Detecting Fake Yelp Reviews Using a Positional LSTM / K-L Divergence Ensemble Approach

Abstract

Online reviews of products and services often have a significant impact on future purchases. Unfortunately, this invites opportunities for fraud –to provide fake reviews to improve a company’s reputation in the eyes of consumers or to disparage the reputation of a competitor. We combine two methods to detect fraud in Yelp restaurant reviews. First, we develop and apply a bi-directional Long-Short Term Memory (LSTM), a type of recurrent neural network, to take advantage of the positional relevancy of comments within reviews. LSTMs are well-suited to examine linguistic features in text and evaluating different regions of text within the review enhances our model’s accuracy. To this component, we apply a Kullback-Leibler (K-L) divergence technique to examine the discrepancy in the term rankings between real and fake Yelp reviews. This ensemble is used to discriminate fake reviews from real ones, achieving an Average Precision (AP) of 0.5402 and an Area Under the Curve (AOC) of 0.866. which is an improvement over other state-of-the-art techniques on the same YelpNYC dataset.