Co-GA: A Bio-inspired Semi-supervised Framework for Fake News Detection on Scarcely Labeled Data

With social media becoming ubiquitous, disinformation such as fake news has become rampant and consequently increased the importance of an effective fake news detection methodology. One of the key challenges faced by researchers is the lack of sufficient labeled data making semi-supervised learning increasingly important for social media data. Non-latent content features differ in fake and real news but at the same time suffer from the high-dimensional nature of text data. In the last decade, metaheuristics research has proved itself to be an effective option for optimization in feature selection. The Co-GA framework described in this paper combined co-training-based semi-supervised learning with bio-inspired metaheuristics for optimization in feature selection. To the best of our knowledge, such an approach has not been attempted so far in disinformation research. The efficacy of this approach has been proved for fake news detection tasks for an article dataset as well as a tweet dataset. This approach can be readily adapted for other learning tasks in social media where sufficient strong supervision is lacking.