📰 Fake News Detection using Deep Learning (TensorFlow)

📌 Overview

In today’s hyperconnected digital ecosystem, the rapid spread of fake news poses serious risks ranging from public misinformation to political and social instability. This project focuses on building a deep learning–based fake news detection system that classifies news articles as real or fake based on their textual content.

The model leverages natural language processing (NLP) techniques combined with a hybrid CNN–LSTM architecture implemented in TensorFlow to capture both local textual patterns and long-term contextual dependencies.

---

🎯 Objectives

• Preprocess and clean raw news text data
• Convert textual data into numerical representations using tokenization and padding
• Utilize pre-trained GloVe word embeddings for semantic richness
• Design and train a CNN–LSTM neural network
• Evaluate model performance using accuracy, confusion matrix, and classification report
• Demonstrate an end-to-end deep learning workflow for text classification

---

📂 Dataset Description

• Dataset name: fake_news.csv
• Total samples: 6,335
• Classes:
  • REAL
  • FAKE

Features

Column	Description
title	News headline
text	Full news article
label	Target variable (REAL / FAKE)

The dataset is well balanced across classes, making it suitable for supervised classification.

---

🛠 Tools & Technologies

• Language: Python

• Framework: TensorFlow / Keras

• Libraries:

  • NumPy
  • Pandas
  • scikit-learn
  • Matplotlib
  • TensorFlow (Keras API)

• NLP Techniques:

  • Tokenization
  • Sequence padding
  • Pre-trained word embeddings (GloVe)

---

🧠 Model Architecture

The model follows a hybrid CNN–LSTM design:

1. Embedding Layer

   • Initialized with pre-trained GloVe (300-dimensional)
   • Non-trainable to preserve semantic structure

2. Dropout Layer

   • Prevents overfitting

3. 1D Convolution Layer

   • Extracts local n-gram patterns

4. Max Pooling Layer

   • Reduces dimensionality and computation

5. LSTM Layer

   • Captures long-term dependencies in text sequences

6. Dense Output Layer

   • Sigmoid activation for binary classification

Loss function: binary_crossentropy
Optimizer: Adam

---

📁 Repository Structure

fake-news-detection/
│
├── data/
│ └── fake_news.csv
│
├── notebooks/
│ └── fake_news_detection.ipynb
│
├── results/
│ ├── fake_news_detection.html
│ └── fake_news_detection.pdf
│
├── LICENSE
└── README.md

---

📊 Results Summary

• The model achieves approximately 78% classification accuracy on the test set.
• Performance is balanced across REAL and FAKE classes.
• Training accuracy is high, while validation accuracy highlights realistic generalization behavior.
• Confusion matrix and learning curves provide transparent performance evaluation.
• Generated outputs are available in the results/ directory.

---

🧠 Key Insights

• Hybrid CNN–LSTM models effectively capture both local and sequential text features.
• Pre-trained embeddings significantly improve semantic understanding.
• Deep learning models can overfit quickly without regularization.
• Even with balanced data, validation performance reflects the inherent complexity of fake news detection.

---

📜 License

This project is licensed under the MIT License.
See the LICENSE file for details.

---

📌 Notes

• This project is intended for educational and portfolio purposes.
• The code prioritizes clarity, reproducibility, and interpretability.
• Possible future extensions include:
  • Attention mechanisms
  • Bidirectional LSTM architectures
  • Handling class imbalance
  • Real-time news stream integration

🧾 Author

Mr Rup GitHub: https://github.com/Mr-Rup

---