Fake News Detection

Problem Statement

Misinformation spreads faster than corrections. The goal was a real-time NLP classifier that could distinguish fake from real news with high precision - usable by anyone without ML knowledge via a simple paste-and-check interface.

Dataset

44,898 articles from the ISOT Fake News Dataset: 23,481 real Reuters articles and 21,417 fake articles from PolitiFact/unreliable sources. Text was lowercased, punctuation stripped, stop-words removed, and lemmatised before vectorisation.

Architecture

TF-IDF vectoriser (max 10K features, 1–2 n-grams) feeding a Linear SVM classifier. Streamlit front end accepts raw article text, runs the pipeline in memory, and returns a REAL/FAKE verdict with probability score.

Model Selection

Linear SVM was benchmarked against Multinomial Naïve Bayes and Logistic Regression. SVM achieved the best accuracy (99.29%) and F1 (0.99) on the held-out test set and is fastest at inference - ideal for Streamlit deployment.

Training Process

Trained on an 80/20 stratified split. Hyperparameter tuning explored regularisation parameter C (0.1–10) with cross-validation. Final model and vectoriser saved with joblib for instant loading on Streamlit cold start.

Evaluation Metrics

Results

Live at fake-news-detection-nlp-01.streamlit.app - paste any news article and get an instant real/fake verdict backed by SVM confidence scores.

Key Learnings

• 1Linear SVM is remarkably effective for high-dimensional TF-IDF spaces.
• 2n-gram features (bigrams) significantly improve fake-news pattern detection.
• 3Deployment via Streamlit makes ML demos instantly accessible without a backend.