Abstract
The proliferation of social media has significantly increased the threat of fake news, necessitating the development of efficient and timely rumor detection methods. While Graph Neural Networks (GNNs) have become the standard for modeling post-propagation patterns, they suffer from a lack of interpretability and a susceptibility to noise within the propagation graph. This paper introduces the Diverse Counterfactual Evidence framework for Rumor Detection (DCE-RD). The framework identifies multiple, distinct subgraphs within an event graph that serve as counterfactual evidence—components whose removal would fundamentally alter the model's prediction. By aggregating these diverse "views," DCE-RD achieves both robust detection and human-interpretable explanations. Experimental results on the Weibo and MC-Fake datasets demonstrate that DCE-RD outperforms existing state-of-the-art models, particularly in early detection and noisy environments.
Introduction
As social media platforms become primary news sources, the speed at which rumors spread has outpaced traditional fact-checking. Automated rumor detection typically relies on GNNs to analyze how a story propagates through retweets and replies. However, these models are often viewed as "black boxes," providing little insight into why a specific post was flagged. Furthermore, propagation graphs are inherently noisy, containing irrelevant interactions or bot-driven activity that can lead to spurious correlations.
To address these challenges, we propose DCE-RD. Our approach moves beyond single-view explanations by seeking a diverse set of counterfactual subgraphs. This multi-view strategy ensures that the model's decision is based on several independent pieces of evidence, making it more resilient to noise and more transparent to human moderators.
The DCE-RD Framework
The DCE-RD framework operates on the principle that a rumor is best understood through the lens of what is essential to its classification. It consists of four primary components:
1. Subgraph Generation
To identify critical propagation patterns, the model employs a Top-K Node Sampling strategy. By utilizing the Gumbel-max trick, the framework can select the most influential nodes in a differentiable manner. This allows the subgraph generation process to be trained end-to-end alongside the detection model.
2. Counterfactual Evidence Identification
A subgraph is defined as "counterfactual evidence" if its removal from the original propagation graph causes a maximum drop in the model's confidence in its original prediction. Essentially, the model identifies the smallest set of interactions that, if absent, would flip the classification from "rumor" to "non-rumor" (or vice versa).
3. Diversity via Determinantal Point Processes (DPP)
To ensure the model does not simply find redundant versions of the same explanation, we implement a diversity loss inspired by Determinantal Point Processes (DPP). This mathematical constraint encourages the selection of subgraphs that are distinct in the feature space, capturing different facets of the rumor's spread—such as a cluster of skeptical replies in one view and a suspicious burst of bot-like retweets in another.
4. Robust Aggregation
Rather than making a judgment based on the entire noisy graph, DCE-RD aggregates the insights from these diverse counterfactual subgraphs. This filtering process naturally ignores task-irrelevant nodes, leading to a more stable and accurate final veracity judgment.
Experimental Evaluation
Datasets
We evaluated DCE-RD on two large-scale datasets:
- Weibo: A benchmark dataset from the Chinese social media platform containing 4,664 events.
- MC-Fake: A comprehensive Twitter dataset featuring 28,334 events. This dataset is particularly challenging as it covers diverse topics such as COVID-19 and geopolitical conflicts.
Performance and Robustness
DCE-RD achieved state-of-the-art results across both datasets. On MC-Fake, it reached an accuracy of 90.31%, a 2.6% improvement over the strongest baseline. On Weibo, it achieved 93.28% accuracy.
Beyond raw accuracy, the model demonstrated two critical strengths:
- Early Detection: Rumor detection is most valuable in the first few hours of a post's life. DCE-RD maintained high accuracy even with only 0–2 hours of propagation data, significantly outperforming models that require a more developed graph structure.
- Noise Resistance: In tests where Gaussian noise was added to simulate real-world data corruption, DCE-RD's performance remained stable, whereas standard GNN-based models saw significant declines.
Conclusion
The DCE-RD framework represents a significant step toward making automated rumor detection both more reliable and more explainable. By leveraging diverse counterfactual evidence, the model aligns more closely with human reasoning—asking "what if" certain interactions had not occurred. While future work may expand this to include user-to-user social networks and more complex interaction types, the current framework provides a robust solution for identifying misinformation in the messy, high-speed environment of modern social media.
Summary updated