import numpy as np
from sklearn.metrics import precision_recall_curve, precision_recall_fscore_support,\
average_precision_score,auc
from matplotlib import pyplot as plt

#create some example labels
y=np.hstack([np.ones((100,)),np.zeros((900,))])

#create fake uniform probabilities of a dummy classifier
prob_random=np.ones_like(y)*.5

#Note that when recall is 0 (all samples assigned to class 0) precision returned is 1.
#However, this case is mathematically undefined (0/0 limit) and could make sense to set returned #precision to 0 rather than one as it is done in other function within sklearn
precision,recall,_ = precision_recall_curve(y,prob_random)

print('precision at 0 recall = {}'.format(precision[1]))

#For example, this function returns precision 0 and recall 0
P,R,_,_ =precision_recall_fscore_support(y,np.zeros_like(y))
print('precision at 0 recall = {}'.format(P[1]))

#A problem with the first definition is that 
# the AUC under the precision recall curve of a random classifier appears  artificially very good
plt.plot(recall,precision)
recall_def2=[1,0]
precision_def2=[.1,0]
plt.plot(recall_def2,precision_def2)
print('AUC current definition {}'.format(auc(recall,precision)))
print('AUC second definition {}'.format(auc(recall_def2,precision_def2)))