Embracing Domain Differences in Fake News: Cross-domain Fake News Detection using Multi-modal Data

With the rapid evolution of social media, fake news has become a significant problem, which cannot be addressed in timely manner using manual investigation. This motivated numerous studies on automating detection. Most explore supervised training models with different modalities (e.g., text, images, and propagation networks) records to identify news. However, performance such techniques generally drops if are coming from domains politics, entertainment), especially for that unseen or rarely-seen during training. As motivation, we empirically show have significantly word usage patterns. Furthermore, due sheer volume unlabelled records, it is challenging select labelling so domain-coverage labelled dataset maximised. Hence, this work: (1) proposes novel framework jointly preserves domain-specific cross-domain knowledge detect domains; (2) introduces an unsupervised technique set informative labelling, can ultimately used train detection model performs well many while minimizing cost. Our experiments integration proposed selective annotation approach achieves state-of-the-art datasets, yielding notable improvements rarely-appearing datasets.