📰 Fake News Detection using BiLSTM

This project applies a Bidirectional LSTM (BiLSTM) neural network to detect fake news based on the content of news articles. It includes steps for data visualization, preprocessing, model building, training, and evaluation.

📁 Project Structure

.
├── data_analysis.ipynb       # Notebook for visualizing the dataset
├── lstm.ipynb                # Main notebook with BiLSTM model
├── requirements.txt          # Required Python libraries
├── News-_dataset             # Folder containing True.csv and Fake.csv after download
├── images/                   # Folder containing output plots 

🔧 Setup Instructions

1. Clone the repository

   git clone https://github.com/yourusername/fake-news-lstm.git
   cd fake-news-lstm

2. Install dependencies

   pip install -r requirements.txt

3. Download and place the dataset

   • Go to: ISOT Fake News Dataset
   • Download and unzip the file.
   • Make sure that the unzipped folder, News-_dataset, contains True.csv and Fake.csv. Make sure the folder is kept in the root directory.

📊 Data Analysis

Open data_analysis.ipynb and run all cells to visualize the dataset. This notebook provides:

• 📈 Number of true vs. fake news articles
• ❗ Null value check for both True and Fake news files
• ☁️ Word clouds showing most common words in each category
• 📏 Histograms showing the distribution of article lengths

Word Cloud (True News)	Word Cloud (Fake News)

🧠 LSTM Model

Open lstm.ipynb to run the full model pipeline.

🛠 Preprocessing steps:

• Convert text to lowercase and remove punctuation
• Filter non-alphabetic characters
• Stopword removal using NLTK
• Tokenize the cleaned text and apply padding for uniform input shape

📐 Model Architecture:

• Embedding Layer
• Bidirectional LSTM
• Global Max Pooling
• Dense Layers
• Sigmoid Output Layer for binary classification

Model Architecture

📈 Evaluation Metrics

The model’s performance was evaluated using the following metrics:

• ✅ Accuracy: 99.81%
• 🎯 Precision: 100.00%
• 📡 Recall: 99.60%
• 📊 F1 Score: 99.80%
• 📉 AUC (Area Under ROC Curve): 1.0000
• ⚖️ EER (Equal Error Rate): 0.0014 at threshold 0.0000

ROC Curve

Confusion Matrix

🧾 Conclusion

We implemented a Bidirectional LSTM model for fake news detection, inspired by the approach proposed in the paper “Detecting opinion spams and fake news using text classification”.

Our results demonstrate excellent performance on the ISOT dataset, with an AUC of 1.0 and an accuracy of 99.81%, showing that the model is highly capable of distinguishing between real and fake news articles. The results validate the effectiveness of Bidirectional LSTM architectures in text classification tasks.

📚 References

• Kadek Sastrawan, I.P.A. Bayupati, Dewa Made Sri Arsa, Detection of fake news using deep learning CNN–RNN based methods, ICT Express, Volume 8, Issue 3, 2022, Pages 396-408, ISSN 2405-9595, https://doi.org/10.1016/j.icte.2021.10.003.