Sparse Reconciliation

[mcsqr]

Using sparse matrices to reduce the computational time and memory appetite of a subset of the reconciliation methods.

The PR contains MinTraceSparse which supports the diagonal methods (ols, wls_struct, and wls_var), the speed-up is very considerable for datasets of ~2k-20k time series (the most I checked so far).

It also contains BottomUpSparse, which doesn't really show speed-up until 20k, but should hopefully help with even bigger datasets.

Note: This version still doesn't guarantee that no O(N^2) dense matrix is instantiated, it could probably be checked/improved by testing on even bigger datasets.

Also likely some docs and tests are missing, tbd.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[CLAassistant]

All committers have signed the CLA.

[kdgutier]

Thanks, @mcsqr
This is awesome; we will take a look.

From the execution of the CI tests, we will need to fix the scikit-learn version to have the updated sparse OneHotEncoder. This would need to be done here:
https://github.com/Nixtla/hierarchicalforecast/blob/main/environment.yml
https://github.com/Nixtla/hierarchicalforecast/blob/main/settings.ini
I don't know if we will need to increase min_python > 3.7

[mcsqr]

CI tests are fixed. Note that sparse argument in OneHotEncoder is deprecated and scheduled for removal, so at some point it'll also break, but for now it looks ok.

[kdgutier]

Awesome @mcsqr,

My most important comment is on the declaration of S_sparse in core.py.
Thanks a lot for your contribution.