Analysis of a Partly Supervised Learning Model for the Identification of Spammer Classes

Abstract

Online product ratings play a key part in customer buying decisions. A high proportion of favorable feedback can carry significant revenue increase, whilst poor feedback can result in a reduction in revenue. Driven by large financial gain, many spammers seek to advertise their goods or demote their rivals' goods by publishing false and negative web reviews. Via a variety of accounts and activities, multiple spammers can be mobilised as a spammer at crowdsourcing sites communities to collectively exploit user ratings which may be more harmful. Current research on spammer group identification removes spammer community members from analysis data and distinguishes actual spammer groups using unsupervised spam rating methods. Moreover, according to previous research, it is easier than expected to mark a small number of spammers but few methods attempt to use these useful branded findings effectively. In this paper, a partially controlled spammer class recognition learning model (PSGD) is proposed. In particular, in terms of favourable circumstances and individual characteristics  have the clear negative selection. By mixing positive instances, negative cases and unknown cases, they are transforming the PU problem into a well-known semi-supervised learning Problems and trains a spammer group recognition classification with an EM algorithm and the Naive Bayesian model. Real-life experiments on Amazon.cn data sets show that the proposed PSGD is efficient and exceeds state-of-the-art community spammer detection procedures.