Skip to content

Transfer learning for fake news detection

Transfer learning in fake news detection addresses the challenge that models trained on one set of events often fail to generalize to newly emerged events. The core problem is that deep learning models tend to learn event-specific features (e.g., keywords, image patterns) that are tied to the training events and do not transfer to unseen events.

Approaches to transfer learning in this domain include:

  • Event-invariant feature learning: Learn representations that are shared across events, removing event-specific information while preserving discriminative signals for fake detection. This is the approach taken by EANN via adversarial learning.
  • Domain adaptation: Adapt models trained on one source domain (e.g., events with abundant labeled data) to a target domain (e.g., new events with little or no labeled data) using techniques like feature alignment or adversarial training.
  • Few-shot / zero-shot learning: Train models that can detect fake news on new events with minimal or no labeled examples from those events.

Key papers

  • Nan et al. (2021) — MDFEND: Addresses multi-domain transfer via mixture-of-experts with domain gating; learns domain-agnostic base features and domain-specific expert representations; gate adaptively weights experts using domain and content signals; introduces Weibo21 dataset spanning 9 domains; F₁ 0.9137 outperforms EANN (0.8975) and other multi-domain baselines.

  • Wang et al. (2018) — EANN: First to explicitly formulate fake news detection on newly emerged, time-critical events as a transfer learning problem. Proposes using an event discriminator with adversarial learning to remove event-specific features. Demonstrates that models trained on known events can transfer to unseen events when using event-invariant features.

  • A Heuristic-driven Uncertainty based Ensemble Framework for Fake News Detection in Tweets and News Articles: Demonstrates transfer learning effectiveness via fine-tuning of multiple pre-trained language models (BERT, RoBERTa, XLNet, DeBERTa, ERNIE 2.0, ELECTRA) on fake news detection tasks; shows that pre-trained models adapt well to both social media tweets and news articles with minimal task-specific training.

Connections