Ml antipattern validator

ML Antipattern Validator

Overview

AI/ML 개발에서 30+ 안티패턴을 감지하고 방지하는 스킬입니다.

Key Principle: Honest evaluation > Impressive metrics.

When to Activate

Automatic Triggers:

• ML training code (train*.py, model training)
• Dataset preparation or splitting
• Model evaluation or testing
• Production deployment planning

Manual Triggers:

• @validate-ml - Full validation
• @check-leakage - Data leakage detection
• @verify-eval - Evaluation methodology

---

Pre-Implementation Checklist

✅ Requirements:
□ Problem clearly defined with success metrics
□ Train/test split strategy defined
□ Evaluation methodology matches business objective

✅ Data Integrity:
□ No temporal leakage (future → past)
□ No target leakage (answer in features)
□ No preprocessing leakage (fit on all data)
□ No group leakage (related samples split)

✅ Evaluation Setup:
□ Test set completely held out
□ Metrics aligned with business objective
□ Baseline models defined

---

Critical Antipatterns

Category 1: Data Leakage 🚨

1.1 Target Leakage

❌ WRONG: Using "refund_issued" to predict "purchase_fraud"
✅ CORRECT: Only use features available at purchase time

1.2 Temporal Leakage

❌ WRONG: train = df[df['date'] > '2024-06-01']  # Future data
✅ CORRECT: train = df[df['date'] < '2024-06-01']  # Past for training

1.3 Preprocessing Leakage

❌ WRONG: X_scaled = scaler.fit_transform(X); train_test_split(X_scaled)
✅ CORRECT: Split first, then scaler.fit(X_train)

1.4 Group Leakage

❌ WRONG: train_test_split(df)  # Same user in both sets
✅ CORRECT: GroupShuffleSplit(groups=df['user_id'])

1.5 Data Augmentation Leakage

❌ WRONG: augment(X) → train_test_split()
✅ CORRECT: train_test_split() → augment(X_train)

---

Category 2: Evaluation Mistakes ⚠️

2.1 Testing on Training Data

❌ WRONG: evaluate(model, training_data)
✅ CORRECT: evaluate(model, unseen_test_data)

2.2 Metric Misalignment

Business Objective → Appropriate Metric:
- Ranking → NDCG, MRR, MAP
- Imbalanced → F1, Precision@K, AUC-PR
- Balanced → Accuracy, AUC-ROC

2.3 Accuracy Paradox

❌ WRONG: 99% accuracy on 99:1 imbalanced data
✅ CORRECT: Check per-class metrics with classification_report()

2.4 Invalid Time Series CV

❌ WRONG: cross_val_score(model, X, y, cv=5)  # Shuffles time!
✅ CORRECT: TimeSeriesSplit(n_splits=5)

2.5 Hyperparameter Tuning on Test Set

❌ WRONG: grid_search(model, X_test, y_test)
✅ CORRECT: train/validation/test three-way split

---

Category 3: Training Pitfalls 🔧

3.1 Batch Norm Inference Error

❌ WRONG: predictions = model(X_test)  # Still in train mode
✅ CORRECT: model.eval(); with torch.no_grad(): predictions = model(X_test)

3.2 Early Stopping Overfitting

❌ WRONG: EarlyStopping(patience=50)
✅ CORRECT: EarlyStopping(patience=5, min_delta=0.001, restore_best_weights=True)

3.3 Learning Rate Warmup

✅ CORRECT: get_linear_schedule_with_warmup(num_warmup_steps=1000)

3.4 Class Imbalance

❌ WRONG: CrossEntropyLoss()  # Biased toward majority
✅ CORRECT: CrossEntropyLoss(weight=class_weights)

---

Detection Patterns

Leakage Detection

# Check feature-target correlation
correlation = df[features].corrwith(df['target'])
if (correlation.abs() > 0.95).any():
    raise DataLeakageError("Suspiciously high correlation")

# Check temporal ordering
if train['date'].min() > test['date'].max():
    raise TemporalLeakageError("Training on future, testing on past")

# Check group overlap
if train_groups & test_groups:
    raise GroupLeakageError("Overlapping groups")

Mode Check

if model.training:
    raise InferenceModeError("Model in training mode during evaluation")

---

Validation Checklist

Before deployment:

• No data leakage detected
• Test set never seen during training
• Metrics aligned with business objective
• model.eval() called for inference
• Class imbalance handled
• Covariate shift monitoring planned

---

References

상세 예시 및 시나리오는 references/REFERENCE.md 참조.