Supervised Learning Deep Dive

Project: projects/sota-supervised-learning-showcase

Why This Deep Dive

Use this project when you want one end-to-end supervised workflow that covers:

• class imbalance handling,
• binary, multiclass, multilabel, and multioutput tasks,
• threshold-aware evaluation,
• model selection curves,
• regression baselines.

Quickstart

cd projects/sota-supervised-learning-showcase
make sync
make run
make test

Optional model boosters:

make sync-boosting

What You Should Inspect

Artifact	Why it matters
artifacts/binary_metrics.csv	compare precision/recall/F1 under imbalance strategies
artifacts/pr_curves.csv and artifacts/roc_curves.csv	check threshold behavior and score ranking quality
artifacts/classification_benchmark.csv	compare model families (tree/boosting/ensemble)
artifacts/validation_curve.csv and artifacts/learning_curve.csv	diagnose underfitting vs overfitting
artifacts/regression_benchmark.csv	verify advanced regressors beat simple baselines

Example: Fast Artifact Inspection

cd projects/sota-supervised-learning-showcase
python - <<'PY'
import pandas as pd
bench = pd.read_csv("artifacts/classification_benchmark.csv")
print(bench.sort_values("pr_auc", ascending=False).head(5).to_string(index=False))
PY

Example: Threshold Decision Framing

# Pseudocode for converting model scores into policy:
# if score >= threshold:
#     approve_action()
# else:
#     hold_or_review()
#
# Choose threshold by business tradeoff:
# - higher threshold -> higher precision, lower recall
# - lower threshold -> higher recall, lower precision

How To Interpret Outputs

1. If ROC-AUC is strong but PR-AUC is weak, class imbalance is likely dominating practical quality.
2. If train score is high and validation score plateaus early, increase regularization or simplify model.
3. If ensemble gains are marginal, prefer simpler models for explainability and maintenance.
4. If baseline regressors are close to advanced models, feature quality may be the bottleneck.

Next Step

Move to MLOps Drift Deep Dive to operationalize similar supervised models in a production-style workflow.