Hyperparameter tuning by randomized-search

In the previous notebook, we showed how to use a grid-search approach to search for the best hyperparameters maximizing the generalization performance of a predictive model.

However, a grid-search approach has limitations. It does not scale well when the number of parameters to tune increases. Also, the grid imposes a regularity during the search which might miss better parameter values between two consecutive values on the grid.

In this notebook, we present a different method to tune hyperparameters called randomized search.

Our predictive model

Let us reload the dataset as we did previously:

import pandas as pd

adult_census = pd.read_csv("../datasets/adult-census.csv")

We extract the column containing the target.

target_name = "class"
target = adult_census[target_name]
target

0         <=50K
1         <=50K
2          >50K
3          >50K
4         <=50K
          ...  
48837     <=50K
48838      >50K
48839     <=50K
48840     <=50K
48841      >50K
Name: class, Length: 48842, dtype: object

We drop from our data the target and the "education-num" column which duplicates the information with "education" columns.

data = adult_census.drop(columns=[target_name, "education-num"])
data.head()

	age	workclass	education	marital-status	occupation	relationship	race	sex	capital-gain	capital-loss	hours-per-week	native-country
0	25	Private	11th	Never-married	Machine-op-inspct	Own-child	Black	Male	0	0	40	United-States
1	38	Private	HS-grad	Married-civ-spouse	Farming-fishing	Husband	White	Male	0	0	50	United-States
2	28	Local-gov	Assoc-acdm	Married-civ-spouse	Protective-serv	Husband	White	Male	0	0	40	United-States
3	44	Private	Some-college	Married-civ-spouse	Machine-op-inspct	Husband	Black	Male	7688	0	40	United-States
4	18	?	Some-college	Never-married	?	Own-child	White	Female	0	0	30	United-States

Once the dataset is loaded, we split it into a training and testing sets.

from sklearn.model_selection import train_test_split

data_train, data_test, target_train, target_test = train_test_split(
    data, target, random_state=42
)

We create the same predictive pipeline as done for the grid-search section.

from sklearn.compose import ColumnTransformer
from sklearn.preprocessing import OrdinalEncoder
from sklearn.compose import make_column_selector as selector

categorical_columns_selector = selector(dtype_include=object)
categorical_columns = categorical_columns_selector(data)

categorical_preprocessor = OrdinalEncoder(
    handle_unknown="use_encoded_value", unknown_value=-1
)
preprocessor = ColumnTransformer(
    [("cat_preprocessor", categorical_preprocessor, categorical_columns)],
    remainder="passthrough",
)

from sklearn.ensemble import HistGradientBoostingClassifier
from sklearn.pipeline import Pipeline

model = Pipeline(
    [
        ("preprocessor", preprocessor),
        (
            "classifier",
            HistGradientBoostingClassifier(random_state=42, max_leaf_nodes=4),
        ),
    ]
)

model

Pipeline(steps=[('preprocessor',
                 ColumnTransformer(remainder='passthrough',
                                   transformers=[('cat_preprocessor',
                                                  OrdinalEncoder(handle_unknown='use_encoded_value',
                                                                 unknown_value=-1),
                                                  ['workclass', 'education',
                                                   'marital-status',
                                                   'occupation', 'relationship',
                                                   'race', 'sex',
                                                   'native-country'])])),
                ('classifier',
                 HistGradientBoostingClassifier(max_leaf_nodes=4,
                                                random_state=42))])

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

Tuning using a randomized-search

With the GridSearchCV estimator, the parameters need to be specified explicitly. We already mentioned that exploring a large number of values for different parameters quickly becomes untractable.

Instead, we can randomly generate the parameter candidates. Indeed, such approach avoids the regularity of the grid. Hence, adding more evaluations can increase the resolution in each direction. This is the case in the frequent situation where the choice of some hyperparameters is not very important, as for the hyperparameter 2 in the figure below.

Indeed, the number of evaluation points needs to be divided across the two different hyperparameters. With a grid, the danger is that the region of good hyperparameters may fall between lines of the grid. In the figure such region is aligned with the grid given that hyperparameter 2 has a weak influence. Rather, stochastic search samples the hyperparameter 1 independently from the hyperparameter 2 and find the optimal region.

