Learning to Rank Items of Minimal Reviews Using Weak Supervision

Abstract

Customer reviews and star ratings are widely used on E-commerce and reviewing sites for the public to express their opinions. To help the online public make decisions, items (e.g., products, services, movies, books) are typically represented and ordered by an aggregated star rating from all reviews. Existing approaches simply average star ratings or use other statistical functions to aggregate star ratings. However, these approaches rely on the existence of large numbers of reviews to work effectively. On the other hand, many new items have few reviews. In this paper, we argue that at the core of review aggregation is ranking items, hence, we cast the problem of ranking a set of items as a learning to rank (L2R) problem to address the issue of reviews scarcity. We devise a rank-oriented loss function to directly optimize the ranking of groups of items. Standard L2R models require ranking labels for training, but item ranking ground-truth information is not always available. Therefore, we propose to aggregate star ratings for items with large numbers of reviews to automatically generate weak supervision ranking labels for training. We further propose to extract features from review contents, rating distributions and helpfulness information to train the ranking model. Extensive experiments on an Amazon dataset showed that our model is very effective compared to state-of-the-art heuristic aggregation approaches, regression and standard L2R approaches.