Unsupervised contrastive domain adaptive rumor detection with test-time classifier adjustment

Abstract

Domain-adaptive rumor detection faces significant challenges in mitigating distributional shifts between the source and target domains. Although contrastive learning-based models have shown promise, they exhibit two fundamental shortcomings. Firstly, neglecting the impact of source content on feature alignment may hinder discriminative feature learning. Secondly, relying on unbiased classifier assumptions despite inherent distributional discrepancies in target data. To address these challenges, we propose a novel method called Unsupervised Contrastive Domain Adaptive Rumor Detection with Test-Time Classifier Adjustment (CDTT). Our contrastive domain adaptation framework utilizes a stance-based contrastive learning mechanism to align latent stance features across domains while maintaining content independence. Additionally, to address label unavailability in the target domain, we devise a pseudo-label generation strategy that aggregates nearest-neighbor probabilities through feature-space distance-based batch soft voting. Finally, we implement a test-time adaptation strategy that refines the source-trained classifier by constructing class-wise pseudo-prototypes from unlabeled target data and optimizing prediction through distance-based sample classification. Extensive experiments conducted on four groups of cross-domain datasets and a cross-event dataset showcase that our model surpasses the state-of-the-art baselines.