Unsupervised Content-Based Identification of Fake News Articles with Tensor Decomposition Ensembles

Social media provide a platform for quick and seamless access to information. However, the propagation of false information, especially during the last year, raises major concerns, especially given the fact that social media are the primary source of information for a large percentage of the population. False information may manipulate people’s beliefs and have real-life consequences. Œerefore, one a major challenge is to automatically identify false information by categorizing it into di‚erent types and notify users about the credibility of di‚erent articles shared online. Existing primarily focuses on feature generation and selection from various sources, including corpus-related features. However, so far, prior work has not paid considerable aŠention to the following question: how accurately can we distinguish di‚erent categories of false news, solely based on the content? In this paper we work on answering this question. In particular, we propose a tensor modeling of the problem, where we capture latent relations between articles and terms, as well as spatial/contextual relations between terms, towards unlocking the full potential of the content. Furthermore, we propose an ensemble method which judiciously combines and consolidates results form di‚erent tensor decompositions into clean, coherent, and highaccuracy groups of articles that belong to di‚erent categories of false news. We extensively evaluate our proposed method on real data, for which we have labels, and demonstrate that the proposed algorithm was able to identify all di‚erent false news categories within the corpus, with average homogeneity per group of up to 80%.