Weighted Contrastive Teacher: A Teacher-Student Framework for Unsupervised Cross-Domain Rumour Detection

Unsupervised cross-domain rumour detection aims to identify rumours in the target domain by utilizing a detection model trained on source domain data, where the key challenge lies in addressing the domain gap between these domains. Most existing methods adopt transformers as their backbone models, which effectively represent semantic features but cannot extract the structural features of tweet propagation, resulting in cross-domain detection limitations. In this study, we propose a teacher-student framework named a weighted contrastive teacher. This method utilizes the consistency regularization loss between the teacher and student model to overcome the domain gap. It combines pre-trained language models and graph convolutional networks as the teacher model to integrate semantic and propagation structural features. The effectiveness of the consistency regularization loss is enhanced through dynamic pseudo-label filtering and loss function reweighting. Additionally, we introduce categorical and sample-wise contrastive learning methods to better align the feature distributions between the domains. We conduct experiments on four groups of rumour detection datasets, and the results demonstrate that in comparison to baseline models, our model demonstrates exceptional performance.