def adjust_count(y_pred, cm):
ncm = cm / cm.sum(1, dtype=numpy.float64)[:, numpy.newaxis]
tpr = ncm[1,1]
fpr = ncm[0,1]
pp = DE.to_bin_dist(y_pred)[1]
if tpr - fpr < .25:
raise TooLittleDifferenceException
new_pos = (pp - fpr) / float(tpr - fpr)
dist_est = numpy.array([1-new_pos, new_pos])
dist_est[dist_est<0] = 0
dist_est[dist_est>1] = 1
return dist_est / dist_est.sum()
def adjust_count(self, y_pred, cm):
ncm = cm / cm.sum(1, dtype=numpy.float64)[:, numpy.newaxis]
tpr = ncm[1,1]
fpr = ncm[0,1]
pp = DE.to_bin_dist(y_pred)[1]
if tpr == fpr:
new_pos = .5
else:
new_pos = (pp - fpr) / float(tpr - fpr)
dist_est = numpy.array([1-new_pos, new_pos])
if self.cap:
dist_est[dist_est<0] = 0
dist_est[dist_est>1] = 1
return dist_est / dist_est.sum()
def predict_binary(self, X_population, params):
clf, pos2neg, X, y = params
cost_fp = 1.0
cost_fn = 1.0
for i in range(self.itr_count):
new_label = clf.predict(X_population)
cn = Counter(new_label)
# We add a small prior (1) to prevent divide by zero error. This is
# not mentioned in the original paper, but we add this for fair
# comparison.
pos2neg_new = (cn[1] + 1) / float(cn[0] + 1)
cost_fp = pos2neg / pos2neg_new
# High cost means the examples are important, thus we should weigh
# them more.
clf = self.base_clf_class(class_weight = {0: cost_fn, 1:cost_fp})
clf.fit(X, y)
y_pred = clf.predict(X_population)
return DE.to_bin_dist(y_pred)
def predict_binary(self, X_population, params):
''' prediction outcome is length-2 array. '''
clf = params
y_pred = clf.predict(X_population)
return DE.to_bin_dist(y_pred)
