Fraudulent Job Posting Detection with CGCL

This repository contains the implementation of our proposed Centroid-Guided Contrastive Loss (CGCL) for fraudulent job posting detection.
CGCL unifies classification and clustering objectives to learn discriminative and compact latent-space representations, improving both accuracy and structural interpretability.

---

🚀 Features

• Centroid-Guided Contrastive Loss (CGCL) with top-k push-and-pull mechanism.
• Integration of Cross-Entropy Loss for strong class discrimination.
• Supports multiple feature extraction techniques:
  • Word2Vec
  • GloVe
  • TF-IDF
• Extensive experiments with hyperparameter tuning.
• Evaluation on both classification and clustering metrics.

---

📂 Project Structure

.
├── data/                 # Dataset and preprocessing scripts
├── models/              # Model definitions and loss functions
├── utils/               # Helper functions

---

⚙️ Installation

Clone this repository and install the required dependencies:

git clone https://github.com/your-username/fraudulent-job-detection-cgcl.git
cd fraudulent-job-detection-cgcl
pip install -r requirements.txt

---

📊 Results

Our final GloVe-based model achieved:

• Accuracy: 99.2%
• F1-score: 98.7%
• Strong clustering performance with:
  • Silhouette Score: 0.701
  • Adjusted Rand Index: 0.953
  • Normalized Mutual Information: 0.903

Performance Comparison

Embedding	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)	Silhouette	ARI	NMI
Word2Vec	94.8	89.6	93.2	91.4	0.425	0.723	0.712
GloVe	99.2	98.9	98.5	98.7	0.701	0.953	0.903
TF-IDF	97.1	95.3	96.8	96.0	0.593	0.847	0.825

Hyperparameter Analysis

Optimal Configuration (GloVe)

• Learning Rate: 1e-4
• Hidden Dimensions: [512, 256, 128]
• Temperature (τ): 0.1
• Top-k: 10
• Epochs: 2000
• Loss Weight (α): 0.7

Learning Rate Impact

Learning Rate	Accuracy (%)	F1-Score (%)
1e-3	92.4	90.2
1e-4	99.2	98.7
1e-5	96.8	95.4

---

📈 Visualizations

• AUC-ROC curve for the best-performing model.
• Latent-space visualization showing compact and separable clusters.
• Loss convergence plots during training.
• Hyperparameter sensitivity analysis.

---

🔬 Key Technical Contributions

Centroid-Guided Contrastive Loss (CGCL)

L_CGCL = L_CE + α · L_contrastive

where:
L_contrastive = (1/N) Σ [pull_loss + push_loss]
pull_loss = d(z_i, c_y_i)²
push_loss = max(0, margin - d(z_i, c_j))²

Architecture Overview

1. Feature Extraction Layer: Processes text embeddings (Word2Vec/GloVe/TF-IDF)
2. Multi-layer Perceptron: [Input → 512 → 256 → 128 → 64 → 2]
3. Dual Objective Training: Classification + Contrastive Learning
4. Top-k Selection: Focuses on hardest negatives for efficient training

---

📊 Dataset Information

Metric	Value
Total Samples	17,880
Fraudulent Posts	866 (4.8%)
Legitimate Posts	17,014 (95.2%)
Train/Test Split	80/20
Features	18 (text + metadata)

Feature Categories

• Textual Features: Job title, description, requirements, benefits
• Metadata Features: Company profile, location, employment type, experience level

---

🏷️ Keywords

Fraud Detection, Contrastive Learning, Clustering, Word2Vec, GloVe, TF-IDF, Anomaly Detection, Representation Learning

---

🤝 Contributing

Pull requests are welcome! For major changes, please open an issue first to discuss what you would like to change.

Development Setup

# Create virtual environment
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install development dependencies
pip install -r requirements-dev.txt

# Run tests
python -m pytest tests/

---

📧 Contact

For any queries, please contact: k224058@nu.edu.pk

---

📜 License

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