Paper Abstract

Abstract (Draft)

In recent years, it has been shown that falsification of online reviews can have a large, quantifiable effect on the success of the subject. This creates a large enticement for sellers to participate in review deception to boost their own SEO score or hinder the competition.

Most current efforts to detect review deception are based on supervised classifiers trained on syntactic and lexical patterns. However, recent neural classifiers have been shown to match or outperform SOTA methods. Generative models have demonstrated an ability to learn these underlying patterns in an unsupervised fashion. Thus, we propose DeceptGAN, a novel classification approach based on Generative Adversarial Networks.

Standard GAN models involve a single generator and discriminator network, however, it has been shown that two discriminator networks can solve a common issue GAN issue known as mode collapse. DeceptGAN will therefore be based on this same architecture. The specifics of our architecture are still a WIP, however it looks as though our generator will be modelled as a stochastic policy agent in reinforcement learning (RL), and our discriminators will use a search algorithm to pass action values as a RL reward to the generator.