Detecting fake reviews through topic modelling

Against the uncertainty caused by the information overload in the online world, consumers can benefit greatly by reading online product reviews before making their online purchases. However, some of the reviews are written deceptively to manipulate purchasing decisions. The purpose of present study is to determine which feature combination is most effective in fake review detection among the features of sentiment scores, topic distributions, cluster distributions and bag of words. In this study, additional feature combinations to a sentiment analysis are searched to examine the critical problem of fake reviews made to influence the decision-making process using review from amazon.com dataset. Results of the study points that behavior-related features play an important role in fake review classifications when jointly used with text-related features. Verified purchase is the only behavior related feature used comparatively with other text-related features.

Automated summary

Consumers can greatly benefit from reading online product reviews to counter the uncertainty caused by information overload in the online world. However, some reviews are deceptively written to manipulate purchasing decisions. This study aims to determine the most effective feature combination for detecting fake reviews, including sentiment scores, topic distributions, cluster distributions, and bag of words. The results suggest that behavior-related features, particularly verified purchases, play a crucial role in accurately classifying fake reviews in conjunction with text-related features.