Modelpublic

Rakshi1/fake-news-detector

sourceHugging Facemitupdated 2d agoView on Hugging Face

0likes48downloads

Model Card

📰 Fake News Detection Using Deep Learning (LSTM)

![Open In Colab](https://colab.research.google.com/github/Rakshi1) ![Hugging Face Space](https://huggingface.co/spaces/Rakshi1/fake-news-detector-app) ![Hugging Face Model](https://huggingface.co/Rakshi1/fake-news-detector)

🚀 Live Interactive Web App: https://huggingface.co/spaces/Rakshi1/fake-news-detector-app

NLP • Supervised Learning • Binary Text Classification • Long Short-Term Memory

This repository hosts a production-grade Deep Learning text classification model that reads news headlines and body text to classify whether an article is Fake News (0) or Real News (1).

📊 Model Evaluation Performance

Evaluated on a 20% stratified test split from the ISOT Fake and Real News Dataset:

Metric	Score	Description
Accuracy	91.83%	Overall percentage of correct classifications across test set
Precision	94.04%	Of all articles predicted as Real, 94.04% were actually Real
Recall	89.33%	Of all actual Real articles, 89.33% were successfully detected
F1-Score	91.62%	Balance between precision and recall

Confusion Matrix

—True Fake Detected: 283 / 300 (94.3%)
—True Real Detected: 268 / 300 (89.3%)

🧠 Neural Network Architecture

text

Input News Headline & Article Body
      │
      ▼
Text Preprocessing (Lowercasing, symbol removal, whitespace normalization)
      │
      ▼
Tokenizer (Vocab size: 20,000, Sequence length: 300)
      │
      ▼
Embedding Layer (dim=128, mask_zero=True)
      │
      ▼
LSTM Layer (units=64, return_sequences=False)
      │
      ▼
Dropout Layer (rate=0.3)
      │
      ▼
Dense Layer (units=1, activation='sigmoid')
      │
      ▼
Output: Fake News (0) [< 0.5] or Real News (1) [>= 0.5]

🚀 How to Test & Predict Using this Hugging Face Model

You can load and test this model directly in Python or Google Colab with 4 lines of code:

python

import pickle
from huggingface_hub import hf_hub_download
from tensorflow.keras.models import load_model
from tensorflow.keras.preprocessing.sequence import pad_sequences

# 1. Download model weights and tokenizer from Hugging Face
model_file = hf_hub_download(repo_id="Rakshi1/fake-news-detector", filename="models/fake_news_lstm.keras")
tok_file = hf_hub_download(repo_id="Rakshi1/fake-news-detector", filename="models/tokenizer.pickle")

# 2. Load artifacts
model = load_model(model_file)
with open(tok_file, 'rb') as f:
    tokenizer = pickle.load(f)['tokenizer']

# 3. Test on any news text
def predict(title, text):
    content = f"{title} {text}".lower().strip()
    seq = tokenizer.texts_to_sequences([content])
    padded = pad_sequences(seq, maxlen=300, padding='post', truncating='post')
    prob = float(model.predict(padded)[0][0])
    label = "Real News" if prob >= 0.5 else "Fake News"
    conf = (prob if prob >= 0.5 else 1.0 - prob) * 100
    return label, round(conf, 2)

# Quick Test Example
label, confidence = predict(
    title="White House signs executive order on cybersecurity",
    text="WASHINGTON (Reuters) - Government announces new infrastructure regulations."
)
print(f"Prediction: {label} ({confidence}% confidence)")

📁 Repository Contents

—models/fake_news_lstm.keras: Trained Keras neural network weights.
—models/tokenizer.pickle: Fitted tokenizer dictionary.
—notebooks/fake_news_detection.ipynb: Complete 12-step notebook for Google Colab and Jupyter.
—src/: Complete source code (model.py, preprocessor.py, train.py, evaluate.py, predict.py).
—app/app.py: Interactive Streamlit web interface.
—reports/: Confusion matrix heatmap and training curves.
—requirements.txt: Environment dependencies.