model-evaluation

Model Evaluation

When to Use

When assessing whether a model is good enough to deploy, fairly represents all user groups, and won't degrade in production.

Core Jobs

1. Choose the Right Metrics

Match metric to problem type:

• Binary classification: AUC-ROC, F1, precision, recall (choose based on cost of FP vs FN)
• Multi-class: macro/weighted F1, confusion matrix
• Regression: RMSE, MAE, MAPE (use MAE when outliers shouldn't dominate)
• Ranking: NDCG, MRR
• Generation (LLM): BLEU/ROUGE (weak), human eval, LLM-as-judge

Business metric matters more than ML metric — always connect model performance to business outcome.

2. Evaluate Across Segments

Don't report only aggregate metrics. Slice by:

• User demographics (age, region, language)
• Data subgroups (product category, request type)
• Time (recent vs older data — look for drift)
• Edge cases (short inputs, rare labels)

3. Bias and Fairness Checks

• Equal opportunity: equal TPR across groups?
• Demographic parity: equal positive prediction rates?
• Use tools: Fairlearn, IBM AI Fairness 360

4. Pre-Production Validation

• Performance on holdout test set
• Performance on recent data (last 30 days)
• Latency at P50/P95/P99 (meets SLA?)
• Memory footprint (fits in serving environment?)
• Slice analysis (no group significantly underperforms)
• Shadow mode test (run alongside current system)

Key Outputs

• Evaluation report with chosen metrics and rationale
• Slice analysis (breakdown by key segments)
• Bias/fairness assessment
• Pre-production validation checklist

Anti-Patterns

• Optimizing for accuracy on imbalanced datasets
• Never slicing results by subgroup
• Declaring a model "ready" without latency testing
• Using test set for model selection (leakage)