DutchDraw 0.0.1

Determine (optimal) baselines for binary classification

DutchDraw

DutchDraw is a Python package for binary classification.

Paper

This package is an implementation of the ideas from INSERTONZEPAPER, where VERHAALWATWEINDEPAPERDOEN.

Citation

If you have used the DutchDraw package, please also cite: INSERTONZEBIBTEX

Installation

Use the package manager pip to install the package

pip install DutchDraw

---

Windows users

python -m pip install --upgrade  --index-url https://test.pypi.org/simple/ DutchDraw

or

py -m pip install --upgrade  --index-url https://test.pypi.org/simple/ DutchDraw

Method

To properly assess the performance of a binary classification model, the score of a chosen measure should be compared with the score of a 'simple' baseline. E.g. an accuracy of 0.9 isn't that great if a model (without knowledge) attains an accuracy of 0.88.

Basic baseline

Let M be the total number of samples, where P are positive and N are negative. Let θ_star = round(θ * M) / M. Randomly shuffle the samples and label the first θ_star * M samples as 1 and the rest as 0. This gives a baseline for each θ in [0,1]. Our package can optimize (maximize and minimize) the baseline.

Reasons to use

This package contains multiple functions. Let y_true be the actual labels and y_pred be the labels predicted by a model.

If:

• You want to determine an included measure --> measure_score(y_true, y_pred, measure)
• You want to get statistics of a baseline given theta --> baseline_functions_given_theta(theta, y_true, measure)
• You want to get statistics of the optimal baseline --> optimized_baseline_statistics(y_true, measure)
• You want the baseline without specifying theta --> baseline_functions(y_true, measure)

List of all included measures

Measure	Definition
TP	TP
TN	TN
FP	FP
FN	FN
TPR	TP / P
TNR	TN / N
FPR	FP / N
FNR	FN / P
PPV	TP / (TP + FP)
NPV	TN / (TN + FN)
FDR	FP / (TP + FP)
FOR	FN / (TN + FN)
ACC, ACCURACY	(TP + TN) / M
BACC, BALANCED ACCURACY	(TPR + TNR) / 2
FBETA, FSCORE, F, F BETA, F BETA SCORE, FBETA SCORE	((1 + β2) * TP) / ((1 + β2) * TP + β2 * FN + FP)
MCC, MATTHEW, MATTHEWS CORRELATION COEFFICIENT	(TP * TN - FP * FN) / (sqrt((TP + FP) * (TN + FN) * P * N))
BM, BOOKMAKER INFORMEDNESS, INFORMEDNESS	TPR + TNR - 1
MK	PPV + NPV - 1
COHEN, COHENS KAPPA, KAPPA	(Po - Pe) / (1 - Pe) with Po = (TP + TN) / M and Pe = ((TP + FP) / M) * (P / M) + ((TN + FN) / M) * (N / M)
G1, GMEAN1, G MEAN 1, FOWLKES-MALLOWS, FOWLKES MALLOWS, FOWLKES, MALLOWS	sqrt(TPR * PPV)
G2, GMEAN2, G MEAN 2	sqrt(TPR * TNR)
TS, THREAT SCORE, CRITICAL SUCCES INDEX, CSI	TP / (TP + FN + FP)
PT, PREVALENCE THRESHOLD	(sqrt(TPR * FPR) - FPR) / (TPR - FPR)

Usage

As example, we first generate the true and predicted labels.

import random
random.seed(123) # To ensure similar outputs

y_pred = random.choices((0,1), k = 10000, weights = (0.9, 0.1))
y_true = random.choices((0,1), k = 10000, weights = (0.9, 0.1))

---

Measure performance

In general, to determine the score of a measure, use measure_score(y_true, y_pred, measure, beta = 1).

Input

• y_true (list or numpy.ndarray): 1-dimensional boolean list/numpy.ndarray containing the true labels.

• y_pred (list or numpy.ndarray): 1-dimensional boolean list/numpy containing the predicted labels.

• measure (string): Measure name, see all_names_except(['']) for possible measure names.

• beta (float): Default is 1. Parameter for the F-beta score.

Output

• float: The score of the given measure evaluated with the predicted and true labels.

