Social media has greatly enabled people to participate in online activities\nat an unprecedented rate. However, this unrestricted access also exacerbates\nthe spread of misinformation and fake news online which might cause confusion\nand chaos unless being detected early for its mitigation. Given the rapidly\nevolving nature of news events and the limited amount of annotated data,\nstate-of-the-art systems on fake news detection face challenges due to the lack\nof large numbers of annotated training instances that are hard to come by for\nearly detection. In this work, we exploit multiple weak signals from different\nsources given by user and content engagements (referred to as weak social\nsupervision), and their complementary utilities to detect fake news. We jointly\nleverage the limited amount of clean data along with weak signals from social\nengagements to train deep neural networks in a meta-learning framework to\nestimate the quality of different weak instances. Experiments on realworld\ndatasets demonstrate that the proposed framework outperforms state-of-the-art\nbaselines for early detection of fake news without using any user engagements\nat prediction time.\n