Back to publications

arXiv preprint · 2025

E-CaTCH: Event-Centric Cross-Modal Attention with Temporal Consistency and Class-Imbalance Handling for Misinformation Detection

View resource →

  • misinformation
  • multimodal
  • attention
  • temporal modeling

Summary

E-CaTCH is an interpretable, scalable framework for multimodal misinformation detection built on a simple insight: misinformation usually spreads through clusters of related posts (pseudo-events) rather than isolated items.

The model clusters posts into pseudo-events by textual similarity and temporal proximity, then processes each event on its own. Within an event, text and image features (from BERT and ResNet) are refined with intra-modal self-attention and aligned through bidirectional cross-modal attention, fused by a soft gating mechanism. A trend-aware LSTM, enhanced with semantic-shift and momentum signals over overlapping time windows, models how the narrative evolves. Classification happens at the event level, and the loss combines adaptive class weighting, temporal-consistency regularization, and hard-example mining to handle severe class imbalance.

On Fakeddit, IND, and COVID-19 MISINFOGRAPH, E-CaTCH consistently beats state-of-the-art baselines, with cross-dataset tests showing strong robustness and generalizability.

Authors

A. Mousavi, Y. Abdollahinejad, R. Corizzo, N. Japkowicz, Z. Boukouvalas

Links