It's really amazing!! Why are the calculation results of the AUC (recall, precision) function and the average precision score function significantly different?

[kourenke]

Describe the bug

Method 1:
Precision, recall, _=Precision-Recall_curve()
A=auc (recall, precision)
Method 2:
B=average precision score ()
Method 1 and Method 2 both calculate the area under the PR curve, but why are the results significantly different, i.e. a ≠ b

Steps/Code to Reproduce

Method 1:
Precision, recall, _=Precision-Recall_curve()
A=auc (recall, precision)
Method 2:
B=average precision score ()
Method 1 and Method 2 both calculate the area under the PR curve, but why are the results significantly different, i.e. A ≠ B

Expected Results

A = B

Actual Results

A ≠ B

Versions

sklearn: 1.0.2

[MichaelSkowronek]

AUC uses the trapezoidal rule for interpolation between the points. The average precision score utilizes an interpolation scheme that can be described as a step function. Therefore the difference. For more information, see the user guide: https://scikit-learn.org/stable/modules/model_evaluation.html#precision-recall-and-f-measures .

[glemaitre]

Indeed this is linked with the following comment:

The average precision (cf. average_precision_score) in scikit-learn is computed without any interpolation. To be consistent with this metric, the precision-recall curve is plotted without any interpolation as well (step-wise style).

You can change this style by passing the keyword argument drawstyle="default". However, the curve will not be strictly consistent with the reported average precision.

The computation of average precision is therefore not trivial :)