Friendship Preference: Scalable and Robust Category of Features for Social Bot Detection

Social bots are intelligent programs that control the behavior of fake accounts in an online social network(OSN). They pass themselves off as human accounts and manipulate the health of the ecosystem of OSNs. Thus, it is crucial to distinguish social bots from human accounts. Despite recent advances in social bot detection, state-of-the-art methods face challenges in scalability, generalization, and robustness. Considering these drawbacks, in this article, a new category of features, called friendship preference, is proposed. Friendship preference features are extracted from the profile attributes of the followers. The proposed feature extraction formula is designed to be scalable as the number of followers grows. Two classifiers are developed to evaluate the proposed category of features in terms of efficiency and usefulness. The classifiers are benchmarked on several real datasets. The experimental results indicate that the classifiers outperform Botometer in terms of classification performance. Scalability is assessed by inspecting the detection performance of classifiers when the attributes of a limited number of the accounts in the neighborhood of the users with large egonets are available. Furthermore, generalization is validated by crossover validation on various datasets. Finally, robustness and the ability for early detection of social bots are discussed.