🏦 Credit Risk Assessment DNN
A custom Deep Neural Network with Feature Attention Gate trained on the Taiwan Credit Card Default dataset (30,000 samples, 23 features) to predict credit default risk.
Model Architecture
CreditRiskDNN( FeatureAttention(23 → 23) # soft attention gate per feature Linear(23 → 128) + BatchNorm + ReLU + Dropout(0.3) Linear(128 → 64) + BatchNorm + ReLU + Dropout(0.3) Linear(64 → 32) + BatchNorm + ReLU Linear(32 → 1) # binary output )
Performance
| Metric | Score |
|---|---|
| ROC-AUC | ~0.79 |
| F1 (default) | ~0.68 |
| Threshold | 0.4 |
Usage
from huggingface_hub import hf_hub_download
import torch, pickle
# Download artifacts
hf_hub_download("pankajkapri/credit-risk-dnn", "best_model.pth", local_dir="outputs")
hf_hub_download("pankajkapri/credit-risk-dnn", "scaler.pkl", local_dir="outputs")
# Load
scaler = pickle.load(open("outputs/scaler.pkl", "rb"))
model = CreditRiskDNN(n_features=23)
model.load_state_dict(torch.load("outputs/best_model.pth", map_location="cpu"))
model.eval()
# Predict
raw = [200000, 2, 2, 2, 35, -1, -1, -1, -1, -1, -1,
50000, 48000, 45000, 43000, 41000, 40000,
5000, 5000, 5000, 5000, 5000, 5000]
scaled = scaler.transform([raw]).tolist()
X = torch.tensor([scaled], dtype=torch.float32)
with torch.no_grad():
prob = torch.sigmoid(model(X)).item()
print(f"Default probability: {prob:.1%}")
Live Demo
- Frontend: https://your-app.vercel.app
Fairness
Audited across age groups using fairlearn —
demographic parity difference and equalized odds reported.
Explainability
SHAP KernelExplainer used for global feature importance and per-prediction waterfall charts.