Introduction:

In today’s tutorial, we will see various techniques of cross-validation in Sklearn such as K-

fold cross-validation, Stratified K-fold cross-validation, Leave one out cross-
validation(LOOCV), and Repeated random train test splits. We will look into how these
cross-validation techniques help better evaluation of machine learning models along with
examples.

  What is Cross-Validation in Machine Learning

When we create a machine learning model with supervised learning techniques it is very
important to evaluate the performance and robustness of the model before moving it to
production.

Key aspects of evaluating the quality of the model are –

 How accurate the model is

 How generalized the model is
When we start building a model and train it with the ‘entire’ dataset, we can very well
calculate its accuracy on this training data set. But we cannot test how this model will
behave with new data which is not present in the training set, hence its generalization
cannot be determined.
Hence we need techniques to make use of the same data set for both training and testing
of the models.
In machine learning, Cross-Validation is the technique to evaluate how well the model has
generalized and its overall accuracy. For this purpose, it randomly samples data from the
dataset to create training and testing sets. There are multiple cross-validation approaches
as follows –

1. Hold Out Approach

2. Leave One Out Cross-Validation
3. K-Fold Cross-Validation
4. Stratified K-Fold Cross-Validation
5. Repeated Random Train Test Split
1. Hold Out Approach
In the hold-out approach, the data set is split into the train and test set with random
sampling. The train set is used for training the model and the test set is used to test its
accuracy with unseen data. If the training and accuracy are almost the same then the
model is said to have generalized well. It is common to use 80% of data for training and
the remaining 20% for testing.

Advantages

 It is simple and easy to implement

 The execution time is less.

Disadvantages

 If the dataset itself is small, setting aside portions for testing would reduce the
robustness of the model. This is because the training sample may not be
representative of the entire dataset.
 The evaluation metrics may vary due to the randomness of the split between the
train and test set.
 Although 80-20 split for train test is widely followed, there is no thumb rule for the
split and hence the results can vary based on how the train test split is done.

2. Leave One Out Cross Validation (LOOCV)

In this technique, if there are n observations in the dataset, only one observation is
reserved for testing, and the remaining data points are used for training. This is repeated
n times till all data points have been used for testing purposes in each iteration. Finally,
the average accuracy is calculated by combining the accuracies of each iteration.

Advantage

 Since every data participates both for training and testing, the overall accuracy is
more reliable.
 It is very useful when the dataset is small.

Disadvantage

 LOOCV is not practical to use when the number of data observations n is huge.
E.g. imagine a dataset with 500,000 records, then 500,000 model needs to be
created which is not really feasible.
 There is a huge computational and time cost associated with the LOOCV approach.

3. K-Fold Cross-Validation
In the K-Fold Cross-Validation approach, the dataset is split into K folds. Now in 1st
iteration, the first fold is reserved for testing and the model is trained on the data of the
remaining k-1 folds.
In the next iteration, the second fold is reserved for testing and the remaining folds are
used for training. This is continued till the K-th iteration. The accuracy obtained in each
iteration is used to derive the overall average accuracy for the model.

Advantages

 K-Fold cross-validation is useful when the dataset is small and splitting it is not
possible to split it in train-test set (hold out approach) without losing useful data
for training.
 It helps to create a robust model with low variance and low bias as it is trained on
all data

Disadvantages

 The major disadvantage of K-Fold Cross Validation is that the training needs to be
done K times and hence it consumes more time and resources,
 Not recommended to be used with sequential time series data.
 When the dataset is imbalanced, K-fold cross-validation may not give good results.
This is because some folds may have just a few or no records for the
minority class.

4. Stratified K-Fold Cross-Validation

Stratified K-fold cross-validation is useful when the data is imbalanced. While sampling
data into K-folds it makes sure that the distribution of all classes in each fold is
maintained. For example, if in the dataset 98% of data belongs to class B and 2% to class
A, the stratified sampling will make sure each fold contains the two classes in the same
ratio of 98% to 2%.

Advantage
Stratified K-fold cross-validation is recommended when the dataset is imbalanced.

5. Repeated Random Test-Train Split

Repeated random test-train split is a hybrid of traditional train-test splitting and the k-fold
cross-validation method. In this technique, we create random splits of the data into the
training-test set and then repeat this process multiple times, just like the cross-validation
method.

