Unsupervised Cross-Domain Rumor Detection with Contrastive Learning and Cross-Attention

Hongyan Ran, Caiyan JiaAAAI 2023 · DOI

Abstract

Detecting rumors on social media is a critical task, yet models often fail when encountering new, breaking news topics due to "domain shift." This paper presents UCD-RD, an unsupervised framework designed to bridge this gap without requiring labeled data from the target domain. By leveraging contrastive learning to align source and target representations and a cross-attention mechanism to extract domain-invariant features, UCD-RD achieves state-of-the-art performance across diverse social media datasets.

Introduction

Rumor detection models typically excel within the specific domains they were trained on. For instance, a model trained on data from a terrorist attack learns the specific vocabulary and linguistic markers of that event. However, when a new crisis—such as a global pandemic—emerges, these models often fail because the language and context have shifted significantly. This "domain shift" problem is a major hurdle for real-time misinformation mitigation.

In practice, manually labeling thousands of tweets for every new trending topic is unfeasible. There is a pressing need for Unsupervised Cross-Domain (UCD) detection: the ability to adapt a model trained on a labeled "Source" domain to an unlabeled "Target" domain. The core challenge lies in identifying features that remain constant across domains—such as skeptical crowd responses—while ignoring domain-specific jargon.

Conceptual overview of domain shift and the proposed matching strategies.

The UCD-RD Framework

The proposed Unsupervised Cross-Domain Rumor Detection (UCD-RD) model addresses the domain gap through four integrated modules:

  1. Rumor Representation Module (RRM): Rather than analyzing isolated tweets, UCD-RD processes "propagation trees"—the structural evolution of a rumor and its associated replies. Multi-head Attention (MHA) is employed to capture long-range dependencies within these conversation paths.
  2. Contrastive Learning Module (CLM): This module performs instance-wise and prototype-wise contrastive learning. It pulls representations of the same class (e.g., "Rumor") closer together while pushing different classes ("Non-Rumor") apart, regardless of their domain of origin.
  3. Pseudo-Labeling: To handle the lack of labels in the target domain, the model utilizes K-means clustering. Crucially, these clusters are initialized using "prototypes" (centroids) from the source domain, ensuring that target-domain clusters align with known categories.
  4. Cross-Attention Module (CAM): By pairing source samples with target samples sharing the same (pseudo) label, the CAM identifies shared, domain-invariant patterns. This forces the model to focus on universal indicators of rumors, such as patterns of denial or questioning.
Detailed architecture of the UCD-RD framework.

Experimental Evaluation

The researchers evaluated UCD-RD across four major cross-domain scenarios:

  • PHEME: Transitioning between 'Terrorist' and 'Gossip' events.
  • Twitter to Twitter-COVID19: Moving from historical rumors to pandemic-specific misinformation.
  • Twitter15 to Twitter16: Testing across different years of social media data.
  • Weibo to Weibo-COVID19: Evaluating cross-domain performance on Chinese social media.

Key Results

UCD-RD significantly outperformed existing state-of-the-art methods across all test groups. On the Terrorist → Gossip task, the model achieved an accuracy of 84.88%, representing a 14.69% improvement over the best baseline (PPA-WAE). Furthermore, F1-scores improved by 10-20% compared to traditional supervised models that lacked domain adaptation. Ablation studies confirmed that both the contrastive learning and cross-attention components are essential for maintaining high performance.

Conclusion

UCD-RD demonstrates that focusing on propagation patterns and aligning domain representations through contrastive learning can effectively overcome the domain shift problem in rumor detection. While the model currently relies on K-means for pseudo-labeling—which can introduce noise—it represents a significant step toward robust, real-time misinformation detection. Future work will explore incorporating multi-source information, such as user metadata and external knowledge graphs, to further enhance detection capabilities.

Summary updated

BibTeX
@inproceedings{ran2023unsupervised,
  author = {Hongyan Ran and Caiyan Jia},
  title = {{Unsupervised Cross-Domain Rumor Detection with Contrastive Learning and Cross-Attention}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.1609/aaai.v37i11.26584},
}
Figure