Domain adaptation

Domain adaptation refers to machine learning techniques that adjust models trained in one domain to perform well in a different domain, typically when labeled data is scarce or unavailable in the target domain. In fake news detection, domain adaptation enables models to handle shifts in topic, source, platform, language, or temporal distribution without full retraining.

Approaches

• Adversarial domain adaptation: Learn domain-invariant feature representations via minimax training, forcing a feature extractor to fool a domain classifier. Example: EANN.
• Domain discovery: Unsupervised identification of distinct domains (e.g., clusters of topically similar records) in unlabeled data, enabling targeted training. Example: Silva et al. use propagation network community structure to discover domains.
• Instance weighting: Reweight labeled source-domain data to match the target domain's distribution, reducing distribution shift penalties.
• Fine-tuning: Pretrain on source domain(s), then fine-tune on small labeled target-domain sample.

Key papers

• Wang et al. (2018) — EANN: Foundational work formulating fake news detection as domain adaptation; adversarial training to learn event-invariant features; achieves 71.5% accuracy on Twitter.

• Silva et al. (2021): Extends adversarial approaches with explicit preservation of domain-specific knowledge (via domain-specific decoder) and cross-domain knowledge (via shared decoder); unsupervised domain discovery eliminates need for manual domain labels.

Connections

• Cross-domain learning for fake news — Domain adaptation is a strategy for achieving cross-domain transfer.
• Multimodal fake news detection — Some multimodal methods incorporate domain adaptation (e.g., EANN).
• Fake news detection methods — Domain adaptation is a technique within the broader detection toolkit.