prc implements a few functions to compute Precision-Recall curves, and their AUC (area under curve). The implementation is checked against scikit-learn.

Latest release

opam install prc

Latest commit

opam pin add -y prc https://github.com/pveber/prc.git

Precision-Recall curves are a useful representation of the performance of binary classification methods in the case there are many negative items. Problem is they are difficult to estimate from a finite sample. As an illustration, compare the following graph obtained in a particular case where the true precision-recall curve can be computed analytically (so-called binormal model):

The red curve shows the "true" precision-recall curve while the grey ones are the empirical estimated obtained from samples of size 1000. We see that there is a lot of variability, particularly in the low-recall region.

As a consequence, displaying a precision-recall curve can be misleading, and it is safer to report an estimate of its area under curve (AUC) along with a confidence interval. This is what this library provides.

Demo code is available in the demo directory. To run it, simply launch an interpreter via a dune utop demo command, and then call functions in Prc_demo. For instance to get an overview of the sampling distribution of several estimators under the binormal model, just type:

# Prc_demo.estimator_sampling_distribution ~sample_size:1000 ();;

This yields:

where each boxplot represents the distribution of an estimator (resp. binormal, trapezoidal and average precision) for a binormal model when drawing finite samples of size 1000.