HGR+++: Metadata‐Conditioned Hierarchical Gated Residual Fusion for Fine‐Grained Multimodal Fake News Detection

The rapid proliferation of misinformation on social media poses significant societal challenges, while many existing detection approaches focus on binary classification or use relatively static multimodal fusion strategies. In practice, misinformation can appear in multiple forms, motivating fine‐grained modeling of textual, visual, and contextual information. To address this challenge, we propose Hierarchical Gated Residual Fusion (HGR+++), a multimodal framework for fine‐grained fake news detection under the six‐class Fakeddit setting. The proposed approach integrates semantically aligned text–image representations from pretrained CLIP encoders with structurally encoded metadata capturing engagement and source‐level context. A metadata‐conditioned gated residual fusion mechanism performs featurewise adaptive recalibration of the fused multimodal representation. Experiments on the Fakeddit benchmark show that HGR+++ achieves an average accuracy of 96.6% and macro‐F1 of 94.9% across five random seeds, with higher mean performance than the evaluated internal baselines. Paired statistical analysis against the CLIP + metadata baseline and ablation experiments further support the contribution of the proposed fusion mechanism and structural metadata. These findings demonstrate the potential of metadata‐conditioned gated multimodal fusion for fine‐grained misinformation detection while highlighting the importance of contextual information within the evaluated benchmark.