Multi-input BiLSTM deep learning model for social bot detection

Abstract

The recent emergence of social bot detection tech-niques on social media has lately garnered immense attention. These fake automated accounts can post content and interact with other accounts as if they were hosted by a real person. In fact, automation in the wrong hands is a threat, opening up the opportunity to some malicious users and manipulators to spread fake news and misleading information. Various approaches and techniques are used for bot detection making a diversity of choices for relevant feature selection. Therefore, exploiting the accounts auxiliary information and textual features is challenging of itself because their combination produce incomplete, unstructured, and noisy data. This research offers a new architecture that incorporates multiple inputs based on the tweet content and the user metadata merged then fed into a Bidirectional Long-Short-Term Memory (BiLSTM) network. We obtain very satisfactory results as regard to performance metrics (over 97% for accuracy, precision, fl-score, 98% for recall and 99% of ROC/AUC). Experiments with real-world data reveals that it is complex to identify the impact of each feature in bot detection problem and gives accurate detection results.