Paper Related Work

Added after initial draft

Collective Opinion Spam Detection: Bridging Review Networks and Metadata

behavioral features are superior to text features.

---

The initial study of opinion spam and trustworthiness of reviews branched from existing research on classifying the opinion of reviews [1], this study invents a logistic regression model that is capable of detecting certain types of fake reviews effectively. Large amounts of work have been done since using statistical, supervised learning [2, 3] showing highly accurate results in-domain on certain datasets. Recently studies using neural modelling have been created that claim to improve on baseline methods [4, 5]. A study has recently been shown that GAN models can be made stable using two discriminators [6], and also that they can rival SOTA results at 89.1% accuracy. This was the first paper to study GANs for opinion spam detection, and considering the highly accurate results it indicates further studies would be worthwhile.

1. Jindal and Liu, 2008.
2. Hernández-Castañeda et al, 2016: Hernández-Castañeda et al, 2016: 'Cross-domain deception detection using support vector networks'
3. Kumar et al, 2018: 'Detecting Review Manipulation on Online Platforms with Hierarchical Supervised Learning'
4. You et al, 2018: 'An Attribute Enhanced Domain Adaptive Model for Cold-Start Spam Review Detection'
5. Wang et al, 2017: 'Fake Review Detection on Yelp'
6. Aghakhani et al, 2018: 'Detecting Deceptive Reviews using Generative Adversarial Networks'