A metaheuristic based filter-wrapper approach to feature selection for fake news detection.

Due to ease of dissemination, humankind is facing an "infodemic" that has spread through electronic and social media. Therefore, there is a need to combat fake news using text classification techniques. However, textual data contains a lot of redundant useless features which can cause issues during the learning and classification phase. Therefore, an effective feature selection method is required to select the important features only. Filter-based methods exist in the literature for feature selection but their performance is average at best. Similarly, many wrapper-based methods also exist but very few are specialized for textual features. In this study, a meta-heuristic based filter-wrapper approach to feature selection is proposed for fake news classification. The proposed algorithm combines three filter-based measures with Binary Dragonfly Algorithm. Moreover, a mechanism for dynamically adjusting the exploratory and exploitative behavior of the said algorithm is also proposed. The hybrid model is evaluated on three datasets of fake news. Additionally, it is also evaluated on the task of sentiment analysis of news. Both binary and multi-class classification tasks were used in our experiments. The proposed algorithm has been compared with several state-of-the-art wrapper-based and filter-based feature selection methods. For fake news detection, Macro F-1 Scores of 0.897, 0.782 and 0.667 were achieved on the three datasets. Moreover, for multi-class sentiment analysis task, Macro F-1 Scores of 0.553 and 0.597 were achieved. [ABSTRACT FROM AUTHOR]