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\nsocial problem, which cannot be addressed in a timely manner using manual\ninvestigation. This has motivated numerous studies on automating fake news\ndetection. Most studies explore supervised training models with different\nmodalities (e.g., text, images, and propagation networks) of news records to\nidentify fake news. However, the performance of such techniques generally drops\nif news records are coming from different domains (e.g., politics,\nentertainment), especially for domains that are unseen or rarely-seen during\ntraining. As motivation, we empirically show that news records from different\ndomains have significantly different word usage and propagation patterns.\nFurthermore, due to the sheer volume of unlabelled news records, it is\nchallenging to select news records for manual labelling so that the\ndomain-coverage of the labelled dataset is maximized. Hence, this work: (1)\nproposes a novel framework that jointly preserves domain-specific and\ncross-domain knowledge in news records to detect fake news from different\ndomains; and (2) introduces an unsupervised technique to select a set of\nunlabelled informative news records for manual labelling, which can be\nultimately used to train a fake news detection model that performs well for\nmany domains while minimizing the labelling cost. Our experiments show that the\nintegration of the proposed fake news model and the selective annotation\napproach achieves state-of-the-art performance for cross-domain news datasets,\nwhile yielding notable improvements for rarely-appearing domains in news\ndatasets.\n