Towards Generating Spam Queries for Retrieving Spam Accounts in Large-Scale Twitter Data

Abstract

Twitter, as a top microblogging site, has became a valuable source of up-to-date and real-time information for a wide range of social-based researches and applications. Intuitively, the main factor of having an acceptable performance in those recherches and applications is the working and relying on information having an adequate quality. However, given the painful truth that Twitter has turned out a fertile environment for publishing noisy information in different forms. Consequently, maintaining the condition of high quality is a serious challenge, requiring great efforts from Twitter’s administrators and researchers to address the information quality issues. Social spam is a common type of the noisy information, which is created and circulated by ill-intentioned users, so-called social spammers. More precisely, they misuse all possible services provided by Twitter to propagate their spam content, leading to have a large information pollution flowing in Twitter’s network. As Twitter’s anti-spam mechanism is not both effective and immune towards the spam problem, enormous recherches have been dedicated to develop methods that detect and filter out spam accounts and tweets. However, these methods are not scalable when handling large-scale Twitter data. Indeed, as a mandatory step, the need for an additional information from Twitter’s servers, limited to a few number of requests per 15 min time window, is the main barrier for making these methods too effective, requiring months to handle large-scale Twitter data. Instead of inspecting every account existing in a given large-scale Twitter data in a sequential or randomly fashion, in this paper, we explore the applicability of information retrieval (IR) concept to retrieve a sub-set of accounts having high probability of being spam ones. Specifically, we introduce a design of an unsupervised method that partially processes a large-scale of tweets to generate spam queries related to account’s attributes. Then, the spam queries are issued to retrieve and rank the highly potential spam accounts existing in the given large-scale Twitter accounts. Our experimental evaluation shows the efficiency of generating spam queries from different attributes to retrieve spam accounts in terms of precision, recall, and normalized discounted cumulative gain at different ranks.