Stop 07 · Model Evaluation & TuningTopics · 5/8

Hyperparameter Tuning (Grid / Random Search)

In short

Tree depth, learning rate, number of neighbours: these settings aren't learned from data, you choose them. Tuning is the organised way to find good values instead of guessing.

Systematically searching for the best model settings — grid search tries every combination in a defined range, random search samples combinations randomly, often nearly as effective and much faster.

The mascot tastes biryani from one of six small pots lined up on a kitchen counter.

Think of it like this

Perfecting the biryani

You want the best biryani. Rice soaking time and dum time are your knobs. Grid search cooks every combination on your list. Random search cooks a few random combinations from a wider range. Either way, a panel of friends tastes each pot fairly before you pick.

Watch it happen

Two ways to search

Grid search

Tries every combination you list
Cost grows fast with each parameter
Good for 2 or 3 small lists

Random search

Samples combinations from ranges
You fix the budget with n_iter
Covers wide ranges more cheaply

Use grid search for a small, known set of values. Use random search when you have many parameters or wide ranges.

The big ideas

What you need to know

Parameters vs hyperparameters

Weights are learned during fit. Hyperparameters, like max_depth, are set by you before training starts.

Score with cross-validation

Each candidate is scored on several folds of training data, so one lucky split doesn't fool you.

Keep the test set out

Tune only on training data. Touch the test set once, after tuning is finished.

Budget matters

3 values for 4 parameters with 5 folds is 405 fits. Random search lets you cap that.

How it works

What the search does for each candidate

1
Pick settings
Take the next combination from the grid, or sample one at random.
2
Cross-validate
Train on k-1 folds, score on the held-out fold, repeat k times.
3
Average the score
Store the mean score for this combination.
4
Refit the winner
After all candidates, refit the best settings on all training data.

See it in code

Walk through the code

python

1from scipy.stats import randint2from sklearn.datasets import load_breast_cancer3from sklearn.ensemble import RandomForestClassifier4from sklearn.model_selection import GridSearchCV, RandomizedSearchCV5 6X, y = load_breast_cancer(return_X_y=True)7rf = RandomForestClassifier(random_state=0)8 9grid = GridSearchCV(rf, {"max_depth": [3, 5, None], "n_estimators": [100, 300]}, cv=5)10grid.fit(X, y)11 12rand = RandomizedSearchCV(rf, {"max_depth": randint(2, 20), "n_estimators": randint(50, 500)},13                          n_iter=10, cv=5, random_state=0)14rand.fit(X, y)15print(grid.best_params_, rand.best_params_)

1/3Grid: 3 depths times 2 sizes = 6 combinations, each fitted 5 times.

Both searches cross-validate every candidate. In a real project, run them on the training split only.

Save yourself the pain

Common mistakes

Oops

Tuning against the test set and then reporting that same test score.

Do this instead

Tune with cross-validation on training data. The test score is only honest if it never guided a choice.

Oops

Building a giant grid with six parameters and ten values each.

Do this instead

Start with random search over wide ranges, then try a small grid around the best region.

Your mission

Grid vs random on a budget

~30 min

Quick check

Why does GridSearchCV use cross-validation for each candidate?

Go deeper

1Run a grid search over 2-3 hyperparameters for a model you've trained.
2Compare grid search results to a random search over the same range.
3Learn why tuning must always be validated on held-out data, never the training set.