scipy isotonic regression

[nicholasloveday]

Currently sklearn is only used in isotonic regression when functional="mean". This is because the sklearn implementation is about 4 times faster than implementing it ourselves.

Recently scipy added an isotonic regression function https://docs.scipy.org/doc/scipy/reference/generated/scipy.optimize.isotonic_regression.html .

We should investigate if we can use that and drop the sklearn dependency

[tennlee]

For anyone wanting to do this, there should be plenty of test coverage in scores already. The work would include:

• Checking whether the scipy implementation has the same functionality as that of sklearn
• Checking whether the scipy implementation is as fast/efficient as that of sklearn
• Change the functional code in scores to use the new scipy method
• Check the unit tests all pass, both with and without dask being installed in the environment
• If successful, remove sklearn from the dependencies in pyproject.toml

[nicholasloveday]

Additionally, the current implementation uses scipy's interpolate.interp1d which is a legacy function that won't be supported in the future. We should explore replacements

https://docs.scipy.org/doc/scipy/tutorial/interpolate/1D.html#legacy-interface-for-1-d-interpolation-interp1d