def q2(y_true, y_pred, contingency=None, cond_entrop=None, a_i=None, b_j=None):
if contingency is None:
contingency = utility.compute_contingency(y_true, y_pred)
if cond_entrop is None:
cond_entrop = cond_entropy(y_true, y_pred, contingency)
if a_i is None:
a_i = utility.compute_a_i(contingency)
if b_j is None:
b_j = utility.compute_b_j(contingency)
J = np.ma.size(b_j)
N = np.sum(a_i)
combi = comb(a_i + J - 1, J - 1)
log = np.ma.log(combi)
log = log.filled(0)
q0 = cond_entrop + (np.sum(log) / N)
logmin = comb(b_j + J - 1, J - 1)
logmin = np.ma.log(logmin)
logmin = logmin.filled(0)
maxq0 = sklearn.metrics.cluster.entropy(y_true) + np.log(J)
minq0 = np.sum(logmin) / N
return (maxq0 - q0) / (maxq0 - minq0)
def adjusted_mutual_information(y_true,
y_pred,
contingency=None,
a_i=None,
b_j=None,
mi=None,
real_clustering_entropy=None,
predicted_clustering_entropy=None):
if contingency is None:
contingency = utility.compute_contingency(y_true, y_pred)
if a_i is None:
a_i = utility.compute_a_i(contingency)
if b_j is None:
b_j = utility.compute_b_j
if mi is None:
mi = mutual_information(y_true, y_pred, contingency, a_i, b_j)
if real_clustering_entropy is None:
real_clustering_entropy = sklearn.metrics.cluster.entropy(y_true)
if predicted_clustering_entropy is None:
predicted_clustering_entropy = sklearn.metrics.cluster.entropy(y_pred)
J_a_i = np.repeat(a_i, np.ma.size(b_j), axis=1)
I_b_j = np.repeat(b_j, np.ma.size(a_i), axis=0)
N = np.sum(contingency)
emi = expected_mutual_information(contingency, N)
maxmi = np.sqrt(real_clustering_entropy * predicted_clustering_entropy)
return (mi - emi) / (maxmi - emi)
def cond_entropy(y_true, y_pred, contingency=None, a_i=None):
# Compute the average of each predicted-cluster's entropy, which is based on membership diversity of each data on reals clusters
# Must be distinguished from entropy
if contingency is None:
contingency = utility.compute_contingency(y_true, y_pred)
if a_i is None:
a_i = utility.compute_a_i(contingency)
J_a_i = np.repeat(a_i, np.size(contingency, axis=1), axis=1)
log_value = np.true_divide(contingency, J_a_i)
log_value = np.ma.log(log_value) # Required to deal with 0's
log_value = log_value.filled(0)
N = np.ma.sum(a_i)
return -np.sum(contingency * log_value) / N
def mutual_information(y_true, y_pred, contingency=None, a_i=None, b_j=None):
if contingency is None:
contingency = utility.compute_contingency(y_true, y_pred)
if a_i is None:
a_i = utility.compute_a_i(contingency)
if b_j is None:
b_j = utility.compute_b_j
J_a_i = np.repeat(a_i, np.ma.size(b_j), axis=1)
I_b_j = np.repeat(b_j, np.ma.size(a_i), axis=0)
N = np.sum(a_i)
tmp = np.true_divide(N * contingency, J_a_i * I_b_j)
log_value = np.ma.log(tmp)
log_value = log_value.filled(0)
return np.sum(contingency * log_value) / N
def normalized_mutual_information(y_true,
y_pred,
contingency=None,
a_i=None,
b_j=None,
mi=None,
real_clustering_entropy=None,
predicted_clustering_entropy=None):
if contingency is None:
contingency = utility.compute_contingency(y_true, y_pred)
if a_i is None:
a_i = utility.compute_a_i(contingency)
if b_j is None:
b_j = utility.compute_b_j
if mi is None:
mi = mutual_information(y_true, y_pred, contingency, a_i, b_j)
if real_clustering_entropy is None:
real_clustering_entropy = sklearn.metrics.cluster.entropy(y_true)
if predicted_clustering_entropy is None:
predicted_clustering_entropy = sklearn.metrics.cluster.entropy(y_pred)
return mi / np.sqrt(real_clustering_entropy * predicted_clustering_entropy)
def adjusted_rand_index(y_true,
y_pred,
contingency=None,
PBV=None,
a_i=None,
b_j=None):
if contingency is None:
contingency = compute_contingency(y_true, y_pred)
if PBV is None:
PBV = utility.pair_based_values(contingency)
if a_i is None:
a_i = utility.compute_a_i(contingency)
if b_j is None:
b_j = utility.compute_b_j(contingency)
tp, fp, fn, tn = PBV
N = np.sum(a_i)
ri = np.sum(comb(contingency,
2)) # note : not truly ri, missing division by comb(N,2)
