Collective Classification for Social Opinion Spam Detection

With increasingly more firms using online social media to market their products and services, so are the widely spread attacks to the consumer opinions posted to social media, namely the social opinion spam. Fake social opinions may inflate firms' own product reputation or defame competitors' product reputation. Accordingly, there is a pressing need to develop effective detection method to identify and remove social opinion spam in order to facilitate fair online trading and improve the effectiveness of consumer decision-making. The main contribution of our research work is to design and evaluate several collective classification methods to detect social opinion spam on online social media. In particular, experiments based on the Yelp social opinion dataset reveal that state-of-the-art collective classification algorithms can achieve the detection performance of 72.5% in terms of F-score. However, selecting the effective relational features is critical for achieving good performance of collective classification. Moreover, our experiments also show that more recent deep learning techniques such as DeepWalk can be incorporated into the iterative collective classification to further bootstrap the performance of social opinion spam detection. More studies are needed to further examine deep learning techniques and investigate into why they cannot bootstrap social opinion spam detection. As a whole, our work contributes to develop a new generation social opinion spam detection methodology to improve the hygiene of social opinions and facilitate fair online trading and effective consumer decision-making.