Adaptive deep Q-learning model for detecting social bots and influential users in online social networks

Abstract

In an online social network (like Twitter), a botmaster (i.e., leader among a group of social bots) establishes a social relationship among legitimate participants to reduce the probability of social bot detection. Social bots generate fake tweets and spread malicious information by manipulating the public opinion. Therefore, the detection of social bots in an online social network is an important task. In this paper, we consider social attributes, such as tweet-based attributes, user profile-based attributes and social graph-based attributes for detecting the social bots among legitimate participants. We design a deep Q-network architecture by incorporating a Deep Q-Learning (DQL) model using the social attributes in the Twitter network for detection of social bots based on updating Q-value function (i.e., state-action value function). We consider each social attribute of a user as a state and the learning agent’s movement from one state to another state is considered as an action. For Q-value function, we consider all the state-action pairs in order to construct the state transition probability values between the state-action pairs. In the proposed DQL algorithm, the learning agent chooses a specific learning action with an optimal Q-value in each state for social bot detection. Further, we also propose an approach that identifies the most influential users (which are influenced by the social bots) based on tweets and the users’ interactions. The experimentation using the datasets collected from Twitter network illustrates the efficacy of proposed model.