How to manually perform what PrecisionRecallDisplay library does

iacisme · 6 May 2022 05:05

Hi guys;

I’m interested in knowing how to manually generate this precision-recall curve. I want to use the Plotly graphing library, not the matplotlib one.

As such, I’m trying to break these steps down and would appreciate some guidance in doing so, as all the libraries that I would have used to do this are deprecated in this version of scikit-learn.

My goal in doing this course was to replicate every graph shown, using the Plotly graphing library. I’ve gotten to this point in the course figuring things out on my own, but now I’ve hit a wall.

I’m more than willing to share all my notebooks for this course, as they show how to do these graphs both in matplotlib, as well as plotly.

The plotly library has helped me understand concepts like hyper parameters, and what hyper parameter tuning does, in addition to other features.

Thus, I’d really like to figure out how to do this graph, but all references online are either depreciated (They worked in the previous version of the course, but since then there’s a new version of scikit-learn), or all point to PrecisionRecallDisplay library.

Thank you!

glemaitre58 · 6 May 2022 08:53

You can probably check the source code of the display:

github.com

scikit-learn/scikit-learn/blob/baf828ca1/sklearn/metrics/_plot/precision_recall_curve.py#L12

    
      
          from .base import _get_response
          
          
from .. import average_precision_score
          from .. import precision_recall_curve
          from .._base import _check_pos_label_consistency
          from .._classification import check_consistent_length
          
          
from ...utils import check_matplotlib_support, deprecated
          
          

          
class PrecisionRecallDisplay:
              """Precision Recall visualization.
          
          
    It is recommend to use
              :func:`~sklearn.metrics.PrecisionRecallDisplay.from_estimator` or
              :func:`~sklearn.metrics.PrecisionRecallDisplay.from_predictions` to create
              a :class:`~sklearn.metrics.PredictionRecallDisplay`. All parameters are
              stored as attributes.
          
          
    Read more in the :ref:`User Guide <visualizations>`.

But in short, you need to call precision_recall_curve as there: scikit-learn/precision_recall_curve.py at baf828ca126bcb2c0ad813226963621cafe38adb · scikit-learn/scikit-learn · GitHub

Then use this information and plot it. So look at the plot method then: scikit-learn/precision_recall_curve.py at baf828ca126bcb2c0ad813226963621cafe38adb · scikit-learn/scikit-learn · GitHub

The important point for precision-recall curve is to not linear interpolate between the point that is the default. You can see in the implementation that we use the following option in matplotlib:

line_kwargs = {"drawstyle": "steps-post"}

iacisme · 6 May 2022 16:46

Nice!

Thank you for this suggestion, it didn’t even cross my mind to look at source code.

I will look at this and break it down.

I appreciate you pointing me in the right direction.

Cheers.

–Igor