mmschlk / shapiq

Shapley Interactions for Machine Learning

Home Page:https://shapiq.readthedocs.io

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Allow multiple instances to explain/compute

mmschlk opened this issue · comments

Description

Add functionality to the package to allow for calculation of interactions for multiple instances/games to allow for the following scenario.

A user wants to put a 2d matrix of instances/games to explain into the explainer and expects interaction scores for each of those.

Potential Solutions

At the moment it is a bit unclear how to address this. Depending how deeply we integrate this will require more or less refactoring of the package. The two extremes of solutions are as follows:

  • Lazy/Light Integration: We can abstract the 2d input away inside the explainer/approximators and just loop over the instances (potentially in parallel). This requires not much rework but will probably be not that efficient.
  • Tight Integration: We can integrate the computations directly into our approximators (or explainers) by allowing the approximate method to accept a list of games instead of a single game. Then the approximate code needs to be adjusted. Parallelization under the hood inside the games/imputers can improve the performance even more.

This is an important issue that will be tackled at some point. It touches on the whole library.