Enhancing Social Media Bot Detection with Cross-Feature Gating and Residual Learning

The growing presence of malicious bot accounts on social media poses a threat to the authenticity of online communities, as they amplify misinformation, spread spam, and manipulate engagement. Reliable detection of these accounts is therefore essential to protect the integrity of platforms such as Instagram. This study introduces a deep learning–based detection framework built on the CrossGatedTabular (CGT) architecture, designed to learn complex patterns in user activity. To strengthen evaluation, two publicly available datasets of Instagram accounts were merged into a comprehensive benchmark representing diverse user behaviors. Natural language processing (NLP) was applied to refine textual content and metadata, enhancing the quality of feature representation. For classification, a CrossGatedTabular neural architecture is employed, which integrates cross-feature interactions, gated multilayer perceptron (MLP) layers, and regularization mechanisms to effectively capture complex patterns within the dataset. Experimental results demonstrate that the proposed approach achieves 0.9340 accuracy along with high precision, recall, and F1-scores, consistently outperforming baseline classifiers. These findings highlight the effectiveness of deep neural architectures in malicious account detection and provide a scalable solution for enhancing the trustworthiness of social media platforms.