Sentiment aware tensor model for multi-criteria recommendation

With the advance of sentiment analysis techniques, several studies have been on Multi-Criteria Recommender Systems (MCRS) leveraging sentiment information. However, partial preferences quite and naturally happen in MCRS and negatively affect the predictive performances of sentiment analysis and multi-criteria recommendation. In this paper, we propose a Sentiment Aware Tensor Model-based MCRS named SATM. It maps between i) a set of multiple classes from explicit user feedbacks and ii) sentiments extracted from free texts in user reviews. In particular, we found the four patterns of the partial preferences and applied a rule-based function to detect them and fill their incomplete ratings intuitively. Lastly, we introduce a mapping function of the misinterpretable patterns into sentiment scores in order to generate virtual user preferences that construct the SATM. Experiments on three datasets (i.e., hotel and restaurant reviews) collected from TripAdvisor show that the SATM is superior to various baseline techniques, including state-of-the-art approaches. Additionally, the experimental evaluation of the SATM's variants reveals that the rule-based and mapping functions can handle the partial preferences and improve the MCRS' performance, regardless of target domains.

Automated summary

This paper proposes a sentiment aware tensor model-based multi-criteria recommender system (SATM) that maps user feedback and sentiment information to handle partial preferences and improve the system's performance.