A Self-Supervised Adaptive Tuning Approach for Out-of-Distribution Social Media Bot Detection

With the rapid advancement of generative AI and deepfake technologies, social media bots have become increasingly sophisticated, engaging in spreading misinformation. Detecting these bots early and effectively is critical for enhancing users' online experience. Existing approaches primarily rely on fixed model parameters obtained during training, which often struggle with the Out-of-Distribution (OOD) problem caused by discrepancies between training and test data. To overcome the above challenges, this paper introduces a novel adaptive feature tuning approach that incorporates a test-time fine-tuning mechanism. By leveraging self-supervised contrastive learning, this approach dynamically adjusts model parameters, enabling the extraction of more robust feature representations and better adaptation to unseen data distributions. Extensive experiments on commonly used baseline models demonstrate the effectiveness of the proposed method.