Multi-Source Domain Adaptation with Weak Supervision for Early Fake News Detection

Recently, the massive and diverse fake news from politics to entertainment and health has amplified the social distrust problem and has become a big challenge for the society and research community. The existing fake news detection methods are mostly designed for either a specific domain or require huge labeled data from various domains. If there is not enough labeled data in a certain domain, existing models may not work well for detecting fake news from that domain. To overcome these limitations we propose a novel framework based on multi-source domain adaptation and weak supervision for early fake news detection. The framework transfers sufficient labeled source domains' knowledge into a target/new domain with limited or even no labeled data by the multi-source domain adaptation, and applies researchers' prior knowledge about fake news to the target domain by the weak supervision. The weak supervision assigns the weak labels to the unlabeled samples in the target domain through known heuristic rules. Our experimental results show that our approach outperforms 7 state-of-the-art methods in three real-world datasets. In particular, our model achieves, on average, 5.2% higher accuracy than the best baseline. Our model with a more advanced encoder can further boost the performance by 3.7%. The code is available at this clickable link.