Social Media Fake News Detection Using Machine Learning Models and Feature Extraction Techniques

Due to continuous technological advances, fake news is rampant on social media platforms, making it an urgent area of research. This research investigates various Machine Learning (ML) algorithms and feature extraction techniques to classify news as real or fake. The study compares different ML models, such as Logistic Regression (LR), Support Vector Machine (SVM), Decision Tree (DT), K-Nearest Neighbors (K-NN), Naïve Bayes (NB), Multilayer Perceptron (MLP), and Stacked ensemble models. Six feature extraction methods were utilized, measured, and compared, including Term Frequency-Inverse Document Frequency (TF-IDF), Bag of Words (BOW), 2-grams, 3-grams, (1-2) grams, and (1-2-3) grams. The experiments were on a merged dataset for supervised learning and a separate test dataset for blind testing. The results showed that SVM, DT, and Stacked models achieve up to 99% accuracy with TF-IDF and BOW on the merged dataset. For the blind testing, K-NN for BOW and MLP for TF-IDF had an accuracy of 64% on separate test datasets. Lastly, the results demonstrated that n-grams helped to understand the context and the order of words, but a higher value of $\mathbf{n}$ increased dimensionality and caused sparsity.