Comment Filtering Based Explainable Fake News Detection

Abstract

Fake News Detection is one of the most currently researched areas over the globe; many methods have come to light using different features as their sources. Hence, there are also methods using existing comments on any news article which can be used to determine the credibility of the news article as fake or real. Here, we have introduced a hypothesis that uses a machine learning approach to check the credibility of comments before they can be analyzed for further fake news detection. So, we have used various text classification algorithms to check for our hypothesis that filtering comments since there is a high possibility that the comments used for any analysis can be useless and full of useless stuff. For example, the comments showing only the emotion of readers like ‘Yesss’ or ‘Nooo!’ and likewise or the comments built using only the curse words. Such comments would prove useless as a contributing factor for fake news detection and might also affect the results of fake news detection for any news article. These text classifiers are—Complement Naïve Bayes, Logistic Regression, Multinomial Naïve Bayes, and Support Vector Machine. Out of these, the best accuracy is provided by the MultinomialNB method of 75.7% and Decision Tree with 75.4% as opposed to the original algorithm with an accuracy of 73.3% using the same dataset. Since the MultinomialNB has provided the best improvement in all the metrics compared to the original method, and we are focusing our paper on this method. This hypothesis aims to classify comments as junky (useless) comments and utility (useful) comments. These utility comments will be further used for analysis to identify fake news. Also, since the size of comments per article may vary from a few tens to a few hundred or thousands, we have used the semi-supervised approach to classify the comments in junky or utility comments classes. We have also collected data from various sources and collaborated them to fetch ourselves from a usable dataset. It contains 415 records with contents or article data for each record, along with many comments for each record. Moreover, we have also classified those comments into junky and utility comment classes using the basic definition of spam filtering. This can be improvised for different uses using different criteria. Hence, eradicating the useless comments and only analyzing the useful comments for better identification of fake news is fake news detection.