Catch Me If You Can — Detecting Fraudulent Online Reviews of Doctors Using Deep Learning

Abstract

Fake online reviews are becoming more prevalent and are a significant concern for consumer protection groups and regulatory authorities. However, identifying fake reviews has been a challenge in IS, marketing, and computer science. In this study, we design a deep learning approach to capture the linguistic traits that differentiate between genuine and fake reviews. Our deep learning model is evaluated on a dataset of 181,951 doctor reviews, 8% of which are fake. Since a natural honeypot existed at one point on the platform that hosted these reviews, we are able to label the reviews that exploited the natural honeypot as fraudulent, thus overcoming the major challenge in constructing the ground truth for training the model. Our model shows a significant improvement in accuracy when compared to traditional machine learning algorithms such as logistic regression and random forest. Interestingly, we also find that human evaluators perform much worse than machine learning approaches. Compared to 200 human evaluators, our deep learning approach has a true positive rate (14.29% vs. 8.70%) that is twice as high, and it also achieves a much lower false positive rate (0.63% vs. 11.68%). We also observe that these evaluators are susceptible to human bias, as they are more likely to label fake reviews as genuine than they are to label genuine reviews as genuine. Our study offers further explanations for the advantages of deep learning and is the first to construct a deep learning model to detect fraudulent online reviews, an approach that can help curb fake reviews and increase information quality and market efficiency.