# DECAF: Dynamic Entropy-guided Credit Assessment Framework

This repository contains the implementation of DECAF, a deep learning framework for credit assessment that combines Capsule Networks (CapsNet) and Soft Actor-Critic (SAC) algorithms.

## Environment Requirements

```python
python==3.9
torch==2.1.0
numpy==1.21.0
pandas==1.4.0
scikit-learn==1.0.2
xgboost==1.5.0
matplotlib==3.5.0
```

## Project Structure

```
├── data/
│   └── car_insurance.csv    # Car Insurance Dataset
├── models/
│   ├── capsnet.py          # CapsNet implementation
│   └── sac.py              # SAC implementation
├── utils/
│   ├── data_utils.py       # Data preprocessing utilities
│   └── eval_utils.py       # Evaluation metrics
├── train.py                # Main training script
└── evaluate.py             # Evaluation script
```

## Data Preprocessing

The Car Insurance Dataset is used in this project. You can download it from [Kaggle](https://www.kaggle.com/datasets/sagnik1511/car-insurance-data).

Main preprocessing steps include:
1. Numerical feature normalization
2. Categorical feature encoding
3. Train/validation/test split
4. Class imbalance handling using SMOTE

## Model Training

To train the DECAF model:

```bash
python train.py --data_path data/car_insurance.csv --epochs 50 --batch_size 256
```

Key hyperparameters are configured in `train.py`:
- Learning rate: 1e-4
- CapsNet layers: 3
- SAC target entropy: -2
- Replay buffer size: 5e5

## Evaluation

To evaluate the trained model:

```bash
python evaluate.py --model_path checkpoints/best_model.pth --test_data data/car_insurance.csv
```

The evaluation script will output:
- ROC curves
- AUC scores
- Accuracy metrics
- Loss curves

## References

If you use this code, please cite our paper:
```
@article{decaf2024,
  title={DECAF: Dynamic Entropy-guided Credit Assessment Framework},
  year={2024}
}
```