Tabular ML workflow

Decision checklist from request to monitored use.

Quick reference. Print friendly.

1. Frame

Decision: what action changes? Unit: what does one row mean? Target: what outcome and horizon? Prediction time: when must inputs exist? Baseline: what simple rule sets minimum value?

2. Audit data

Check label origin, missingness, duplicates, impossible values, time coverage, population coverage, and features unavailable at prediction time.

3. Split first

Match deployment: random for independent stable rows, grouped for repeated entities, temporal for future prediction. Keep test data untouched until choices freeze.

4. Build one pipeline

pipeline = make_pipeline(preprocessor, estimator)
pipeline.fit(X_train, y_train)
scores = cross_validate(pipeline, X_train, y_train, cv=splitter)

5. Evaluate decisions

Choose metrics from false-positive and false-negative costs. Compare baseline. Inspect fold variation, train-validation gap, threshold behavior, and concrete errors.

6. Decide and monitor

Record purpose, data, split, metrics, limitations, owners, fallback, drift signals, outcome delay, and retraining trigger. Choose go, bounded pilot, or no-go.

Never let model choice outrun problem framing, data validity, or evaluation design.

GlossaryStart course