Bridging Claims and News for Cross-Dataset Fake News Detection: Hierarchical Veracity Learning and Evidence-Guided Expert Routing

Abstract

Cross-dataset fake news detection remains challenging because claim-level fact verification and news-level misinformation recognition differ substantially in input granularity, contextual richness, and label structure. This study proposes C2N-HIVE, a unified framework that combines hierarchical veracity learning, split-safe dense evidence retrieval, communication-pattern graph modeling, pre-routing uncertainty estimation, and evidence-routed memory-based expert specialization. The term “intent” in the model name denotes a latent, model-derived communication-pattern factor rather than verified author intent. Rather than collapsing heterogeneous data into a flat binary representation, C2N-HIVE retains LIAR’s six-level veracity structure as auxiliary supervision and transfers the learned hierarchy to article-level detection. Under the leakage-controlled protocol, the model achieved 83.2% Macro-F1 on LIAR under the strict mapping, 81.0% under the inclusive mapping, 86.4% on FakeNewsNet-PolitiFact, and 87.0% on FakeNewsNet-GossipCop. In zero-shot transfer restricted to LIAR and GossipCop, it obtained 74.7% Macro-F1 for LIAR-to-GossipCop and 69.2% for GossipCop-to-LIAR-inclusive under a shared-only path with a fixed external evidence corpus, while retaining an advantage under low-resource adaptation. Temporal, publisher-held-out, ablation, controlled evidence-index degradation, ordinal-freeze, routing-utilization, sensitivity, perturbation, masking, missing-evidence, calibration, and human-rating analyses further support the evaluated contributions of hierarchical supervision, structured evidence integration, communication-pattern modeling, and adaptive routing. The findings position C2N-HIVE as an evidence-supported screening framework, while also identifying evidence coverage, mapping choices, benchmark structural homogeneity, human-evaluation sample size, and domain-specific knowledge as important limitations.