The RandomizedSearchCV class allows for such stochastic search. It is used similarly to the GridSearchCV but the sampling distributions need to be specified instead of the parameter values. For instance, we can draw candidates using a log-uniform distribution because the parameters we are interested in take positive values with a natural log scaling (.1 is as close to 1 as 10 is).

Note

Random search (with RandomizedSearchCV) is typically beneficial compared to grid search (with GridSearchCV) to optimize 3 or more hyperparameters.

We now optimize 3 other parameters in addition to the ones we optimized in the notebook presenting the GridSearchCV:

• l2_regularization: it corresponds to the strength of the regularization;

• min_samples_leaf: it corresponds to the minimum number of samples required in a leaf;

• max_bins: it corresponds to the maximum number of bins to construct the histograms.

We recall the meaning of the 2 remaining parameters:

• learning_rate: it corresponds to the speed at which the gradient-boosting corrects the residuals at each boosting iteration;

• max_leaf_nodes: it corresponds to the maximum number of leaves for each tree in the ensemble.

Note

scipy.stats.loguniform can be used to generate floating numbers. To generate random values for integer-valued parameters (e.g. min_samples_leaf) we can adapt it as follows:

from scipy.stats import loguniform


class loguniform_int:
    """Integer valued version of the log-uniform distribution"""

    def __init__(self, a, b):
        self._distribution = loguniform(a, b)

    def rvs(self, *args, **kwargs):
        """Random variable sample"""
        return self._distribution.rvs(*args, **kwargs).astype(int)

Now, we can define the randomized search using the different distributions. Executing 10 iterations of 5-fold cross-validation for random parametrizations of this model on this dataset can take from 10 seconds to several minutes, depending on the speed of the host computer and the number of available processors.

%%time
from sklearn.model_selection import RandomizedSearchCV

param_distributions = {
    "classifier__l2_regularization": loguniform(1e-6, 1e3),
    "classifier__learning_rate": loguniform(0.001, 10),
    "classifier__max_leaf_nodes": loguniform_int(2, 256),
    "classifier__min_samples_leaf": loguniform_int(1, 100),
    "classifier__max_bins": loguniform_int(2, 255),
}

model_random_search = RandomizedSearchCV(
    model,
    param_distributions=param_distributions,
    n_iter=10,
    cv=5,
    verbose=1,
)
model_random_search.fit(data_train, target_train)

Fitting 5 folds for each of 10 candidates, totalling 50 fits

CPU times: user 20.1 s, sys: 267 ms, total: 20.4 s
Wall time: 20.4 s

RandomizedSearchCV(cv=5,
                   estimator=Pipeline(steps=[('preprocessor',
                                              ColumnTransformer(remainder='passthrough',
                                                                transformers=[('cat_preprocessor',
                                                                               OrdinalEncoder(handle_unknown='use_encoded_value',
                                                                                              unknown_value=-1),
                                                                               ['workclass',
                                                                                'education',
                                                                                'marital-status',
                                                                                'occupation',
                                                                                'relationship',
                                                                                'race',
                                                                                'sex',
                                                                                'native-country'])])),
                                             ('classifier',
                                              HistGradientBoostingC...
                   param_distributions={'classifier__l2_regularization': <scipy.stats._distn_infrastructure.rv_continuous_frozen object at 0x711c4be1fdf0>,
                                        'classifier__learning_rate': <scipy.stats._distn_infrastructure.rv_continuous_frozen object at 0x711c4af777c0>,
                                        'classifier__max_bins': <__main__.loguniform_int object at 0x711c4af7e100>,
                                        'classifier__max_leaf_nodes': <__main__.loguniform_int object at 0x711c4af71d00>,
                                        'classifier__min_samples_leaf': <__main__.loguniform_int object at 0x711c4af77310>},
                   verbose=1)

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

Then, we can compute the accuracy score on the test set.

accuracy = model_random_search.score(data_test, target_test)

print(f"The test accuracy score of the best model is {accuracy:.2f}")

The test accuracy score of the best model is 0.86

from pprint import pprint

print("The best parameters are:")
pprint(model_random_search.best_params_)

The best parameters are:
{'classifier__l2_regularization': 0.0042545515275358116,
 'classifier__learning_rate': 0.029495676031538985,
 'classifier__max_bins': 30,
 'classifier__max_leaf_nodes': 35,
 'classifier__min_samples_leaf': 65}

We can inspect the results using the attributes cv_results as we did previously.

