Low-frequency, high-impact: Discovering important rare events from UGC

Abstract

Rare events are those with a small occurrence probability, but might have a substantial impact. Despite their potential importance, previous research hardly considers rare events because of the difficulty of handling long-tail information. We propose a solution that incorporates a large-scale topic model to extract them from customer opinions and network inference to select high-impact variables. We test our model by analyzing hotel reviews. The result indicates that some of the rare events (e.g., “misleading description”, “home feeling”) can be highly influential to customer ratings or recommendations, which helps marketers sharpen their understanding of consumers and redesign their services accordingly.