cv

cv(pool=None,
   params=None,
   dtrain=None,
   iterations=None,
   num_boost_round=None,
   fold_count=3,
   nfold=None,
   inverted=False,
   partition_random_seed=0,
   seed=None,
   shuffle=True,
   logging_level=None,
   stratified=None,
   as_pandas=True,
   metric_period=None,
   verbose=None,
   verbose_eval=None,
   plot=False,
   early_stopping_rounds=None,
   folds=None,
   type='Classical',
   return_models=False)

Purpose

Perform cross-validation on the dataset.

The dataset is split into N folds. N–1 folds are used for training, and one fold is used for model performance estimation. N models are updated on each iteration K. Each model is evaluated on its' own validation dataset on each iteration. This produces N metric values on each iteration K.

The cv function calculates the average of these N values and the standard deviation. Thus, these two values are returned on each iteration.

If the dataset contains group identifiers, all objects from one group are added to the same fold when partitioning is performed.

Parameters

pool

Alias: dtrain

Description

The input dataset to cross-validate.

Possible types

Pool

Default value

Required parameter

params

Description

The list of parameters to start training with.

Note

  • The following parameters are not supported in cross-validation mode: save_snapshot,
    --snapshot-file
    , snapshot_interval.
  • The behavior of the overfitting detector is slightly different from the training mode. Only one metric value is calculated at each iteration in the training mode, while fold_count metric values are calculated in the cross-validation mode. Therefore, all fold_count values are averaged and the best iteration is chosen based on the average metric value at each iteration.

Possible types

dict

Default value

Required parameter

iterations

Aliases: num_boost_round, n_estimators, num_trees

Description

The maximum number of trees that can be built when solving machine learning problems.

When using other parameters that limit the number of iterations, the final number of trees may be less than the number specified in this parameter.

Possible types

int

Default value

1000

fold_count

Alias: nfold

Description

The number of folds to split the dataset into.

Possible types

int

Default value

3

inverted

Description

Train on the test fold and evaluate the model on the training folds.

Possible types

bool

Default value

False

partition_random_seed

Alias: seed

Description

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

shuffle

Description

Shuffle the dataset objects before splitting into folds.

Possible types

bool

Default value

True

logging_level

Description

The logging level to output to stdout.

Possible values:

  • Silent — Do not output any logging information to stdout.

  • Verbose — Output the following data to stdout:

    • optimized metric
    • elapsed time of training
    • remaining time of training
  • Info — Output additional information and the number of trees.

  • Debug — Output debugging information.

Possible types

string

Default value

None (corresponds to the Verbose logging level)

stratified

Description

Perform stratified sampling.

It is turned on (True) by default if one of the following loss functions is selected: Logloss, MultiClass, MultiClassOneVsAll.

It is turned off (False) for all other loss functions by default.

Possible types

bool

Default value

None

as_pandas

Description

Sets the type of return value to pandas.DataFrame.

The type of return value is dict if this parameter is set to False or the pandasPython package is not installed.

Possible types

bool

Default value

True

metric_period

Description

The frequency of iterations to calculate the values of objectives and metrics. The value should be a positive integer.

The usage of this parameter speeds up the training.

Possible types

int

Default value

1

verbose

Alias: verbose_eval

Description

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

  • bool — Defines the logging level:

    • True  corresponds to the Verbose logging level
    • False corresponds to the Silent logging level
  • int — Use the Verbose logging level and set the logging period to the value of this parameter.

Possible types

  • bool
  • int

Default value

False

plot

Description

Plot the following information during training:

  • the metric values;
  • the custom loss values;
  • the loss function change during feature selection;
  • the time has passed since training started;
  • the remaining time until the end of training.
    This option can be used if training is performed in Jupyter notebook.

Possible types

bool

Default value

False

early_stopping_rounds

Description

Sets the overfitting detector type to Iter and stops the training after the specified number of iterations since the iteration with the optimal metric value.

