Unsupervised Deep Learning for Fake Content Detection in Social Media

Fake content is ever increasing in the online environment, driven by various motivations such as gain-ing commercial and political advantages. The interactive and collaborative nature of social media further fuels the growth of fake content by exerting fast and widespread influence. Despite growing and interdisciplinary efforts in detecting fake content in social media, some common research challenges remain to be addressed such as humans’ cognitive bias and scarcity of labeled data for training supervised machine learning models. This study aims to tackle both challenges by developing unsupervised deep learning models for the detection of fake content in social media. In view that traditional linguistic features fail to capture context information, our proposed method learns feature representations from the context in social media content. The empirical evaluation results with fake comments from YouTube demonstrate that our proposed methods not only outperform baseline models with traditional unsupervised machine learning techniques, but also achieve comparable performance to the state-of-the-art supervised models. The proposed analytical pipeline provides an end-to-end solution to detecting fake social media contents, which largely reduce the human labor required in collaborative data science teams (i.e., particularly the data labeling). The findings of this study can be used to facilitate collaboration in data science by reducing humans’ cognitive bias and improve the collaboration efficiency