Ottimizzare con GridSearchCV

Some model settings are not learned from data; you pick them before training. These are hyperparameters, and good choices matter a lot.

Tweaking one value, re-running, and eyeballing the score by hand wastes time and misses better combos. Let a search do it. 🔍

GridSearchCV tries every combination in a grid you define and uses cross-validation to score each one fairly.

You list each hyperparameter and the values to try in a dictionary. Every value pairing becomes one candidate to test.

To target a pipeline step, prefix the name with the step plus a double underscore. This naming points the search at the right knob.

Wrap your estimator and grid, then call fit. It trains and scores every combination across the folds for you.

After fitting, best_params_ tells you which combination won, and best_score_ shows its average cross-validated score.

The search refits the top combo on all training data. Just call predict on the search object to use that best_estimator.

By default it optimizes accuracy. Set scoring to f1 or roc_auc so the search chases the metric your problem cares about.

The grid grows fast: combinations times folds equals fits. When it explodes, try RandomizedSearchCV to sample instead.

Because the whole pipeline is searched, prep refits inside each fold. Tuning stays honest with zero leakage sneaking in.

To tune a pipeline step named model, how do you key the parameter?

You can now let GridSearchCV test parameter combos with cross-validation and hand you the best one. Next, saving your trained pipeline. 💾