Examples of Cross-Validation in Sklearn Library

About Dataset
We will be using Parkinson’s disease dataset for all examples of cross-validation in the
Sklearn library. The goal is to predict whether or not a particular patient has Parkinson’s
disease. We will be using the decision tree algorithm in all the examples.

The dataset has 21 attributes and 195 rows. The various fields of the Parkinson’s Disease
dataset are as follows –
  MDVP:Fo(Hz) – Average vocal fundamental frequency
 MDVP:Fhi(Hz) – Maximum vocal fundamental frequency
 MDVP:Flo(Hz) – Minimum vocal fundamental frequency
 MDVP:Jitter(%),MDVP:Jitter(Abs),MDVP:RAP,MDVP:PPQ,Jitter:DDP – Several
 measures of variation in fundamental frequency
 MDVP:Shimmer,MDVP:Shimmer(dB),Shimmer:APQ3,Shimmer:APQ5,MDVP:APQ,Sh
immer:DDA – Several measures of variation in amplitude
 NHR,HNR – Two measures of ratio of noise to tonal components in the voice
 status – Health status of the subject (one) – Parkinson’s, (zero) – healthy
 RPDE,D2 – Two nonlinear dynamical complexity measures
 DFA – Signal fractal scaling exponent
 spread1,spread2PPE – Three nonlinear measures of fundamental frequency
variation

Importing Necessary Libraries

We first load the libraries required to build our model.

import pandas as pdimport numpy as npfrom sklearn.tree import

DecisionTreeClassifierfrom sklearn.model_selection import
train_test_splitfrom sklearn.model_selection import KFoldfrom
sklearn.model_selection import StratifiedKFold

Reading CSV Data into Pandas

Next, we load the dataset in the CSV file into the pandas dataframes and check the top 5
rows.

df=pd.read_csv(“Parkinsson disease.csv")

df.head()
Out[52]:
MDVP:Fo(Hz MDVP:Fhi(Hz MDVP:Flo(Hz MDVP:Jitter(% MDVP:Jitter(
name
) ) ) ) )
phon_R01_S01_
0 119.992 157.302 74.997 0.00784 0.00007
1

phon_R01_S01_
1 122.400 148.650 113.819 0.00968 0.00008
2

phon_R01_S01_
2 116.682 131.111 111.555 0.01050 0.00009
3

phon_R01_S01_
3 116.676 137.871 111.366 0.00997 0.00009
4

phon_R01_S01_
4 116.014 141.781 110.655 0.01284 0.00011
5
5 rows × 24 columns
Data Preprocessing
The “name” column is not going to add any value in training the model and can be
discarded, so we are dropping it below.

df.drop(df.columns[0], axis = 1, inplace = True)

Next, we will separate the feature and target matrix as shown below.

#Independent And dependent features

X=df.drop('status', axis=1)

y=df['status']

Hold out Approach in Sklearn

The hold-out approach can be applied by using train_test_split module of
sklearn.model_selection

In the below example we have split the dataset to create the test
data with a size of 30% and train data with a size of 70%. The
random_state number ensures the split is deterministic in every run.

from sklearn.model_selection import train_test_split

X_train, X_test, y_train, y_test = train_test_split(X, y,

test_size=0.30, random_state=4)

model = DecisionTreeClassifier()

model.fit(X_train, y_train)

result = model.score(X_test, y_test)print(result)

Out[38]:

0.7796610169491526
 

K-Fold Cross-Validation
K-Fold Cross-Validation in Sklearn can be applied by using cross_val_score module
of sklearn.model_selection.

In the below example, 10 folds are used that produced 10 accuracy scores using which we
calculated the mean score.

In [40]:

from sklearn.model_selection import cross_val_score

model=DecisionTreeClassifier()

kfold_validation=KFold(10)
results=cross_val_score(model,X,y,cv=kfold_validation)

print(results)print(np.mean(results))
Out[40]:

[0.7 0.8 0.8 0.8 0.8

0.78947368

0.84210526 1. 0.68421053 0.36842105]

0.758421052631579

Stratified K-fold Cross-Validation

In Sklearn stratified K-fold cross-validation can be applied by using StratifiedKFold module
of sklearn.model_selection

In the below example, the dataset is divided into 5 splits or folds. It returns 5 accuracy
scores using which we calculate the final mean score.

