predictive power in the context of the recommendation system. Based on ablation experiments, we find that the enhanced predictive power benefits from the preference uncertainty and the attention mechanism. From the qualitative analysis, we suggest that the proposed model can yield many interpretable results.