Toward Understanding Causes of Anomaly in Dynamic Restaurant Rating

Abstract

Rating score and text review are the most common features provided in online review systems to gather the opinions shared by users. Product rating distributions usually evolve dynamically over time and potentially accompany with some unusual changes, namely anomalies, which might be caused by product quality change or spamming attacks. In this preliminary study, we analyze the time-series of rating score distributions by using the data collected from Yelp restaurants, and we apply Principal Component Analysis (PCA) to detect anomalous time points. Through manually checking the corresponding review texts, we further investigate the underlying reasons leading to anomalous rating scores. The potential reasons we identified include food/service quality change, user preference, and review spam. Our study is envisioned to help business owners respond timely to unusual feedbacks and manage their business more efficiently.