Detecting Spammers in E-Commerce Website via Spectrum Features of User Relation Graph
Bo Guo, Hao Wang, Zhaojun Yu, Yu Tao Sun · 2017
The review system promoted the prosperity of the E-commerce market, but it has also been improperly exploited by spammers who found it profitable to write fake reviews to mislead the consumers. Detecting spams accurately has been a challenging problem, since spammers alter their writing style and imitate normal users. The majority of the existing work tried to solve this problem by designing an indicator system using behavioral features, which often suffers from the lack of large-scale labeled datasets. This paper proposes a novel user relation graph model based on a bipartite graph built directly from the review data and introduces two novel algorithms called Finding Abnormal Dimensions by Kurtosis function (FADK) and Finding Abnormal Dimensions by Shapiro-Wilk test (FADSW) to find small groups of spammers in the large user-relation graph. FADK focus on each eigenvector and its neighborhood, which uses the kurtosis function as a crucial measure. In FADSW, the hypothesis test theory is introduced to solve this problem for the first time. This combination is a novel interdisciplinary research methodology. To get an unbiased performance evaluation, we evaluated our two algorithms using two different real-world datasets: 1) Amazon dataset from the US; and 2) JD.com dataset from China. While both algorithms showed relatively high distinguishing power, FADSW outperforms FADK on both the JD and the Amazon datasets.