Example

To examine the performance of the predicted labels, we measure the markedness (MK) and F₂ score (FBETA).

import DutchDraw as bbl

# Measuring markedness (MK):
print('Markedness: {:06.4f}'.format(bbl.measure_score(y_true, y_pred, measure = 'MK')))

# Measuring FBETA for beta = 2:
print('F2 Score: {:06.4f}'.format(bbl.measure_score(y_true, y_pred, measure = 'FBETA', beta = 2)))

This returns as output

Markedness: 0.0061
F2 Score: 0.1053

Note that FBETA is the only measure that requires an additional parameter value.

---

Get basic baseline given theta

To obtain the basic baseline given theta use baseline_functions_given_theta(theta, y_true, measure, beta = 1).

Input

• theta (float): Parameter for the shuffle baseline.

• y_true (list or numpy.ndarray): 1-dimensional boolean list/numpy.ndarray containing the true labels.

• measure (string): Measure name, see all_names_except(['']) for possible measure names.

• beta (float): Default is 1. Parameter for the F-beta score.

Output

The function baseline_functions_given_theta gives the following output:

• dict: Containing Mean and Variance
  • Mean (float): Expected baseline given theta.
  • Variance (float): Variance baseline given theta.

Example

To evaluate the performance of a model, we want to obtain a baseline for the F₂ score (FBETA).

results_baseline = bbl.baseline_functions_given_theta(theta = 0.5, y_true = y_true, measure = 'FBETA', beta = 2)

This gives us the mean and variance of the baseline.

print('Mean: {:06.4f}'.format(results_baseline['Mean']))
print('Variance: {:06.4f}'.format(results_baseline['Variance']))

with output

Mean: 0.2829
Variance: 0.0001

---

Get basic baseline

To obtain the basic baseline without specifying theta use baseline_functions(y_true, measure, beta = 1).

Input

• y_true (list or numpy.ndarray): 1-dimensional boolean list/numpy.ndarray containing the true labels.

• measure (string): Measure name, see all_names_except(['']) for possible measure names.

• beta (float): Default is 1. Parameter for the F-beta score.

Output

The function baseline_functions gives the following output:

• dict: Containing Distribution, Domain, (Fast) Expectation Function and Variance Function.

  • Distribution (function): Pmf of the measure, given by: pmf_Y(y, theta), where y is a measure score and theta is the parameter of the shuffle baseline.

  • Domain (function): Function that returns attainable measure scores with argument theta.

  • (Fast) Expectation Function (function): Expectation function of the baseline with theta as argument. If Fast Expectation Function is returned, there exists a theoretical expectation that can be used for fast computation.

  • Variance Function (function): Variance function for all values of theta.

Example

Next, we determine the baseline without specifying theta. This returns a number of functions that can be used for different values of theta.

baseline = bbl.baseline_functions(y_true = y_true, measure = 'G2')
print(baseline.keys())

with output

dict_keys(['Distribution', 'Domain', 'Fast Expectation Function', 'Variance Function', 'Expectation Function'])

To inspect the expected value of G2 for different theta values, we do:

import matplotlib.pyplot as plt
theta_values = np.arange(0, 1 + 0.01, 0.01)
expected_value_plot = [baseline['Expectation Function'](theta) for theta in theta_values]
plt.plot(theta_values, expected_value_plot)
plt.xlabel('Theta')
plt.ylabel('Expected value')
plt.show()

with output:

The variance can be determined similarly

theta_values = np.arange(0, 1 + 0.01, 0.01)
variance_plot = [baseline['Variance Function'](theta) for theta in theta_values]
plt.plot(theta_values, variance_plot)
plt.xlabel('Theta')
plt.ylabel('Variance')
plt.show()

with output:

Distribution is a function with two arguments: y and theta. Let's investigate the distribution for theta = 0.5 using Domain.

theta = 0.5
pmf_values = [baseline['Distribution'](y, theta) for y in baseline['Domain'](theta)]
plt.plot(baseline['Domain'](theta), pmf_values)
plt.xlabel('Measure score')
plt.ylabel('Probability mass')
plt.show()

with output:

---

Get optimal baseline

To obtain the optimal baseline use optimized_baseline_statistics(y_true, measure = possible_names, beta = 1).

Input

• y_true (list or numpy.ndarray): 1-dimensional boolean list/numpy.ndarray containing the true labels.

• measure (string): Measure name, see all_names_except(['']) for possible measure names.

• beta (float): Default is 1. Parameter for the F-beta score.

