Use Case Fraud Detection Pipeline - GuitarAlchemist/ix GitHub Wiki
Navigation: [Home]] ](/GuitarAlchemist/ix/wiki/[ML-Pipelines) | [Architecture]]
Use Case: Fraud Detection Pipeline
Complete fraud detection: diagnose imbalance → SMOTE → model comparison → confusion matrix + ROC/AUC.
Pipeline Steps
- Diagnose imbalance with
class_distribution— 99.5% legit, 0.5% fraud - SMOTE to balance training data —
Smote::new(5, 42).fit_resample() - Compare models: Logistic Regression vs Random Forest vs Gradient Boosting
- Evaluate:
ConfusionMatrix, precision/recall/F1 (recall is king for fraud) - ROC/AUC: threshold-independent ranking evaluation
- Cross-validate: 5-fold stratified for reliability
- With vs Without SMOTE: demonstrate recall improvement
Key Lessons
| Lesson | Why |
|---|---|
| Accuracy lies | 99.5% accuracy = 0% fraud caught |
| Recall is king | Missing fraud costs money |
| SMOTE train only | Never contaminate test data |
| AUC > accuracy | Threshold-independent evaluation |
Algorithms Used
| Algorithm | ix Crate | Role |
|---|---|---|
| SMOTE | ix-supervised | Balance data |
| Gradient Boosting | ix-ensemble | Primary classifier |
| Random Forest | ix-ensemble | Comparison |
| Confusion Matrix + ROC/AUC | ix-supervised | Evaluation |
| Cross-validation | ix-supervised | Reliability |