Robust fake review detection using uncertainty-aware LSTM and BERT

Abstract

In a web-based world driven by e-commerce, customers are quick to turn to online shopping services. However, the products available for purchase cannot be personally inspected so buyers turn to online product reviews. Consumers trust these reviews and are likely to spend more at stores with good evaluations. Sellers are well aware of this phenomenon and are not averse to using unethical methods to boost the reputation of their own products, or demerit the products of their rivals. In other words, fake reviews are quite rampant and are responsible for heavily affecting consumer purchasing decisions and business profits. In response, many fake review detection models have been extensively explored in the last decade. However, there is still a lack of robustness in these approaches. Our research addresses this gap in the field of fake review detection. We deploy the BERT and LSTM models coupled with the Monte Carlo Dropout (MCD) technique, on the Yelp Labelled Dataset comprising 10,000 hotel reviews from North America. MCD provides a representation of uncertainty by randomly dropping neurons in multiple predictions of the network. This gives us an approximation of the uncertainty. Since, fake review detection is a risky task we employ MCD to make our system robust and reliable. Our study yields an accuracy of 91.75% using the MCD-embedded BERT model. It outperforms the LSTM model overall.