Fake News Classification Using the Multinomial Naive Bayes Algorithm with TF-IDF Weighting on the FakeNewsNet Dataset

Abstract

The proliferation of digital and social media platforms has accelerated the production and dissemination of fake news, creating serious risks to public trust, democratic discourse, and social stability. Automatic detection of fake news at the point of publication is therefore an urgent research problem in text mining and natural language processing. This study develops and evaluates a fake news classification pipeline based on the Multinomial Naive Bayes (MNB) algorithm combined with Term Frequency-Inverse Document Frequency (TF-IDF) feature weighting, using the FakeNewsNet benchmark corpus of news article titles labeled as real or fake. After parsing and cleaning the raw repository, 21,637 unique, non-empty news titles remained, comprising 16,399 real (75.8%) and 5,238 fake (24.2%) instances. The text was preprocessed through case folding, URL removal, numeral removal, punctuation removal, tokenization, stopword removal, and lemmatization prior to TF-IDF vectorization with a 5,000-term vocabulary. The dataset was partitioned using an 80:20 stratified train-test split (17,309 training and 4,328 testing instances), and a Multinomial Naive Bayes classifier was trained on the resulting TF-IDF matrix. On the held-out test set, the model achieved an overall accuracy of 82.62%, a weighted precision of 81.98%, a weighted recall of 82.62%, and a weighted F1-score of 80.41%. Class-level analysis revealed a pronounced asymmetry: the real-news class was recognized with high precision (83.28%) and recall (96.43%), whereas the fake-news class, being the minority class, achieved a lower recall of 39.41% despite a precision of 77.92%. These findings confirm that TF-IDF-weighted Multinomial Naive Bayes is a computationally efficient and reasonably accurate baseline for short-text fake news classification, while also highlighting the class-imbalance challenge that limits fake-news recall. The paper discusses the algorithmic rationale, the practical trade-offs of the approach, and directions for improving minority-class detection in future work.