📰 FakeNewsClassification

This project is focused on building a robust fake news classification model using a pre-trained DistilBERT transformer model, implemented with TensorFlow and fine-tuned on a labeled dataset. The classifier identifies whether a given text is fake or real news, handling imbalanced data and optimizing the model's performance.

🎯 Project Overview

• Goal: To classify news articles as either fake or real.
• Model: Fine-tuned DistilBERT model.
• Dataset: WELFake Dataset on Kaggle.
• Challenges Addressed:
  • Imbalanced data handling using SMOTE and class weighting.
  • Hyperparameter tuning for improved precision and recall.
  • Jupyter Notebook processing for seamless integration with GitHub.

📂 Files

• notebook.ipynb : Main notebook file containing data preprocessing, model training, and evaluation.
• README.md : Project overview, usage instructions, and requirements.

🛠️ Installation

Clone the repository and install dependencies in a virtual environment.

git clone https://github.com/yourusername/yourrepo.git
cd yourrepo
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

🏗️ Project Setup and Requirements

1. Dependencies: Required libraries include tensorflow, transformers, scikit-learn, imbalanced-learn, matplotlib, and seaborn. Install all with:

bash pip install tensorflow transformers scikit-learn imbalanced-learn matplotlib seaborn 3. Dataset: Download the dataset from Kaggle and place it in the root directory, or specify the correct path in notebook.ipynb.

🔍 Model Training and Evaluation

1. Data Preprocessing: The text data is tokenized with the DistilBERT tokenizer, truncated to a maximum length of 100 tokens.
2. SMOTE and Class Weighting: Applied to manage class imbalance, improving recall and precision.
3. Model Architecture: Fine-tuned DistilBERT model with dropout layers for regularization.
4. Training: The model is trained with balanced cross-entropy loss and an Adam optimizer.

🚀 Running the Notebook

Open the Jupyter Notebook to explore, train, and evaluate the model:

bash jupyter notebook notebook.ipynb

💻 Example Usage

After training, load the saved model and make predictions as shown below:

import pickle
from transformers import DistilBertTokenizer
import tensorflow as tf

# Load model and tokenizer info
with open('./model/info.pkl', 'rb') as f:
    MODEL_NAME, MAX_LEN = pickle.load(f)
tokenizer = DistilBertTokenizer.from_pretrained(MODEL_NAME)
model = tf.keras.models.load_model('./model/clf.keras')

# Sample prediction
text = "Sample news article text to classify."
inputs = tokenizer(text, max_length=MAX_LEN, truncation=True, padding='max_length', return_tensors="tf")
logits = model.predict([inputs["input_ids"], inputs["attention_mask"]])
print("Predicted class:", tf.argmax(logits, axis=1).numpy()[0])

📊 Results

The model achieved a balanced accuracy of approximately 83% on the validation set. Performance can be improved with further tuning.

Evaluation Metrics

Metric	Score	Score	Score	Score
Precision	0.79	0.87	0.83	0.83
Recall	0.88	0.78	0.83	0.83
F1 Score	0.83	0.82	0.83	0.83

🔮 Future Improvements

• Experiment with other oversampling techniques.
• Test additional transformer architectures for improved accuracy.
• Further optimize batch size and learning rate to reduce overfitting.