def recall(y_true, y_pred, contingency=None, PBV=None):
if contingency is None:
contingency = utility.compute_contingency(y_true, y_pred)
if PBV is None:
PBV = utility.pair_based_values(contingency)
tp, _, fn, _ = PBV
return tp / (tp + fn)
def fowlkes_mallows(y_true, y_pred, contingency=None, PBV=None):
if contingency is None:
contingency = utility.compute_contingency(y_true, y_pred)
if PBV is None:
PBV = utility.pair_based_values(contingency)
tp, fp, fn, tn = PBV
return np.sqrt((tp * tp) / ((tp + fp) * (tp + fn)))
def jaccard_index(y_true, y_pred, contingency=None, PBV=None):
if contingency is None:
contingency = utility.compute_contingency(y_true, y_pred)
if PBV is None:
PBV = utility.pair_based_values(contingency)
tp, fp, fn, tn = PBV
return tp / (tp + fp + fn)
def accuracy(y_true, y_pred, contingency=None, PBV=None):
# Note : Identical to Rand Index criterion, in context of pair counting
if contingency is None:
contingency = utility.compute_contingency(y_true, y_pred)
if PBV is None:
PBV = utility.pair_based_values(contingency)
tp, fp, fn, tn = PBV
return (tp + tn) / (tp + tn + fp + fn)
def balanced_accuracy(y_true, y_pred, contingency=None, PBV=None):
if contingency is None:
contingency = utility.compute_contingency(y_true, y_pred)
if PBV is None:
PBV = utility.pair_based_values(contingency)
tp, fp, fn, tn = PBV
a = tp / (tp + fn)
b = tn / (tn + fp)
return 0.5 * a + 0.5 * b
def false_alarm_rate(y_true, y_pred, contingency=None, PBV=None):
if contingency is None:
contingency = utility.compute_contingency(y_true, y_pred)
if PBV == None:
PBV = utility.pair_based_values(contingency)
_, fp, fn, _ = PBV
if fp + fn == 0:
return 0
return fp / (fp + fn)
def clustering_error(y_true, y_pred, contingency=None, PBV=None):
# Note : It might be different to consider CE and Accuracy, especially considering subtle variation like micro/macro/weighted averaging.
# However, in data pairs context, CE is always equal to 1 - Accuracy
if contingency is None:
contingency = utility.compute_contingency(y_true, y_pred)
if PBV is None:
PBV = utility.pair_based_values(contingency)
print(
"'Clustering Error' is redundant criterion ; Please use Accuracy criterion instead, and compute 1 - Accuracy"
)
def goodness(y_true, y_pred, contingency=None, PBV=None, pre=None, rec=None):
if contingency is None:
contingency = utility.compute_contingency(y_true, y_pred)
if PBV is None:
PBV = utility.pair_based_values(contingency)
if pre is None:
pre = precision(y_true, y_pred, contingency, PBV)
if rec is None:
rec = recall(y_true, y_pred, contingency, PBV)
return 0.5 * (pre + rec)
def f_beta_score(y_true,
y_pred,
beta,
contingency=None,
PBV=None,
pre=None,
rec=None):
if contingency is None:
contingency = utility.compute_contingency(y_true, y_pred)
if PBV is None:
PBV = utility.pair_based_values(contingency)
if pre is None:
pre = precision(y_true, y_pred, contingency, PBV)
if rec is None:
rec = recall(y_true, y_pred, contingency, PBV)
beta2 = np.power(beta, 2)
if pre * rec == 0:
return 0
return ((1 + beta2) * pre * rec) / (beta2 * pre + rec)
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 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