Enhancing News Authenticity Prediction with Machine Learning Approach

The rapid proliferation of misinformation across digital platforms poses a critical threat to public discourse, democratic processes, and societal trust. Manual verification systems cannot keep pace with the volume and velocity of fake news being generated, necessitating the development of automated detection mechanisms. This work presents a lightweight machine learning approach for fake news detection using a PassiveAggressive Classifier combined with TF-IDF vectorization, trained and evaluated on the WELFake benchmark dataset comprising 72,119 news articles sourced from multiple platforms spanning 2016 to 2020. The proposed system achieves 96.19% classification accuracy on 14,424 test articles, with precision and recall scores of 0.96-0.97 for both classes and a balanced F1-score of 0.96, validated using a CalibratedClassifierCV wrapper providing calibrated probability outputs. Beyond classification, the system incorporates LIME-based explainability for word-level prediction reasoning, real-time URL verification with source credibility analysis, clickbait detection, and evidence-backed verdict generation using the Gemini API.