Introduction

Hyperparameter tuning is a technique used to improve the accuracy of models. If you make a model with scikit-learn and do not set parameters, it will be set with appropriate complexity.

What are hyperparameters?

Parameters that are specified before training to determine the training method, speed, and model complexity.

Method (type)

Manual
Grid search
Random search

Source: http://www.jmlr.org/papers/volume13/bergstra12a/bergstra12a.pdf

Grid search

It is a method to decide candidates for each parameter and try all combinations of them. Since all are tried, it is not possible to increase the number of parameter candidates.

Random search

It is a method of deciding a candidate for each parameter and repeating a random combination of parameters n times. It may not be possible to search for a better combination of parameters because we do not try all of them.

Parameter combination

import numpy as np

params_list01 = [1, 3, 5, 7]
params_list02 = [1, 2, 3, 4, 5]

#Grid search
grid_search_params = []
for p1 in params_list01:
    for p2 in params_list02:
        grid_search_params.append(p1, p2)
# append():Add an element to the end of the list

#Random search
random_search_params = []
count = 10
for i in range(count):
    p1 = np.random.choice(params_list01)  # random.choice():Get the contents of the array randomly
    p2 = np.random.choice(params_list02)
    random_search_params.append(p1, p2)

scikit-learn

Click here for scikit-learn reference

from sklearn.model_selection import GridSearchCV, RandomizedSearchCV
params = {
    "max_depth": [2, 4, 6, 8, None],
    "n_estimators": [50,100,200,300,400,500],
    "max_features": range(1, 11),
    "min_samples_split": range(2, 11),
    "min_samples_leaf": range(1, 11)
}

#Grid search
gscv = GridSearchCV(RandomForestRegressor(), params, cv=3, n_jobs=-1, verbose=1)
gscv.fit(X_train_valid, y_train_valid)
            
print("Best score: {}".format(gscv.best_score_))
print("Best parameters: {}".format(gscv.best_params_))

#Random search
rscv = RandomizedSearchCV(RandomForestRegressor(), params, cv=3, n_iter=10, n_jobs=-1, verbose=1)
rscv.fit(X_train_valid, y_train_valid)
            
print("Best score: {}".format(rscv.best_score_))
print("Best parameters: {}".format(rscv.best_params_))

in conclusion

When asking which one should be adopted, a random search is performed. It seems that a good combination of parameters can be found efficiently.

reference

Book: Data analysis technology that wins with Kaggle (Technical Review)

What is hyperparameter tuning?