def compute_qmean(Y_label, Y_pred):
pi_array = []
for i in range(2):
pi_a = np.count_nonzero(Y_label == i) / len(Y_label)
pi_array.append(pi_a)
confusion_matrix = get_confusion_matrix(Y_label, Y_pred)
sum_sq_form = 0
for i in range(2):
sum_sq_form += (1 - (confusion_matrix[i][i] / pi_array[i]))**2
return np.sqrt(sum_sq_form / 2)
def compute_gmean(Y_label, Y_pred):
pi_array = []
for i in range(2):
pi_a = np.count_nonzero(Y_label == i) / len(Y_label)
pi_array.append(pi_a)
confusion_matrix = get_confusion_matrix(Y_label, Y_pred)
prod_form = 1
for i in range(2):
prod_form *= confusion_matrix[i][i] / pi_array[i]
return 1 - (prod_form)**(1 / 2)
def compute_hmean(Y_label, Y_pred):
pi_array = []
for i in range(2):
pi_a = np.count_nonzero(Y_label == i) / len(Y_label)
pi_array.append(pi_a)
confusion_matrix = get_confusion_matrix(Y_label, Y_pred)
sum_formulae = 0
for i in range(2):
sum_formulae += pi_array[i] / confusion_matrix[i][i]
return 1 - (2 / sum_formulae)
def compute_KLD(Y_label, Y_pred):
n_class = 2
cm = get_confusion_matrix(Y_label, Y_pred)
pi_array = []
for i in range(n_class):
pi_a = np.count_nonzero(Y_label == i)
pi_array.append(pi_a / len(Y_label))
kld_error = 0
for i in range(n_class):
sum_col_i = 0
for j in range(n_class):
sum_col_i += cm[j, i]
kld_error += pi_array[i] * np.log(pi_array[i] / sum_col_i)
return kld_error
def compute_EOpp(Y_label, Y_pred, X_pro_feature):
n_class = 2
final_conf = np.zeros(shape=(n_class, n_class))
cm = []
unique_a = len(np.unique(X_pro_feature))
for a in np.unique(X_pro_feature).astype(int):
indices_a = X_pro_feature == a
Y_label_a = Y_label[indices_a]
Y_pred_a = Y_pred[indices_a]
cm_a = get_confusion_matrix(Y_label_a, Y_pred_a)
cm_a = cm_a / (cm_a[1, 0] + cm_a[1, 1])
final_conf += cm_a
cm.append(cm_a)
odds_per_a_1 = np.zeros(unique_a)
for i in range(unique_a):
cm_a = cm[i]
odds_per_a_1[i] = abs(cm_a[1, 1] - (1 / unique_a) * (final_conf[1, 1]))
return max(odds_per_a_1)
def compute_DP(Y_label, Y_pred, X_pro_feature):
final_conf = np.zeros(shape=(2, 2))
cm = []
unique_a = len(np.unique(X_pro_feature))
for a in np.unique(X_pro_feature).astype(int):
indices_a = X_pro_feature == a
Y_label_a = Y_label[indices_a]
Y_pred_a = Y_pred[indices_a]
cm_a = get_confusion_matrix(Y_label_a, Y_pred_a)
final_conf += cm_a
cm.append(cm_a)
parity_per_a = np.zeros(unique_a)
for i in range(unique_a):
cm_a = cm[i]
parity_per_a[i] = abs(cm_a[0, 1] + cm_a[1, 1] - (1 / unique_a) *
(final_conf[0, 1] + final_conf[1, 1]))
return max(parity_per_a)
def compute_linear_loss(Y_label, Y_pred):
confusion_matrix = get_confusion_matrix(Y_label, Y_pred)
return 1 - confusion_matrix[0][0] - confusion_matrix[1][1]
def compute_COV(Y_label, Y_pred):
n_class = 2
cm = get_confusion_matrix(Y_label, Y_pred)
return cm[0, 1] + cm[1, 1]