True... and that is inconsistent between roc_curve (which includes all thresholds) and precision_recall_curve. Maybe we should just add the smallest values as threshold for precision_recall_curve, too?

That would be backward-incompatible but would actually do what the docs say.

I can try this one.

Can you expand on "Maybe we should just add the smallest values as threshold for precision_recall_curve, too?"?

well the smallest score value seems to be dropped from threshold. We could not drop it.

my fault, i misread the 0.1 as 1.0 and did not understand the "smallest" part in your statement. taking a look now.

My explanation in #5091 (comment). TLDR: documentation not precisely correct, but code is.

Why is that the correct behavior? I find it surprising, and it seems to be inconsistent with roc_curve, but maybe I'm missing something?

Inspecting the output of the code you gave, never more then one value is dropped (but maybe just because it is very unlikely?)

len(thresholds) depends both on the unique values in y_scores and the index of the final y_scores value that is a true positive (after y_scores has been sorted desc.). When y_true values are generated uniformly, it's likely that even indices with low y_scores can be true positives which keeps len(y_scores) - len(thresholds) small.

right, len(np.unique(y_scores)) - len(thresholds) can be arbitrary. but len(precision) - len(thresholds) will always be 1.
I thought this wasn't about discrepancy between the unique scores and thresholds but between precisions and thresholds. I understand that we don't give the last threshold, I just don't understand why, and I feel that should be consistent between roc and precision_recall.

But then again, I have to recheck the behavior of roc_curve which seems to have a different behavior of dropping points.

Yes, we possibly should be consistent, but it depends in part on how auc
should be calculated for each, which is disputed elsewhere (e.g. #4223).
The difference in size might indeed be considered a historical quirk, but I
don't think it's harmful as long as documentation is sufficient.

On 15 August 2015 at 04:21, Andreas Mueller [email protected]
wrote:

right, len(np.unique(y_scores)) - len(thresholds) can be arbitrary. but len(precision)

• len(thresholds) will always be 1.
  I thought this wasn't about discrepancy between the unique scores and
  thresholds but between precisions and thresholds. I understand that we
  don't give the last threshold, I just don't understand why, and I feel that
  should be consistent between roc and precision_recall.

But then again, I have to recheck the behavior of roc_curve which seems
to have a different behavior of dropping points.

—
Reply to this email directly or view it on GitHub
#4996 (comment)
.

I'm considering this a duplicate of #4223, #5073 and closing it, while quick-fixing the documentation to say <= instead of := in 749f2a9.