Early detection effectiveness of fake news/information disorder classifiers using target-based emotions and stances on evolving social networks

Early detection of information disorder – often referred to as fake news – is crucial for mitigating its societal impact before harmful narratives fully diffuse. This study evaluates whether target-based emotion–stance analysis (TESA), grounded in social-psychological theories of persuasion, improves early information disorder detection when combined with graph-based propagation features. Using FibVID and TruthSeeker, we modelled claim-level cascades as temporally evolving graphs and compared four classifier conditions: Bidirectional Encoder Representations from Transformers (BERT), TESA + BERT, graph neural networks (GNNs) with BERT embeddings, and TESA + GNN with BERT embeddings. TESA + GNN achieved the strongest peak-weighted F1 across both datasets, reaching 0.757 in FibVID at 20 min and 0.837 in TruthSeeker on the full graph. Averaged across time slices, TESA + GNN also outperformed the GNN baseline in FibVID (0.712 vs 0.707) and TruthSeeker (0.795 vs 0.783). Beyond classification, TESA features revealed meaningful diffusion patterns: neutral emotion and stance were associated with broader, slower non-IDO cascades, whereas fear uniquely accelerated IDO cascade growth. These findings show that target-specific emotion and stance provide interpretable, persuasion-relevant signals for early IDO detection and can support real-time triage, alerting, and mitigation systems.