Fake News Detection using NLP & Machine Learning

Fake news spreads faster than the truth — studies show false stories can reach thousands of people within hours on social media. This project builds a working solution: a machine learning system that reads a news article and tells you whether it's likely real or fake, using natural language processing. It is a complete fake news detection using machine learning project that doesn't stop at a notebook — it includes a deployed TruthLens web app where anyone can paste article text and get an instant verdict with a confidence score. The heart of the project is a fair, honest model comparison. A classical TF-IDF + Logistic Regression baseline goes head-to-head with an LSTM sequence model and a fine-tuned transformer (BERT/DistilBERT) baseline, all trained and evaluated on the same labeled news dataset with a proper train/validation/test split. You get real metrics — accuracy, precision, recall, F1 — and a real story to tell in your viva about why different models perform the way they do.

1. Data preparation — a labeled dataset of real and fake news articles is cleaned: lowercased, tokenized, stripped of stopwords, URLs and noise, then lemmatized
2. Feature engineering — for the classical model, text is converted to TF-IDF vectors that weight the words most characteristic of fake vs real writing
3. Training — the Logistic Regression baseline trains in minutes; the LSTM learns word sequences via embedding layers; the transformer is fine-tuned with a classification head
4. Evaluation — all three models are scored on a held-out test set with accuracy, precision, recall, F1-score, and ROC-AUC
5. Deployment — the best model is served through the TruthLens Flask web app: paste an article, and it returns a verdict plus a confidence score in under a second

• Paste-and-check web app: paste any news article into the TruthLens app and get a real/fake verdict with a confidence percentage instantly
• Three-model comparison: TF-IDF + Logistic Regression baseline vs LSTM neural network vs fine-tuned transformer (BERT/DistilBERT)
• Complete NLP pipeline: tokenization, lowercasing, stopword removal, lemmatization, and TF-IDF feature extraction built from scratch
• Confusion matrices and classification reports for every model, with visualizations you can drop straight into your report
• Live model-switching in the TruthLens app: compare predictions from the baseline and deep models on the same article
• Trained model files included — no GPU needed to run the app
• REST-style prediction flow inside the Flask app, ready to demo and extend

• Full source code (preprocessing, training notebooks, and the TruthLens Flask app)
• Trained model files for all three approaches — ready to run, no training required
• Labeled dataset access guide and preprocessing scripts
• Project report (PDF) with methodology, model comparison tables, and result analysis
• Presentation (PPT) for your project demo and review
• Viva Q&A document covering NLP concepts, model choices, and likely examiner questions
• README with setup, training, and deployment instructions