Explainer implementation for LCPN #minor

scikit-learn-contrib / hiclass

A python library for hierarchical classification compatible with scikit-learn

BSD 3-Clause "New" or "Revised" License

113 stars 20 forks source link

Explainer implementation for LCPN #minor #108

Closed ashishpatel16 closed 5 months ago

ashishpatel16 commented 6 months ago

This PR aims to add the implementation for LCPN explainer. The major changes in the code are as follows -

Since the code is structured in a way that explain_with_xr method relies on get_traversed_nodes for accessing the correct nodes for traversing the local classifiers, changes are only required in the get_traversed_nodes function to return traversed nodes depending on the Hiclass model provided.
LCPN has the simplest way of returning the traversed nodes since each label in the prediction (lcpn.predict()) represents the node traversed at each hierarchy level. This is implemented in the corresponding method.
This PR also adds new test case for LCPN, unfortunately, the tests are not generalizable, hence not parametrized. This is because the structure and shape of shap values is completely different for different Hiclass models, depending on the way they make predictions.

codecov-commenter commented 6 months ago

Codecov Report

All modified and coverable lines are covered by tests :white_check_mark:

Project coverage is 96.87%. Comparing base (4595264) to head (5aa9998).

Additional details and impacted files

```diff @@ Coverage Diff @@ ## main #108 +/- ## ========================================== + Coverage 96.68% 96.87% +0.18% ========================================== Files 13 13 Lines 1268 1280 +12 ========================================== + Hits 1226 1240 +14 + Misses 42 40 -2 ```

:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.

ashishpatel16 commented 5 months ago

Hi @mirand863, this is the final version of the Explainer API for LCPN and LCPPN models. LCPL will follow soon.

The changes made here are as follows -

Updated readthedocs (Gallery of examples, Installation, API reference, Algorithm Overview)
Updated the xarray Dataset by aligning the explanation object along required coordinate dimensions.
Updated README to include an example of explaining Hiclass models.

I'm looking forward to your remarks. Hoping this will be the finally merged!

ashishpatel16 commented 5 months ago

Hi @mirand863, I have incorporated the following changes -

Dependencies now have shap version hardcoded to 0.44.1
Added more tests (code coverage is now ~96% with Explainer class having 91% coverage which will further increase when all classifiers are merged)
Added xarray and shap dependencies to docs/requirements.txt. Read the docs is now building fine.
Colab notebook for playing around with the explainer code.
Separated out code, tests and docs for LCPPN in #106 so now this PR only contains code for LCPN.

ashishpatel16 commented 5 months ago

Hi, @mirand863, Updated changes in this PR are as follows -

Merged the progress from main branch.
Colab notebook with updated code (refactored to use the hiclass.datasets module).
Gallery of examples > LCPN example.
Updated README with link to LCPN Colab notebook.

mirand863 commented 5 months ago

Hi @ashishpatel16

The colabs notebook does not work unless I update the git checkout to git checkout --track origin/lcpn_explainer Can you please update it?

ashishpatel16 commented 5 months ago

Hi @mirand863, Thanks for the review. I have updated the Colab notebook with the requested changes. Please check it and let me know if it works perfectly now! ✅

mirand863 commented 5 months ago

Hi @mirand863, Thanks for the review. I have updated the Colab notebook with the requested changes. Please check it and let me know if it works perfectly now! ✅

It works now, Thanks!