Towards Characterizing and Detecting Incentivized Reviews on eCommerce Platforms

Customer reviews play an important role in rankings and visibility on e-commerce sites, and also strongly influence a customer's decision to purchase a product. Motivated by this, malicious sellers engage in incentivized review fraud to inflate their product ratings by providing customers with free products in exchange for five-star reviews, thus compromising review integrity. While there is ample prior work on fake reviews in general, there is limited prior work on incentivized review fraud. In this work, we infiltrate an underground market for fake reviews and implement a custom crawler to collect a dataset of malicious products that seek incentivized reviews. We devise and extract a set of features, and show that these are statistically significant in differentiating between benign and malicious products. Using hypothesis testing, we identify characteristics and trends exhibited by malicious products. While we are unable to achieve a high precision when we train standard machine learning models without compromising on the recall, we propose a lightweight two-phase technique that combines high-precision product classifiers with high-recall review classifiers. This technique allows us to minimize the false positives, with only a slight increase in false negatives. Finally, we also audit the effectiveness of two publicly available tools for incentivized review detection and find that they are not reliable. In summary, we contribute a new high-fidelity dataset, characterize products seeking incentivized reviews, audit existing tools for review analysis, and present a superior method for detecting review fraud. We hope that this research could be useful for e-commerce companies and other entities who have a stake in preserving opinion and review integrity online.