Hyperparameter Tuning

In Machine Learning, there is a crucial difference between Parameters and Hyperparameters:

• Parameters: Learned by the model during training (e.g., weights in a regression or coefficients in a neural network).
• Hyperparameters: Set by the engineer before training starts (e.g., the depth of a Decision Tree or the number of neighbors in KNN).

Hyperparameter Tuning is the automated search for the best combination of these settings to minimize error.

1. Why Tune Hyperparameters?

Most algorithms come with default settings that work reasonably well, but they are rarely optimal for your specific data. Proper tuning can often bridge the gap between a mediocre model and a state-of-the-art one.

2. GridSearchCV: The Exhaustive Search

GridSearchCV takes a predefined list of values for each hyperparameter and tries every possible combination.

• Pros: Guaranteed to find the best combination within the provided grid.
• Cons: Computationally expensive. If you have 5 parameters with 5 values each, you must train the model $5^5 = 3,125$ times.

from sklearn.model_selection import GridSearchCV
from sklearn.ensemble import RandomForestClassifier

param_grid = {
    'n_estimators': [50, 100, 200],
    'max_depth': [None, 10, 20],
    'min_samples_split': [2, 5]
}

grid_search = GridSearchCV(RandomForestClassifier(), param_grid, cv=5)
grid_search.fit(X_train, y_train)

print(f"Best Parameters: {grid_search.best_params_}")

3. RandomizedSearchCV: The Efficient Alternative

Instead of trying every combination, RandomizedSearchCV picks a fixed number of random combinations from a distribution.

• Pros: Much faster than GridSearch. It often finds a result almost as good as GridSearch in a fraction of the time.
• Cons: Not guaranteed to find the absolute best "peak" in the parameter space.

from sklearn.model_selection import RandomizedSearchCV
from scipy.stats import randint

param_dist = {
    'n_estimators': randint(50, 500),
    'max_depth': [None, 10, 20, 30, 40, 50],
}

random_search = RandomizedSearchCV(RandomForestClassifier(), param_dist, n_iter=20, cv=5)
random_search.fit(X_train, y_train)

4. Advanced: Successive Halving

For massive datasets, even Random Search is slow. Scikit-Learn offers HalvingGridSearch. It trains all combinations on a small amount of data, throws away the bottom 50%, and keeps "promising" candidates for the next round with more data.

5. Avoiding the Validation Trap

If you tune your hyperparameters using the Test Set, you are "leaking" information. The model will look great on that test set, but fail on new data.

The Solution: Use Nested Cross-Validation or ensure that your GridSearchCV only uses the Training Set (it will internally split the training data into smaller validation folds).

6. Tuning Strategy Summary

Method	Best for...	Resource Usage
Manual Tuning	Initial exploration / small models	Low
GridSearch	Small number of parameters	High
RandomSearch	Many parameters / large search space	Moderate
Halving Search	Large datasets / expensive training	Low-Moderate

References for More Details

• Sklearn Tuning Guide: Deep dive into HalvingGridSearchCV and custom scoring.

---

Now that your model is fully optimized and tuned, it's time to evaluate its performance using metrics that go beyond simple "Accuracy."