In [41]:

from sklearn.model_selection import StratifiedKFold

skfold=StratifiedKFold(n_splits=5)

model=DecisionTreeClassifier()scores=cross_val_score(model,X,y,cv
=skfold)

print(scores)print(np.mean(scores))
Out[41]:

array([0.61538462, 0.79487179, 0.71794872, 0.74358974,

0.71794872])

0.717948717948718
 

Leave One Out Cross Validation(LOOCV)

In Sklearn Leave One Out Cross Validation (LOOCV) can be applied by
using LeaveOneOut module of sklearn.model_selection

From sklearn.model_selection import LeaveOneOut

model=DecisionTreeClassifier()

leave_validation=LeaveOneOut()

results=cross_val_score(model,X,y,cv=leave_validation)

results
 

Out[22]:

array([1., 1., 1., 1., 1., 0., 1., 1., 1., 1., 1., 1., 1., 0.,
1., 1., 1.,

1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 0., 1.,
1., 1., 1.,

1., 1., 1., 1., 1., 0., 0., 1., 1., 1., 0., 1., 1., 1.,
1., 1., 1.,

1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
1., 1., 1.,

0., 1., 1., 1., 1., 0., 1., 1., 1., 1., 1., 1., 1., 0.,
1., 1., 1.,

1., 0., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
0., 1., 1.,

1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
1., 1., 1.,

1., 1., 1., 1., 1., 1., 1., 0., 1., 1., 1., 1., 1., 1.,
1., 1., 0.,

1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
1., 1., 1.,

1., 1., 1., 1., 0., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
1., 1., 1.,

1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
1., 1., 0.,

1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 0., 1., 1., 1.,
1., 1., 1.,

1., 0., 1., 1., 1., 1., 1., 1., 1., 1., 1., 0., 1., 1.,
1., 1., 1.,

1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 0., 1., 1., 1.,
1., 1., 1.,

1., 1., 1., 1., 1., 1., 1., 1., 1., 0., 1., 1., 1., 1.,
1., 1., 1.,

0., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
1., 1., 1.,

1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
1., 1., 1.,
 1., 1., 0., 1., 1., 1., 1., 1., 0., 1., 1., 1., 1., 1.,
1., 1., 1.,

1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
1., 1., 1.,

1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
1., 1., 1.,

0., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
1., 1., 1.,

1., 1., 1., 0., 1., 1., 0., 1., 1., 1., 1., 1., 1., 1.,
1., 1., 1.,

1., 0., 0., 1., 1., 0., 1., 1., 1., 1., 1., 0., 1., 1.,
1., 1., 1.,

1., 1., 1., 1., 1., 0., 1., 1., 1., 1., 1., 1., 1., 1.,
1., 0., 1.,

1., 1., 0., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
1., 1., 1.,

1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
1., 0., 1.,

1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
1., 1., 1.,

1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 0., 1., 0., 0.,
1., 1., 1.,

1., 0., 1., 1., 1., 1., 1., 1., 0., 1., 1., 1., 1., 1.,
1., 0., 1.,

1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
1., 0., 1.,

1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
1., 1., 1.,

1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
1., 1., 1.,

1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
1., 0., 0.,

1., 1., 1., 1., 1., 1., 1., 1.])

In [44]:

print(np.mean(results))
Out[44]:

0.8358974358974359
Repeated Random Test-Train Splits
In Sklearn repeated random test-train splits can be applied by using ShuffleSplit module of
sklearn.model_selection

In [45]:

from sklearn.model_selection import ShuffleSplit

model=DecisionTreeClassifier()

ssplit=ShuffleSplit(n_splits=10,test_size=0.30)

results=cross_val_score(model,X,y,cv=ssplit)

print(results)print(np.mean(results))
Out[45]:

array([0.79661017, 0.71186441, 0.79661017, 0.88135593,

0.72881356,

0.84745763, 0.83050847, 0.77966102, 0.83050847,

0.81355932])

0.8016949152542372