Development version

You are reading the latest (development) docs. For stable documentation, see stable.

Hyperparameter Tuning Recipes

Recipes are short, copy-paste ready solutions to specific problems. Each recipe answers exactly one question — run it, adapt it, move on.

Recipes vs Tutorials vs Examples

	Recipes	Tutorials	Examples
Purpose	Solve one specific problem	End-to-end conceptual walkthrough	Demonstrate a library feature
Length	5–10 min	20–40 min	5–15 min
Code	Copy-paste ready	Guided walkthrough	Feature-focused snippet

Classification

Recipe	What it solves
Tune RandomForestClassifier	7-parameter joint search with class_weight
Tune LogisticRegression	Solver/penalty compatibility, multi-class
Tune SVC	C/gamma with mandatory scaling in a Pipeline
Tune XGBClassifier	9 params, CPU oversubscription fix, n_jobs=1
Tune LGBMClassifier	num_leaves/max_depth interaction
Tune CatBoostClassifier	Categorical columns, bagging_temperature
Tune HistGradientBoostingClassifier	Fast sklearn GBM, max_leaf_nodes
Tune ExtraTreesClassifier	High variance vs RF, max_features

Regression

Recipe	What it solves
Tune RandomForestRegressor	min_samples_leaf vs min_samples_split, MAE scoring
Tune XGBRegressor	Regression with custom eval_metric
Tune LGBMRegressor	RMSE scoring, min_child_samples
Tune CatBoostRegressor	Regression with ordered boosting
Tune ElasticNet	l1_ratio/alpha interaction, log-uniform

Feature Selection

Recipe	What it solves
Select Features on 50+ Columns	Threshold strategy for large feature sets
Combine Feature Selection + Hyperparameter Tuning	Two-stage: select features, then retune
Use a Custom Scorer for Feature Selection	Wrap roc_auc with a feature-count penalty
Feature Selection with Cross-Validation	Per-fold selection to avoid leakage

Pipelines

Recipe	What it solves
Tune a Preprocessing + Estimator Pipeline	StandardScaler + SVC with step prefix
Tune a ColumnTransformer Pipeline	Mixed types, preprocessor__... param paths
Tune Imputer Strategy as a Hyperparameter	Search over mean/median/most_frequent
Tune Polynomial Features Degree	Include degree and interaction-only as params

Scoring Metrics

Recipe	What it solves
Tune for F1 Score (Binary)	scoring="f1", thresholds, confusion matrix
Tune for ROC-AUC	scoring="roc_auc", probability calibration
Tune for Balanced Accuracy	Class imbalance, class_weight as a param
Tune for MAE (Regression)	scoring="neg_mean_absolute_error"
Tune for RMSE (Regression)	Custom neg_root_mean_squared_error scorer

Integrations

Recipe	What it solves
Log Every Candidate to MLflow	MlflowCallback, nested runs, param logging
Parallelize with Joblib	n_jobs, loky vs threading backend
Run in a Jupyter Notebook	verbose=True, tqdm progress, plot inline

Advanced

Recipe	What it solves
Seed a Search with Known-Good Params	warm_start_configs to skip cold start
Stop Early When Fitness Plateaus	ConsecutiveStopping callback setup
Stop After a Time Budget	TimerStopping for wall-clock limits
Resume a Stopped Search	Save/load state, continue from checkpoint
Write a Custom Scoring Function	make_scorer + unsupervised metrics

See Also

• Tutorials — end-to-end model walkthroughs with explanations
• Examples — focused library feature demonstrations
• API Reference — full parameter documentation
• Common Mistakes — pitfalls to avoid