Output

The function optimized_baseline_statistics gives the following output:

• dict: Containing Max Expected Value, Argmax Expected Value, Min Expected Value and Argmin Expected Value.
  • Max Expected Value (float): Maximum of the expected values for all theta.
  • Argmax Expected Value (list): List of all theta_star values that maximize the expected value.
  • Min Expected Value (float): Minimum of the expected values for all theta.
  • Argmin Expected Value (list): List of all theta_star values that minimize the expected value.

Note that theta_star = round(theta * M) / M.

Example

To evaluate the performance of a model, we want to obtain the optimal baseline for the F₂ score (FBETA).

optimal_baseline = bbl.optimized_baseline_statistics(y_true, measure = 'FBETA', beta = 1)

print('Max Expected Value: {:06.4f}'.format(optimal_baseline['Max Expected Value']))
print('Argmax Expected Value: {:06.4f}'.format(*optimal_baseline['Argmax Expected Value']))
print('Min Expected Value: {:06.4f}'.format(optimal_baseline['Min Expected Value']))
print('Argmin Expected Value: {:06.4f}'.format(*optimal_baseline['Argmin Expected Value']))

with output

Max Expected Value: 0.1874
Argmax Expected Value: 1.0000
Min Expected Value: 0.0000
Argmin Expected Value: 0.0000

---

All example code

import DutchDraw as bbl
import random
import numpy as np

random.seed(123) # To ensure similar outputs

# Generate true and predicted labels
y_pred = random.choices((0,1), k = 10000, weights = (0.9, 0.1))
y_true = random.choices((0,1), k = 10000, weights = (0.9, 0.1))

######################################################
# Example function: measure_score
print('\033[94mExample function: `measure_score`\033[0m')
# Measuring markedness (MK):
print('Markedness: {:06.4f}'.format(bbl.measure_score(y_true, y_pred, measure = 'MK')))

# Measuring FBETA for beta = 2:
print('F2 Score: {:06.4f}'.format(bbl.measure_score(y_true, y_pred, measure= 'FBETA', beta = 2)))

print('')
######################################################
# Example function: baseline_functions_given_theta
print('\033[94mExample function: `baseline_functions_given_theta`\033[0m')
results_baseline = bbl.baseline_functions_given_theta(theta = 0.5, y_true = y_true, measure = 'FBETA', beta = 2)

print('Mean: {:06.4f}'.format(results_baseline['Mean']))
print('Variance: {:06.4f}'.format(results_baseline['Variance']))

print('')
######################################################
# Example function: baseline_functions
print('\033[94mExample function: `baseline_functions`\033[0m')
baseline = bbl.baseline_functions(y_true = y_true, measure = 'G2')
print(baseline.keys())


# Expected Value
import matplotlib.pyplot as plt
theta_values = np.arange(0, 1 + 0.01, 0.01)
expected_value_plot = [baseline['Expectation Function'](theta) for theta in theta_values]
plt.plot(theta_values, expected_value_plot)
plt.xlabel('Theta')
plt.ylabel('Expected value')
#plt.savefig('expected_value_function_example.png', dpi= 600)
plt.show()

# Variance
theta_values = np.arange(0, 1 + 0.01, 0.01)
variance_plot = [baseline['Variance Function'](theta) for theta in theta_values]
plt.plot(theta_values, variance_plot)
plt.xlabel('Theta')
plt.ylabel('Variance')
#plt.savefig('variance_function_example.png', dpi= 600)
plt.show()

# Distribution and Domain
theta = 0.5
pmf_values = [baseline['Distribution'](y, theta) for y in baseline['Domain'](theta)]
plt.plot(baseline['Domain'](theta), pmf_values)
plt.xlabel('Measure score')
plt.ylabel('Probability mass')
#plt.savefig('pmf_example.png', dpi= 600)
plt.show()

print('')
######################################################
# Example function: optimized_baseline_statistics
print('\033[94mExample function: `optimized_baseline_statistics`\033[0m')
optimal_baseline = bbl.optimized_baseline_statistics(y_true, measure = 'FBETA', beta = 1)

print('Max Expected Value: {:06.4f}'.format(optimal_baseline['Max Expected Value']))
print('Argmax Expected Value: {:06.4f}'.format(*optimal_baseline['Argmax Expected Value']))
print('Min Expected Value: {:06.4f}'.format(optimal_baseline['Min Expected Value']))
print('Argmin Expected Value: {:06.4f}'.format(*optimal_baseline['Argmin Expected Value']))

License

MIT