# get the parameter names
column_results = [f"param_{name}" for name in param_distributions.keys()]
column_results += ["mean_test_score", "std_test_score", "rank_test_score"]

cv_results = pd.DataFrame(model_random_search.cv_results_)
cv_results = cv_results[column_results].sort_values(
    "mean_test_score", ascending=False
)


def shorten_param(param_name):
    if "__" in param_name:
        return param_name.rsplit("__", 1)[1]
    return param_name


cv_results = cv_results.rename(shorten_param, axis=1)
cv_results

	l2_regularization	learning_rate	max_leaf_nodes	min_samples_leaf	max_bins	mean_test_score	std_test_score	rank_test_score
9	0.004255	0.029496	35	65	30	0.854167	0.002697	1
5	0.007172	1.138836	3	37	16	0.847643	0.003098	2
7	25.255018	0.609597	83	26	16	0.845759	0.002760	3
6	0.000055	0.987865	168	2	121	0.835604	0.001594	4
4	1.095145	1.132941	90	67	4	0.807922	0.001426	5
0	0.000001	0.001246	9	71	139	0.758947	0.000013	6
1	0.000009	0.002443	5	7	6	0.758947	0.000013	6
2	0.001968	0.001487	37	6	54	0.758947	0.000013	6
3	0.000002	0.014813	2	1	3	0.758947	0.000013	6
8	0.002385	5.011593	11	9	14	0.726024	0.034718	10

Keep in mind that tuning is limited by the number of different combinations of parameters that are scored by the randomized search. In fact, there might be other sets of parameters leading to similar or better generalization performances but that were not tested in the search. In practice, a randomized hyperparameter search is usually run with a large number of iterations. In order to avoid the computation cost and still make a decent analysis, we load the results obtained from a similar search with 500 iterations.

# model_random_search = RandomizedSearchCV(
#     model, param_distributions=param_distributions, n_iter=500,
#     n_jobs=2, cv=5)
# model_random_search.fit(data_train, target_train)
# cv_results =  pd.DataFrame(model_random_search.cv_results_)
# cv_results.to_csv("../figures/randomized_search_results.csv")

cv_results = pd.read_csv(
    "../figures/randomized_search_results.csv", index_col=0
)

(
    cv_results[column_results]
    .rename(shorten_param, axis=1)
    .sort_values("mean_test_score", ascending=False)
)

	l2_regularization	learning_rate	max_leaf_nodes	min_samples_leaf	max_bins	mean_test_score	std_test_score	rank_test_score
208	0.011775	0.076653	24	2	155	0.871393	0.001588	1
343	0.000404	0.244503	15	15	229	0.871339	0.002741	2
21	4.994918	0.077047	53	7	192	0.870793	0.001993	3
328	2.036232	0.224702	28	49	236	0.869837	0.000808	4
327	4.733808	0.036786	61	5	241	0.869673	0.002417	5
...	...	...	...	...	...	...	...	...
232	0.000097	9.976823	28	5	3	0.448205	0.253714	496
413	0.000001	8.828574	64	1	144	0.448205	0.253714	497
344	0.000003	7.091079	5	1	95	0.448205	0.253714	497
200	0.000444	6.236325	2	2	30	0.344629	0.207156	499
357	0.000026	3.075318	3	68	31	0.241053	0.000013	500

500 rows × 8 columns

In this case the top performing models have test scores with a high overlap between each other, meaning that indeed, the set of parameters leading to the best generalization performance is not unique.

In this notebook, we saw how a randomized search offers a valuable alternative to grid-search when the number of hyperparameters to tune is more than two. It also alleviates the regularity imposed by the grid that might be problematic sometimes.

In the following, we will see how to use interactive plotting tools to explore the results of large hyperparameter search sessions and gain some insights on range of parameter values that lead to the highest performing models and how different hyperparameter are coupled or not.