Recipe

The general classification recipe

Use when: you have a labeled dataset and need to predict a discrete category

Goal: a tuned, honestly evaluated classification model with a documented baseline comparison

Taught in Which algorithm should I use?

1. 1

   Write down exactly what one sample is. Write down exactly what the target label means in practice. Note which features are safe, and which count as leakage.

   Why: this choice quietly decides the split, the metric, and the imbalance plan. Every later step depends on it.

2. 2

   Compute the majority-class accuracy first, or fit a plain, untuned logistic regression. Do this before touching anything more complex.

   Why: without this step, no later score has an honest reference point.

3. 3

   Fit a regularized logistic regression, a random forest, and a gradient-boosted model. Run all three through the identical Pipeline and CV setup.

   Why: different model families capture different structure. It's rarely obvious ahead of time which one this problem rewards.

4. 4

   Compare candidates by cross-validated precision, recall, F1, or ROC-AUC. Pick the metric based on the real cost of a false positive versus a false negative. Decide this before seeing any results.

   Why: accuracy alone is often misleading under class imbalance. Picking the metric after seeing results is a form of over-tuning.

5. 5

   Apply random search first, then refine with Bayesian optimization via Optuna. Do both only on the single strongest candidate from the last step.

   Why: tuning time is short, and spending it on an already-weaker candidate rarely closes a real gap.

6. 6

   Check feature importance or SHAP values as a final leakage sanity check. Then save the entire fitted Pipeline, not the bare model alone.

   Why: saving only the model risks a preprocessing mismatch at inference time.