stats_inferential.jl

This unit implements inferential statistics (hypothesis tests) for cross-validation analysis.

Content

function	description
`testCV`	Test the average error loss observed in a cross-validation against an expected (chance) level
`testCV`	Test the average error loss observed in two cross-validation obtained by two different models and/or processing pipelines on the same data

See also stats_descriptive.jl

PosDefManifoldML.testCV — Function

(1) function testCV(cv:CVresult, 
			expected::Union{Float64, Symbol} = :randomchance)

(2) function testCV(𝐂::Vector{Vector{Int}}; 
			expected::Union{Float64, Symbol} = :randomchance)

(1) Given a CVres stucture (which is of type CVresult) obtained calling the crval method, performs Bayles's test (Bayles et al., 2020🎓) on the difference in distribution of the average binary error losses obtained in the cross-validation as compared to a given expected average error loss. This can be used to test whether the performance obtained in the cross-validation is superior to a specified chance level.

(2) As method (1) but giving as first argument a vector of f confusion matrices, The format of the confusion matrices (Vector{Vector{Int}}) is the same used to store the confusion matrices in a CVres stucture.

The test is always directional. Denoting $μ$ the average loss and $E$ the expected average, the test if performed according to null hypethesis $H_0: μ1 = E$ against alternative $μ < E$; if the test is significant, the error loss is statistically inferior to the specified expected level, which means that the observed accuracy is statistically superior to the expected accuracy.

By default, expected is set to $1-\frac{1}{z}$, where $z$ is the number of classes. You can pass another value, which must be a real number.

Return the 3-tuple (z, p, ase) holding, the $z$ test-statistic (a standard deviate), the $p$-value and the asymptotic standard error.

Class Inbalance

The error lost reflects the accuracy, not the balance accuracy. If the class are inbalanced, the test will be driven by the majority classes. See also the documentation on crval.

Examples:

using PosDefManifoldML

# Generate a set of random data
P, _dummyP, y, _dummyy = gen2ClassData(10, 60, 80, 30, 40, 0.1)

# Perform a cross-validation
cv = crval(MDM(Fisher), P, y; scoring=:a)

# The `z` and `p` values are printed out as they are computed automatically
# and stored in the created structure. To compute it yourself:

z, p, ase = testCV(cv)

source

(1) function testCV(cv1::CVresult, cv2::CVresult; 
               	direction = PermutationTests.Both())

(2) function testCV(𝐞₁::Vector{BitVector}, 𝐞₂::Vector{BitVector}; 
               	direction = PermutationTests.Both())

(1) Given two CVres stuctures cv1 and cv2, which can be obtaoined calling the crval method, performs Bayles's test (Bayles et al., 2020🎓) on the difference in distribution of the average binary error losses obtained in cross-validations cv1 and cv2. This can be used to compare the performance obtained by different machine learning models and/or different pre-processing, processing or pre-conditoning pipelines on the same data.

Be careful

When you compare two models and/or pipelines with cross-validation, do not use keyword argument shuffle=true in crval, as in this case the actual data composing the folds would not be identical. Also, is you use the seed keyword argument, make sure the same seed is employed for the two cross-validations.

(2) The function also accepts as argument the error losses directly, denoted 𝐞₁ and 𝐞₂ in method (2) here above. In this case 𝐞₁ and 𝐞₂ are vectors holding the vectors of error losses obtained at each fold. The format is the same used to store the error losses in a CVres stucture.

Denoting $μ_1$ and $μ_2$ the average loss for cv1 and cv2 (or 𝐞₁ and 𝐞₂), respectively, the direction keyword argument determines the test directionality:

$H_1: μ_1 > μ_2$ (use direction=PermutationTests.Right())
$H_1: μ_1 < μ_2$ (use direction=PermutationTests.Left())
$H_1: μ_1 <> μ_2$ (use direction=PermutationTests.Both(), the default)

In al cases $H_0: μ_1 = μ_2$.

Test direction in terms of accuracy

Note that if the test is significant with direction=PermutationTests.Right(), the error loss of cv1(or 𝐞₁) is statistically superior to the error loss of cv2(or 𝐞₂), which means that the accuracy in cv1(or 𝐞₁) is statistically inferior to the accuracy in cv2(or 𝐞₂).

Return the 3-tuple (z, p, ase) holding, the $z$ test-statistic (a standard deviate), the $p$-value and the asymptotic standard error.

Class Inbalance

The error lost reflects the accuracy, not the balance accuracy. If the class are inbalanced, the test will be driven by the majority classes. See also the documentation on crval.

Examples:

using PosDefManifoldML

## Test the performance of the same model and pipeline on different data:

# Generate two sets of random data with the same distribution
P1, _dummyP, y1, _dummyy = gen2ClassData(10, 60, 80, 30, 40, 0.01)
P2, _dummyP, y2, _dummyy = gen2ClassData(10, 60, 80, 30, 40, 0.01)

# Perform a cross-validation on the two sets of data
cv1 = crval(MDM(Fisher), P1, y1; scoring=:a)
cv2 = crval(MDM(Fisher), P2, y2; scoring=:a)

z, p, ase = testCV(cv1, cv2) # two-sided test

# This is equivalent to
z, p, ase = testCV(cv1.losses, cv2.losses)

## Test the performance of, for example, different metrics on the same data:

# Generate two sets of random data with the same distribution
P, _dummyP, y, _dummyy = gen2ClassData(10, 40, 40, 30, 40, 0.15)

cv1 = crval(MDM(logEuclidean), P, y; scoring=:a)
cv2 = crval(MDM(Wasserstein), P, y; scoring=:a)

z, p, ase = testCV(cv1, cv2) # two-sided test

source