Possible types

int

Default value

False

folds

Description

Custom splitting indices.

The format of the input data depends on the type of the parameter:

  • generator or iterator — Train and test indices for each fold.
  • object — One of the scikit-learn Splitter Classes with the split method.

This parameter has the highest priority among other data split parameters.

Possible types

  • generator
  • iterator
  • scikit-learn splitter object

Default value

None

type

Description

The method to split the dataset into folds.

Possible values:

  • Classical — The dataset is split into fold_count folds, fold_count trainings are performed. Each test set consists of a single fold, and the corresponding train set consists of the remaining k–1 folds.

  • Inverted — The dataset is split into fold_count folds, fold_count trainings are performed. Each test set consists of the first k–1 folds, and the corresponding train set consists of the remaining fold.

  • TimeSeries — The dataset is split into (fold_count + 1) consecutive parts without shuffling the data, fold_count trainings are performed. The k-th train set consists of the first k folds, and the corresponding test set consists of the (k+1)-th fold.

Possible types

string

Default value

Classical

return_models

Description

If return_models is True, returns a list of models fitted for each CV fold. By default, False.

Possible types

bool

Default value

False

Type of return value

Depends on return_models, as_pandas, and the availability of the pandasPython package:

  • If return_models is False, cv returns cv_results which is a dict or a pandas frame (see a table below).
  • If return_models is True, cv returns a tuple (cv_results, fitted_models) containing, in addition to regular cv_results, a list of models fitted for each fold.
as_pandas value pandasPython package availability Type of return value
True Installed pandas.DataFrame
True Not installed dict
False Unimportant dict

The first key (if the output type is dict) or column name (if the output type is pandas.DataFrame) contains the iteration of the calculated metrics values on the corresponding line. Each following key or column name is formed from the evaluation dataset type (train or test), metric name, and computed characteristic (std, mean, etc.). Each value is a list of corresponding computed values.

For example, if only the RMSE metric is specified in the parameters, then the return value is:

    iterations  test-Logloss-mean  test-Logloss-std  train-Logloss-mean  train-Logloss-std
0            0           0.693219          0.000101            0.684767           0.011851
1            1           0.682687          0.014995            0.674235           0.003043
2            2           0.672758          0.029630            0.655983           0.005906
3            3           0.668589          0.023734            0.648127           0.005204

Each key or column value contains the same number of calculated values as the number of training iterations (or less, if the overfitting detection is turned on and the threshold is reached earlier).

Examples

Perform cross-validation on the given dataset:

from catboost import Pool, cv

cv_data = [["France", 1924, 44],
           ["USA", 1932, 37],
           ["Switzerland", 1928, 25],
           ["Norway", 1952, 30],
           ["Japan", 1972, 35],
           ["Mexico", 1968, 112]]

labels = [1, 1, 0, 0, 0, 1]

cat_features = [0]

cv_dataset = Pool(data=cv_data,
                  label=labels,
                  cat_features=cat_features)

params = {"iterations": 100,
          "depth": 2,
          "loss_function": "Logloss",
          "verbose": False}

scores = cv(cv_dataset,
            params,
            fold_count=2, 
            plot="True")

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

Perform cross-validation and save ROC curve points to the roc-curve output file:

from catboost import Pool, cv

cv_data = [["France", 1924, 44],
           ["USA", 1932, 37],
           ["Switzerland", 1928, 25],
           ["Norway", 1952, 30],
           ["Japan", 1972, 35],
           ["Mexico", 1968, 112]]

labels = [1, 1, 0, 0, 0, 1]

cv_dataset = Pool(data=cv_data,
                  label=labels,
                  cat_features=[0])

params = {"iterations": 100,
          "depth": 2,
          "loss_function": "Logloss",
          "verbose": False,
          "roc_file": "roc-file"}

scores = cv(cv_dataset,
            params,
            fold_count=2)