eri = (np.sum(comb(a_i, 2)) * np.sum(comb(b_j, 2))) / comb(N, 2) # idem
maxri = 0.5 * (np.sum(comb(a_i, 2)) + np.sum(comb(b_j, 2))) # idem
return (ri - eri) / (maxri - eri)
def variation_of_information(y_true,
y_pred,
contingency=None,
a_i=None,
b_j=None):
if contingency is None:
contingency = utility.compute_contingency(y_true, y_pred)
if a_i is None:
a_i = utility.compute_a_i(contingency)
if b_j is None:
b_j = utility.compute_b_j(contingency)
J_a_i = np.repeat(a_i, np.ma.size(b_j), axis=1)
I_b_j = np.repeat(b_j, np.ma.size(a_i), axis=0)
N = np.sum(a_i)
log1 = np.true_divide(contingency, J_a_i)
log1 = np.ma.log(log1)
log1 = log1.filled(0)
log2 = np.true_divide(contingency, I_b_j)
log2 = np.ma.log(log2)
log2 = log2.filled(0)
return -np.sum(contingency * (log1 + log2)) / N
def compute_prediction_metrics(y_true, y_pred):
# Calculate criteria score for each one present in CRITERION_LIST
# Calculating values that are presents in several formula of criterion
contingency = utility.compute_contingency(y_true, y_pred)
a_i = utility.compute_a_i(contingency)
b_j = utility.compute_b_j(contingency)
true_clustering_entropy = sklearn.metrics.cluster.entropy(y_true)
predicted_clustering_entropy = sklearn.metrics.cluster.entropy(y_pred)
PBV = utility.pair_based_values(contingency)
# Getting all criterion values
mi = converted_metrics.mutual_information(y_true, y_pred, contingency, a_i,
b_j)
ari = converted_metrics.adjusted_rand_index(y_true, y_pred, contingency,
PBV, a_i, b_j)
ami = converted_metrics.adjusted_mutual_information(
y_true, y_pred, contingency, a_i, b_j, mi, true_clustering_entropy,
predicted_clustering_entropy)
compl = converted_metrics.homogeneity(y_true, y_pred, contingency, mi,
predicted_clustering_entropy)
homog = converted_metrics.completness(y_true, y_pred, contingency, mi,
true_clustering_entropy)
vmeasure = converted_metrics.v_measure(y_true, y_pred, 1, mi, contingency,
predicted_clustering_entropy,
true_clustering_entropy, homog,
compl)
entropy = new_metrics.cond_entropy(y_true, y_pred, contingency, a_i)
#accuracy = converted_metrics.accuracy(y_true, y_pred, contingency, PBV) Is equal to ARI in pair-based value context
precision = converted_metrics.precision(y_true, y_pred, contingency, PBV)
recall = converted_metrics.recall(y_true, y_pred, contingency, PBV)
falsealarm = converted_metrics.false_alarm_rate(y_true, y_pred,
contingency, PBV)
fm = converted_metrics.fowlkes_mallows(y_true, y_pred, contingency, PBV)
f1 = converted_metrics.f_beta_score(y_true, y_pred, 1, contingency, PBV,
precision, recall)
purity = new_metrics.purity(y_true, y_pred, contingency)
inversed_purity = new_metrics.inversed_purity(y_true, y_pred, contingency)
epratio = new_metrics.ep_ratio(y_true, y_pred, contingency, a_i, entropy,
purity)
jaccard = converted_metrics.jaccard_index(y_true, y_pred, contingency, PBV)
nmi = converted_metrics.normalized_mutual_information(
y_true, y_pred, contingency, a_i, b_j, mi, true_clustering_entropy,
predicted_clustering_entropy)
ri = new_metrics.rand_index(y_true, y_pred, contingency, PBV)
vi = new_metrics.variation_of_information(y_true, y_pred, contingency, a_i,
b_j)
# clustering error not calculated : always equal to 1 - accuracy
goodness = converted_metrics.goodness(y_true, y_pred, contingency, PBV)
bal_accuracy = converted_metrics.balanced_accuracy(y_true, y_pred,
contingency, PBV)
q2 = new_metrics.q2(y_true, y_pred, contingency, entropy, a_i, b_j)
metrics_dictionnary = {
"mi": mi,
"ari": ari,
"ami": ami,
"compl": compl,
"homog": homog,
"vmeasure": vmeasure,
"entropy": entropy,
"precision": precision,
"recall": recall,
"falsealarm": falsealarm,
"fm": fm,
"f1": f1,
"purity": purity,
"inv_purity": inversed_purity,
"epratio": epratio,
"jaccard": jaccard,
"nmi": nmi,
"ri": ri,
"vi": vi,
"goodness": goodness,
"balacc": bal_accuracy,
"q2": q2
}
if set(metrics_dictionnary.keys()) != CRITERION_LIST:
print(
"ERROR : One or several criterion are not computed in main.compute_prediction_metrics"
)
return metrics_dictionnary