Detecting Review Spammer Groups via Bipartite Graph Projection

Online product reviews play an important role in E-commerce websites because most customers read and rely on them when making purchases. For the sake of profit or reputation, review spammers deliberately write fake reviews to promote or demote target products, some even fraudulently work in groups to try and control the sentiment about a product. To detect such spammer groups, previous work exploits frequent itemset mining (FIM) to generate candidate spammer groups, which can only find tightly coupled groups, i.e. each reviewer in the group reviews every target product. In this paper, we present the loose spammer group detection problem, i.e. each group member is not required to review every target product. We solve this problem using bipartite graph projection. We propose a set of group spam indicators to measure the spamicity of a loose spammer group, and design a novel algorithm to identify highly suspicious loose spammer groups in a divide and conquer manner. Experimental results show that our method not only can find loose spammer groups with high precision and recall, but also can generate more meaningful candidate spammer groups than FIM, thus it can also be used as an alternative preprocessing tool for existing FIM-based approaches.