1

I have been trying to fiddle with sklearn metrics, particularly average_precision_score. However, I could only find one example of computing average_precision_score in the sklearn documentation and that too using SVM. Underneath is the code snippet and also the link to the documentation as a reference:

Documentation - Precision Recall Sklearn and code reference

# Create a simple classifier
classifier = svm.LinearSVC(random_state=random_state)
classifier.fit(X_train, y_train)
y_score = classifier.decision_function(X_test)
# Computing the avaerage_precision_score
from sklearn.metrics import average_precision_score
average_precision = average_precision_score(y_test, y_score)

Now my question is in the case above y_score is the output coming through decision_function (which predicts the confidence scores of samples) and y_test are classes. Considering there is no decision_function method for RandomForests as in the case of SVM, how to go about calculating y_score?

I have tried and seen people using both predict(self, X)(Predict class for X) and predict_proba(self, X)(Predict class probabilities for X) methods to compute average_precision_score. And my results have been very different using both these methods. With predict(self, X) I get 0.74 as average_precision_score and using predict_proba I get 0.94. My y_test are class labels with values (1, 0). I am a little confused as to what is the right thing. When to use predict vs predict_proba, and why are they resulting in so much different average precision scores. Any help would be highly appreciated.

Vipul Mehra
  • 141
  • 1
  • 2
  • 8

1 Answers1

0

With predict - you predict labels (assume it's 0 and 1) With predict_proba - you have probabilities of both labels for certain sample (result [0.1, 0.9] => this sample most likely 1 not 0).

The metric AP - is the order metric so the only order of prediction is important. So if ground truth vector [1, 0, 1] and prediction vector (via probabilities) [0.9, 0.7, 0.8] - AP give you 1.0 even probability to second example as 1 is 0.7.

Thus if you model make a mistake (missing the class) it's still a chance that the order will be more smooth. Thus for order metric like AP or ROC AUC it's better to put proba over predict.

Gleb Erofeev
  • 186
  • 2
  • 3