5x2cv combined F test

5x2cv combined F test procedure to compare the performance of two models

from mlxtend.evaluate import combined_ftest_5x2cv

Overview

The 5x2cv combined F test is a procedure for comparing the performance of two models (classifiers or regressors) that was proposed by Alpaydin [1] as a more robust alternative to Dietterich's 5x2cv paired t-test procedure [2]. paired_ttest_5x2cv.md. Dietterich's 5x2cv method was in turn designed to address shortcomings in other methods such as the resampled paired t test (see paired_ttest_resampled) and the k-fold cross-validated paired t test (see paired_ttest_kfold_cv).

To explain how this method works, let's consider to estimator (e.g., classifiers) A and B. Further, we have a labeled dataset D. In the common hold-out method, we typically split the dataset into 2 parts: a training and a test set. In the 5x2cv paired t test, we repeat the splitting (50% training and 50% test data) 5 times.

In each of the 5 iterations, we fit A and B to the training split and evaluate their performance ( and ) on the test split. Then, we rotate the training and test sets (the training set becomes the test set and vice versa) compute the performance again, which results in 2 performance difference measures:

and

Then, we estimate mean and variance of the differences:

and

The F-statistic proposed by Alpaydin (see paper for justifications) is then computed as

which is approximately F distributed with 10 and 5 degress of freedom.

Using the f statistic, the p value can be computed and compared with a previously chosen significance level, e.g., . If the p value is smaller than , we reject the null hypothesis and accept that there is a significant difference in the two models.

References

Example 1 - 5x2cv combined F test

Assume we want to compare two classification algorithms, logistic regression and a decision tree algorithm:

from sklearn.linear_model import LogisticRegression
from sklearn.tree import DecisionTreeClassifier
from mlxtend.data import iris_data
from sklearn.model_selection import train_test_split


X, y = iris_data()
clf1 = LogisticRegression(random_state=1, solver='liblinear', multi_class='ovr')
clf2 = DecisionTreeClassifier(random_state=1)

X_train, X_test, y_train, y_test = \
    train_test_split(X, y, test_size=0.25,
                     random_state=123)

score1 = clf1.fit(X_train, y_train).score(X_test, y_test)
score2 = clf2.fit(X_train, y_train).score(X_test, y_test)

print('Logistic regression accuracy: %.2f%%' % (score1*100))
print('Decision tree accuracy: %.2f%%' % (score2*100))
Logistic regression accuracy: 97.37%
Decision tree accuracy: 94.74%

Note that these accuracy values are not used in the paired f test procedure as new test/train splits are generated during the resampling procedure, the values above are just serving the purpose of intuition.

Now, let's assume a significance threshold of for rejecting the null hypothesis that both algorithms perform equally well on the dataset and conduct the 5x2cv f test:

from mlxtend.evaluate import combined_ftest_5x2cv


f, p = combined_ftest_5x2cv(estimator1=clf1,
                            estimator2=clf2,
                            X=X, y=y,
                            random_seed=1)

print('F statistic: %.3f' % f)
print('p value: %.3f' % p)
F statistic: 1.053
p value: 0.509

Since , we cannot reject the null hypothesis and may conclude that the performance of the two algorithms is not significantly different.

While it is generally not recommended to apply statistical tests multiple times without correction for multiple hypothesis testing, let us take a look at an example where the decision tree algorithm is limited to producing a very simple decision boundary that would result in a relatively bad performance:

clf2 = DecisionTreeClassifier(random_state=1, max_depth=1)

score2 = clf2.fit(X_train, y_train).score(X_test, y_test)
print('Decision tree accuracy: %.2f%%' % (score2*100))


f, p = combined_ftest_5x2cv(estimator1=clf1,
                            estimator2=clf2,
                            X=X, y=y,
                            random_seed=1)

print('F statistic: %.3f' % f)
print('p value: %.3f' % p)
Decision tree accuracy: 63.16%
F statistic: 34.934
p value: 0.001

Assuming that we conducted this test also with a significance level of , we can reject the null-hypothesis that both models perform equally well on this dataset, since the p-value () is smaller than .

API

combined_ftest_5x2cv(estimator1, estimator2, X, y, scoring=None, random_seed=None)

Implements the 5x2cv combined F test proposed by Alpaydin 1999, to compare the performance of two models.

Parameters

Returns

Examples

For usage examples, please see http://rasbt.github.io/mlxtend/user_guide/evaluate/combined_ftest_5x2cv/