One-Class Recommendation with Asymmetric Textual Feedback

Personalized ranking with implicit feedback (e.g. purchases, views, check-ins) is an important paradigm in recommender systems. Such feedback sometimes comes with textual information (e.g. reviews, comments, tips), which could be a useful signal to reveal item properties, identify users’ tastes and interpret their behavior. Although incorporating such information is common in explicit feedback settings (such as rating prediction), it is less common when dealing with implicit feedback, as it is often not available for negative instances (e.g. there is no review associated with the item the user didn’t buy). Thus our goal in this study is to propose a ranking method (PRAST) to incorporate such personalized, asymmetric textual signals in implicit feedback settings. We evaluate our model on two real-world datasets. Quantitative and qualitative results indicate that the proposed approach significantly outperforms standard recommendation baselines, alleviates ‘cold start’ issues, and is able to provide potential textual interpretations for latent feedback dimensions.