Group-Author Model for Latent Social Astroturfers Group Detection

摘要

Online reviews play vital role in influencing people while making decisions in various sectors. However, in the wake of recent news, it is understood that astroturfing reviews or fake reviews caused unpleasant manipulations in the decisions of people. In this backdrop, astroturfing detection has attracted both industry and academia. Even in the presence of prevailing astroturfing, most of the review sites do not provide effective filters for astroturfing reviews. Unfiltered reviews can be deceptive and may promote or discredit certain product or service. Group of individuals who are intentionally spreading opinion either positively or negatively on a chosen service or product in specific period and in an organized fashion is known as astroturfing group. The existing literature focused on topic models and author models using Latent Dirichlet Allocation (LDA). However, there is no model which reflects both author groups that stay latent in the opinion manipulation with astroturfing reviews. In this paper we proposed and implemented a Group-Author model which is based on modified LDA with underlying unsupervised learning method known as clustering technique for text mining. We proposed a variant of LDA which is known as Group-Author model which takes two global parameters pertaining to corpus. They are known as author distributions in corpus α and astroturfing group author distributions in corpus β. We defined two algorithms namely Latent Astroturfer Group Detection (LAGD) and Temporal Filtering Algorithm (TFA) for discovering social astroturfing groups which are latent and validating such groups respectively. We built a prototype application to demonstrate proof of the concept. The empirical results revealed the utility of the proposed model in terms of discovering astroturfing groups besides reducing time and space complexity.