Hyperparameter Tuning

Hyperparameter Tuning

Overview

Tuning squeezes the last 5-15% out of a model — but done carelessly it overfits the validation set and leaks preprocessing. The rules: tune the whole pipeline inside cross-validation, search smart (not grid), and keep a final untouched test set.

When to use

• A reasonable baseline exists and you want to improve it.
• You need to pick model complexity (depth, regularization, lr).

Strategy selection

Optuna pattern (leakage-safe, prunes bad trials)

import optuna
from sklearn.model_selection import cross_val_score, StratifiedKFold

cv = StratifiedKFold(5, shuffle=True, random_state=42)

def objective(trial):
    params = {
        "clf__learning_rate": trial.suggest_float("lr", 1e-3, 0.3, log=True),
        "clf__max_depth": trial.suggest_int("max_depth", 3, 12),
        "clf__l2_regularization": trial.suggest_float("l2", 1e-3, 10, log=True),
    }
    model.set_params(**params)
    scores = cross_val_score(model, X_train, y_train, cv=cv, scoring="roc_auc")
    return scores.mean()

study = optuna.create_study(direction="maximize",
                            sampler=optuna.samplers.TPESampler(seed=42))
study.optimize(objective, n_trials=50, timeout=1800)
print(study.best_params, study.best_value)

Note the clf__ prefix — you're tuning the estimator inside the pipeline, so preprocessing re-fits per fold.

Search-space design

• Sample learning rates and regularization on a log scale.
• Start wide, then narrow around the best region in a second study.
• Tie n_estimators to early stopping rather than tuning it directly.
• Fix the seed in the sampler for reproducible studies.

Budget management

• Set timeout and n_trials ceilings; use Optuna pruning to kill hopeless trials early.
• Tune on a representative subsample first to find the region, then refine on full data.
• Don't tune dozens of params — pick the 3-5 that matter for your model family.

Pitfalls

• Nested leakage: preprocessing fit outside CV, then tuned — inflates scores. Tune the pipeline.
• Tuning on the test set — only ever touch train/validation; report final number on the holdout once.
• Over-tuning to a tiny validation set → great CV, worse production. Prefer simpler models + regularization.
• Comparing tuned-vs-untuned across different CV splits — keep the split fixed.

Hand-off

Best params + a refit pipeline, logged via experiment-tracking and scored once on the holdout via model-evaluation.