Anomaly Detection in Social Networks

Abstract

Social Networks provides a mode of information sharing forindividuals from anywhere in the world. Diverse domains, such as, entertainment, education, business, and medical are using it. Excessive utilization of social media may cause various types of illegal activities like spams, rumor spreading. In order to avoid these activities, it is necessary to detect them. Two types of data can be extracted from social networks: the behavioral data that represents the users' activities on the social media platform; and the structural data that represents the topological aspects of the networks. The useful information contained by data can be studied by graph theory for anomaly detection task. Existing structural features like degree, edge count, average betweenness centrality and brokerage lack in terms of accuracy and produce false positive and false negative. To reduce the amount of false positive and false negative we have used a new graph metric closeness centrality for anomaly detection task andapplied the proposed method on three datasets collected from a social networkrepository. The comparability and efficiency of different methods have been studied using statistical quantities like F-score, precision, andrecall. Experimental results show that the relationship between closeness centrality of an egonet and number of edges in egonet is capable to catch most of the anomalies and have higher F-score value than degree, average betweenness centrality and brokerage.