def test_generalized_average():
a, b = 1, 2
methods = ["min", "geometric", "arithmetic", "max"]
means = [_generalized_average(a, b, method) for method in methods]
assert means[0] <= means[1] <= means[2] <= means[3]
c, d = 12, 12
means = [_generalized_average(c, d, method) for method in methods]
assert means[0] == means[1] == means[2] == means[3]
def test_generalized_average():
a, b = 1, 2
methods = ["min", "geometric", "arithmetic", "max"]
means = [_generalized_average(a, b, method) for method in methods]
assert means[0] <= means[1] <= means[2] <= means[3]
c, d = 12, 12
means = [_generalized_average(c, d, method) for method in methods]
assert means[0] == means[1] == means[2] == means[3]
def _ami(ab_cts, average_method='arithmetic'):
"""Adjusted mutual information between two discrete categorical random variables
based on counts observed and provided in ab_cts.
Code adapted directly from scikit learn AMI to
accomodate having counts/contingency table instead of rows/instances:
https://scikit-learn.org/stable/modules/generated/sklearn.metrics.adjusted_mutual_info_score.html
Parameters
----------
ab_cts : np.ndarray [len(a_classes) x len(b_classes)
Counts for each combination of classes in random variables a and b
organized in a rectangular array.
average_method : str
See sklearn documentation for details
Returns
-------
ami : float
Adjusted mutual information score for variables a and b"""
a_freq = np.sum(ab_cts, axis=1)
a_freq = a_freq / np.sum(a_freq)
b_freq = np.sum(ab_cts, axis=0)
b_freq = b_freq / np.sum(b_freq)
n_samples = np.sum(ab_cts)
""" Calculate the MI for the two clusterings
contingency is a joint count distribution [a_classes x b_classes]"""
mi = mutual_info_score(None, None, contingency=ab_cts)
"""Calculate the expected value for the mutual information"""
emi = expected_mutual_information(ab_cts, n_samples)
"""Calculate entropy"""
h_true, h_pred = _entropy(a_freq), _entropy(b_freq)
normalizer = _generalized_average(h_true, h_pred, average_method)
denominator = normalizer - emi
if denominator < 0:
denominator = min(denominator, -np.finfo('float64').eps)
else:
denominator = max(denominator, np.finfo('float64').eps)
ami = (mi - emi) / denominator
return ami