<p>In recent years, the rapid expansion of the Internet has significantly lead to the increase in number as well as size of online social networks (OSNs) which connect millions of users across the world. While OSNs offer various benefits—such as enhancing communication, social interaction, and information dissemination—they have also become vulnerable to malicious activities in OSNs. Among these threats, the widespread distribution of fake or misleading content/cyberbullying has raised serious societal concerns. Recent studies reveal that such harmful content is often propagated by automated accounts or social bots, which are designed to mimic human behavior as well as manipulate public opinion at scale. As a result, effective detection of social bots has emerged as a critical task in ensuring the security and trustworthiness of social media platforms. Despite progress in this domain, most existing bot detection techniques rely heavily on supervised learning techniques that require large volumes of annotated networked data. These models are often tailored to specific platforms, such as Twitter; as a result, highly depending on manually engineered features which are derived from user profiles and behavioral patterns. Consequently, these graph learning techniques still faced limitations in terms of generalization and adaptability when applied to other OSNs with different structural characteristics. To overcome these challenges, we propose a novel model called AEG4BD, which integrates a pre-trained auto-encoder with a graph neural network (GNN) to facilitate robust and platform-independent social bot detection. Our proposed AEG4BD model can leverage multi-level representation learning to capture both local user attributes as well as global relational structures. Extensive experiments conducted on real-world social bot datasets demonstrate that AEG4BD outperforms existing baselines; as a result, offering improved accuracy, flexibility, as well as generalization across various social network environments.</p>