Ken4962/fake-job-posting-prediction

Fake Job Posting Detection

A machine learning project that uses DeBERTa-v3-base to detect fraudulent job postings using natural language processing techniques.

Project Overview

This project implements a complete pipeline for detecting fake job postings:

1. Data Processing: Comprehensive data cleaning, feature engineering, and class balancing
2. Model Training: Fine-tuned DeBERTa-v3-base transformer model
3. API Development: REST API for model inference (implemented)
4. UI Development: User interface for job posting analysis (implemented)

Features

• Advanced NLP: Uses DeBERTa-v3-base transformer model
• Class Imbalance Handling: SMOTE oversampling for balanced training
• Comprehensive Data Processing: Missing value handling, feature engineering, encoding
• Proper Evaluation: Train/validation/test separation with no data leakage
• Production Ready: Model saved and ready for deployment
• REST API: FastAPI backend for real-time inference
• Modern UI: React frontend for user-friendly predictions

Model Performance

Final Results

• F1-Score: 35.92% on test set
• Recall: 97.63% (catches almost all fraudulent jobs)
• Precision: 22.01% (some false positives, requires human review)
• Accuracy: 34.28% (expected for imbalanced dataset)

Project Structure

fake-job-posting-prediction/
├── data/
│   ├── data_processing.py      # Data preprocessing pipeline
│   ├── data_processing.log     # Processing logs
│   ├── train_data.csv          # Training data (SMOTE applied)
│   ├── val_data.csv            # Validation data
│   ├── test_data.csv           # Test data
│   └── processed_fake_job_postings.csv  # Combined processed data
├── model/
│   ├── train_deberta.py        # Model training script
│   ├── model_training.log      # Training logs
│   ├── final_results.md        # Complete results documentation
│   └── deberta_best_model/     # Best trained model
├── api/
│   ├── main.py                 # FastAPI backend
│   └── ...                     # API code and logs
├── ui/
│   ├── src/                    # React frontend source code
│   └── ...                     # Frontend assets
├── .gitignore                  # Git ignore rules
└── README.md                   # Project documentation

Installation

1. Clone the repository

   git clone <repository-url>
   cd fake-job-posting-prediction
   

2. Set up Python environment (recommended: pyenv + virtualenv)

   pyenv install 3.10.14
   pyenv virtualenv 3.10.14 hf310env
   pyenv activate hf310env
   

3. Install backend dependencies

   pip install -r requirements.txt
   # or
   pip install fastapi uvicorn transformers torch sentencepiece pandas numpy scikit-learn imbalanced-learn pydantic loguru datasets kaggle
   

4. Install frontend dependencies

   cd ui
   npm install
   

Usage

Data Processing

cd data
python data_processing.py

Model Training

cd model
python train_deberta.py

Start the API Backend

cd api
uvicorn main:app --reload

Start the Frontend UI

cd ui
npm start

API Usage

• POST /predict: Send a job posting text and receive a prediction (fraudulent/legitimate) and probability.
• GET /: Health check endpoint.

Frontend Usage

• Access the React app at http://localhost:3000
• Enter a job posting and click "Predict" to see results.

Demo

• The app provides real-time predictions for job postings via a modern web UI.
• Example: Enter "this is a high paying job, no experience needed and it is remote." and see the prediction.

Troubleshooting & Tips

• CORS errors: Ensure CORS middleware is enabled in FastAPI and only one app instance is created.
• Python version: Use Python 3.10.x for best compatibility with ML libraries.
• Environment issues: Always activate your virtualenv before running backend commands.
• Model always predicts one class: Check your model, data balance, and retrain if needed.

Technical Details

Model Configuration

• Model: microsoft/deberta-v3-base
• Parameters: 184M
• Max sequence length: 256
• Batch size: 8
• Learning rate: 2e-5
• Epochs: 3

Data Processing Pipeline

1. Missing value handling (drop columns/rows with >50% missing)
2. Feature selection (remove identifiers, constants, highly correlated)
3. Feature engineering (text length, keyword flags, count features)
4. Categorical encoding (LabelEncoder with unseen category handling)
5. Numerical scaling (StandardScaler)
6. Train/Test split (80/20)
7. Train/Validation split (80/20 of training data)
8. SMOTE oversampling (applied ONLY to training data)

Dataset Statistics

• Original: ~18K samples (94% non-fraudulent, 6% fraudulent)
• Training: 18,730 samples (perfectly balanced with SMOTE)
• Validation: 2,859 samples (original distribution)
• Test: 3,573 samples (original distribution)

Challenges Solved

1. Class Imbalance: Resolved using SMOTE on training data only
2. Data Leakage: Prevented with proper train/validation/test separation
3. MPS Compatibility: Resolved by switching to CPU training
4. Binary Label Conversion: Handled non-binary processed values
5. Zero Division Warnings: Resolved with proper metric handling
6. CORS/API Integration: Fixed CORS and API integration issues for seamless frontend-backend communication

Next Steps

• Data processing pipeline
• Model training and evaluation
• API development and deployment
• Frontend UI integration
• Troubleshooting and bugfixes
• Project documentation

Contributing

1. Fork the repository
2. Create a feature branch
3. Make your changes
4. Add tests if applicable
5. Submit a pull request

License

This project is licensed under the MIT License.

Acknowledgments

• Microsoft for the DeBERTa model
• Hugging Face for the transformers library
• The original dataset providers