Fake News Detector

An AI-powered web app that classifies news headlines and articles as REAL or FAKE using a TF-IDF + Logistic Regression model trained on 72,000+ real-world articles.

---

Project Structure

fake-news-detector/
├── backend/
│   ├── main.py           # FastAPI server — exposes /predict endpoint
│   ├── model.py          # Dataset loading, training, and inference
│   ├── requirements.txt  # Python dependencies
│   └── model.pkl         # Generated automatically on first run
├── frontend/
│   ├── index.html        # UI
│   ├── style.css         # Styles
│   └── script.js         # Calls the API and renders results
└── README.md

---

Getting Started

1. Set up the backend

cd backend
python -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\activate
pip install -r requirements.txt

2. Add a dataset

The model needs a dataset CSV before it can train. Place one of the following inside backend/:

Dataset	Size	File(s) needed	Download
WELFake (recommended)	72,134 articles	WELFake_Dataset.csv	Zenodo · Kaggle · Hugging Face
ISOT	44,898 articles	True.csv + Fake.csv	Kaggle

If no CSV is found, model.py will attempt to auto-download WELFake from Zenodo (requires internet access).

3. Train the model

python model.py

This downloads the dataset (if needed), trains the model, prints a classification report, and saves model.pkl. Takes 1–3 minutes on a standard laptop.

4. Start the API server

uvicorn main:app --reload

Server runs at http://127.0.0.1:8000. On first launch, if model.pkl doesn't exist yet, training runs automatically.

5. Open the frontend

Open frontend/index.html directly in your browser. If you run into CORS issues, serve it locally instead:

cd frontend
python -m http.server 5500
# Open http://localhost:5500

---

API

POST /predict

Request body

{ "text": "Paste a headline or article here" }

Response

{
  "label":      "FAKE",
  "confidence": 91.3,
  "conf_label": "High",
  "fake_prob":  91.3,
  "real_prob":   8.7
}

label is either "REAL" or "FAKE". conf_label is "High" (≥80%), "Medium" (≥60%), or "Low".

---

Model

Vectorizer	TF-IDF · 100k features · unigrams + bigrams · sublinear_tf=True
Classifier	Logistic Regression · C=5 · lbfgs solver
Input	Article title + body concatenated
Accuracy	~98% on WELFake 10% test split
Stack	scikit-learn · FastAPI · Uvicorn

---

Notes

• Paste full article text for best results — headlines alone are less reliable.
• Press Ctrl + Enter in the text box to submit without clicking.
• To retrain from scratch: python model.py (overwrites model.pkl).
• This is a demonstration project. Always verify news with primary sources.