Statistical modeling of Amazon’s rating time series incorporating herding and fake reviews

Abstract

Online shopping sites such as Amazon and restaurant review sites such as Tabelog calculate product and restaurant scores based on user reviews and provide them to users. However, in recent years, reliability issues have arisen as product providers post fake reviews and engage in advertising. Does the increase in fake reviews and rating actions drive product buying behavior ? We report the results of analyzing the review time series data of four Amazon product genres. First, we show that the distribution of the number of reviews follows a power law, and that the probability law for review submission can be described by the Pitman-Yor process. Next, we detected changes in the review time series under the assumption that there were many fake reviews in the first half of the review time series and not in the second half. The ratio of fake reviews for each product, the number of fake reviews, the average number of reviews before and after the change point, and the average score were evaluated.