Behavior-centric deep learning framework for detecting fake or harmful content using user interaction patterns and side information

The rapid spread of deceptive and harmful content on social media has created a need for detection models that can operate effectively under sparse and limited user activity conditions. Many existing detection approaches depend on content features, item-level representations, or dense interaction histories, which may be unavailable, incomplete, or unreliable in early-stage online environments. To address this limitation, this study proposes a behavior-centric deep learning framework for identifying harmful user activity using temporal interaction patterns and contextual side information. The proposed architecture, GRU-Metadata-Contrast-Attention-Fusion (GRU-MCAF) employs a Bidirectional Gated Recurrent Unit (Bi-GRU) encoder to capture sequential user behavior and a dense projection layer to represent user metadata in a compatible latent space. These representations are integrated through a contrast-guided attention mechanism that highlights behavior–metadata deviations useful for distinguishing harmful activity from normal interaction patterns. The proposed model is evaluated using the FakeNewsNet dataset, which provides user profiles and engagement logs associated with verified misinformation labels. Experimental results show that the proposed framework achieved a precision of 95.13%, recall of 94.68%, accuracy of 94.82%, and F1-score of 94.90%. It also obtained an MCC of 0.901, AUC of 0.969, and false positive rate of 4.16%. Comparative evaluation against CNN, GRU, Bi-LSTM, EfficientNet-B0, and attention-based LSTM models indicates that the proposed behavior-centric fusion strategy provides improved detection performance under the considered experimental setting.