Socially-Attentive Representation Learning for Cold-Start Fraud Review Detection

Abstract

Fraud reviews consist one of the most serious issues in cyberspace, which dramatically damage users’ decisions yet have great challenges to be detected. Accordingly, effectively detecting fraud reviews is becoming a critical task for cybersecurity. Although various efforts have been put on fraud review detection, they may fail in the case of cold-start where a review is posted by a new user who just pops up on social media. The main reason lies in lacking sufficient historical information of the new user. Recently, limited research has been conducted on fraud review detection with the cold-start problem, in which, however, advanced methods either ignore complex collaborative review manipulations or overlook fraud-related characteristics. As a result, they may easily be deceived by camouflage fraudsters and have low detection precision. This paper presents a novel socially-attentive user representation learning method for fraud review detection with the cold-start problem, namely SATURN, which leverages the fraud-related user reviewing behavior with comprehensive user social couplings for cold-start fraud review detection. SATURN jointly embeds user-item-attitude-review entities relations, explicit and implicit hierarchical social couplings, and fraud-related information into a user vector representation space, in which the fraud-related representation of a new user can be effectively inferred according to the learned socially-attentive entities relation. Subsequently, SATURN effectively detects fraud reviews with the cold-start problem in its learned representation space. Extensive experiments on four large real-world data sets demonstrate SATURN significantly outperforms three state-of-the-art and two baseline competitors in terms of both general and cold-start fraud review detection tasks.