Abstract

In online marketplaces like Amazon, opinion spam often takes the form of fake reviews intended to either boost or tarnish a brand's reputation. These reviews can be orchestrated by hired groups targeting specific brands. While previous studies have focused on identifying such groups at the product level, our research explores detecting groups that manipulate opinions about entire brands. We collected and analyzed reviews from Amazon, identifying 923 candidate reviewer groups using frequent itemset mining based on similarities among brands they review. Our hypothesis is that these groups exhibit specific features when targeting brands, such as review consistency and sentiment. We developed a supervised model using these features to classify groups as extremist entities. Using a three-layer perceptron-based classifier, we achieved effective classification. Further analysis of these groups revealed surprising findings, including verified reviewers displaying extreme sentiments, suggesting potential loopholes in current measures against incentivized reviews on Amazon. This study aims to better understand and combat brand-level opinion fraud in online marketplaces.