Fake News Detection in Social Media: a Bertpowered Hybrid Classification Approach

A societal problem that is present today is the spread of misinformation in social media, which is difficult to tackle as it can affect individuals as well as the entire society. Curbing this problem is critical in preserving the trust on the platforms. This study proposes a new method of hybrid classification for fake news detection using BERT for feature extraction. Our approach consists of data cleaning, text encoding using BERT, and classification by hard voting ensemble, where the notion of fake news is opposed to real news. Textual features are voted by the classifiers and assigned to the voting classifier that obtains the textual features from all classifiers and makes the decision based on them. Our model is tested on two datasets: the Fake-or-Real News dataset and BuzzFeed News dataset to determine how these models compare to pre-existing models. Our approach brings about an impressive 89.8 % accuracy on the Fake-or-Real-News dataset, and as for the BuzzFeed site, it brings about 72 % accuracy. These results help prove the effectiveness of the classifiers. With an increase in accuracy of fake news detection, there will also be an increase in the reliability of social media information.