Entity Ranking by Learning and Inferring Pairwise Preferences from User Reviews

Abstract

In this paper, we propose a method of ranking entities (e.g. products) based on pairwise preferences learned and inferred from user reviews. Our proposed method finds expressions from user reviews that indicate pairwise preferences of entities in terms of a certain attribute, and learns a function that determines the relative degree of the attribute to rank entities. Since there are a limited number of such expressions in reviews, we further propose a method of inferring pairwise preferences based on attribute dependencies obtained from reviews. As some pairwise preferences are less confident, we also propose a modified version of a learning to rank method, Fuzzy Ranking SVM, which can take into account the uncertainty of pairwise preferences. The experiment was carried out with three categories of products and several attributes specific to each category. The experimental results showed that our approach could learn more accurate pairwise preferences than baseline methods, and inference based on the attribute dependency could improve the performances.