randomized_search

A simple randomized search on hyperparameters.

In contrast to grid search, not all parameter values are tried out, but rather a fixed number of parameter settings is sampled from the specified distributions. The number of parameter settings that are tried is specified in the n_iter parameter.

Method call formatMethod call format

randomized_search(param_distributions,
                  X,
                  y=None,
                  cv=3,
                  n_iter=10,
                  partition_random_seed=0,
                  calc_cv_statistics=True,
                  search_by_train_test_split=True,
                  refit=True,
                  shuffle=True,
                  stratified=None,
                  train_size=0.8,
                  verbose=True,
                  log_cout=sys.stdout,
                  log_cerr=sys.stderr)

ParametersParameters

param_distributionsparam_distributions

DescriptionDescription

Dictionary with parameters names (string) as keys and distributions or lists of parameter settings to try. Distributions must provide a rvs method for sampling (such as those from scipy.stats.distributions).

If a list is given, it is sampled uniformly.

Possible types:

dict

Default value

Required parameter

XX

DescriptionDescription

The description is different for each group of possible types.

Possible types

Default value

Required parameter

yy

DescriptionDescription

The target variables (in other words, the objects' label values) for the training dataset.

Must be in the form of a one- or two- dimensional array. The type of data in the array depends on the machine learning task being solved:

• Regression and ranking — One-dimensional array of numeric values.

• Multiregression - Two-dimensional array of numeric values. The first index is for a dimension, the second index is for an object.

• Binary classification
  One-dimensional array containing one of:

  • Booleans, integers or strings that represent the labels of the classes (only two unique values).

  • Numeric values.
    The interpretation of numeric values depends on the selected loss function:

    • Logloss — The value is considered a positive class if it is strictly greater than the value of the target_border training parameter. Otherwise, it is considered a negative class.
    • CrossEntropy — The value is interpreted as the probability that the dataset object belongs to the positive class. Possible values are in the range [0; 1].

• Multiclassification — One-dimensional array of integers or strings that represent the labels of the classes.

• Multi label classification
  Two-dimensional array. The first index is for a label/class, the second index is for an object.

  Possible values depend on the selected loss function:

  • MultiLogloss — Only {0, 1} or {False, True} values are allowed that specify whether an object belongs to the class corresponding to the first index.
  • MultiCrossEntropy — Numerical values in the range [0; 1] that are interpreted as the probability that the dataset object belongs to the class corresponding to the first index.

Possible types

• list
• numpy.ndarray
• pandas.DataFrame
• pandas.Series

Default value

None

Supported processing units

CPU and GPU

cvcv

DescriptionDescription

The cross-validation splitting strategy.
The interpretation of this parameter depends on the input data type:

• None — Use the default three-fold cross-validation.
• int — The number of folds in a (Stratified)KFold
• object — One of the scikit-learn Splitter Classes with the split method.
• An iterable yielding train and test splits as arrays of indices.

Possible types

• int
• scikit-learn splitter object
• cross-validation generator
• iterable

Default value

None

n_itern_iter

DescriptionDescription

The number of parameter settings that are sampled. This parameter trades off runtime vs quality of the solution.

Possible types

int

Default value

0

partition_random_seedpartition_random_seed

DescriptionDescription

Use this as the seed value for random permutation of the data.
The permutation is performed before splitting the data for cross-validation.
Each seed generates unique data splits.

Possible types

int

Default value

0

calc_cv_statisticscalc_cv_statistics

DescriptionDescription

Estimate the quality by using cross-validation with the best of the found parameters. The model is fitted using these parameters.
This option can be enabled if the search_by_train_test_split parameter is set to True.

Possible types

bool

Default value

True

search_by_train_test_splitsearch_by_train_test_split

DescriptionDescription

Split the source dataset into train and test parts. Models are trained on the train part, while parameters are compared by the loss function score on the test dataset.

It is recommended to enable this option for large datasets and disable it for the small ones.

Possible types

bool

Default value

True

refitrefit

DescriptionDescription

Refit an estimator using the best-found parameters on the whole dataset.

Possible types

bool

Default value

True

shuffleshuffle

DescriptionDescription

Shuffle the dataset objects before splitting into folds.

Possible types

bool

Default value

True

stratifiedstratified

DescriptionDescription

Perform stratified sampling. True for classification and False otherwise.

Possible types

bool

Default value

None

train_sizetrain_size

DescriptionDescription

The proportion of the dataset to include in the train split.

Possible values are in the range [0;1].

Possible types

float

Default value

0.8

verboseverbose

DescriptionDescription

The purpose of this parameter depends on the type of the given value:

• int — The frequency of iterations to print the information to stdout.
• bool — Print the information to stdout on every iteration (if set to “True”) or disable any logging (if set to “False”).

Possible types

• bool
• int

Default value

True

plotplot

DescriptionDescription

Draw train and evaluation metrics for every set of parameters in Jupyter Jupyter Notebook.

Possible types

bool

Default value
False

log_coutlog_cout

Output stream or callback for logging.

Possible types

• callable Python object
• python object providing the write() method

Default value

sys.stdout

log_cerrlog_cerr

Error stream or callback for logging.

Possible types

• callable Python object
• python object providing the write() method

Default value

sys.stderr

Return valueReturn value

Dict with two fields:

• params — dict of best-found parameters.
• cv_results — dict or pandas.core.frame.DataFrame with cross-validation results. Сolumns are: test-error-mean, test-error-std, train-error-mean, train-error-std.

ExamplesExamples

from catboost import CatBoost

train_data = [[1, 4, 5, 6],
              [4, 5, 6, 7],
              [30, 40, 50, 60],
              [20, 30, 70, 60],
              [10, 80, 40, 30],
              [10, 10, 20, 30]]
train_labels = [10, 20, 30, 15, 10, 25]
model = CatBoost()

grid = {'learning_rate': [0.03, 0.1],
        'depth': [4, 6, 10],
        'l2_leaf_reg': [1, 3, 5, 7, 9]}

randomized_search_result = model.randomized_search(grid,
                                                   X=train_data,
                                                   y=train_labels,
                                                   plot=True)

The following is a chart plotted with Jupyter Notebook for the given example.