Semi-SGD: Semi-supervised Learning based Spammer Group Detection in Product Reviews

The purchase decision of customers in e-commerce platforms is strongly influenced by product ratings and reviews. Driven by the profits, review spammers post fake reviews to promote their products or demote their competitors' products. Differ from individual spammers, the spammer groups manipulate reviews together and can be more damaging. Existing work for spammer group detection extract candidate groups from review data and identify the spammer groups using unsupervised spamicity ranking methods. However, the labeled and unlabeled data are existing simultaneously in practice and no method makes good use of both these data in spammer group detection. In this paper, we propose a semi-supervised learning based spammer group detection method (Semi-SGD), which trains a Naive Bayes classifier on a small set of labeled data as an initial classifier, and then incorporates unlabeled data with Expectation Maximization (EM) algorithm to improve the initial classifier iteratively. Experiments on Amazon.cn datasets show that our proposed Semi-SGD is efficient and effective.