Metrics support for tabular data

[annahedstroem]

Description of the problem

• Support tabular data for different categories of explanation quality such as Robustness, Faithfulness
• Lay out the technical requirements of the different categories:
  • [] Faithfulness
  • [] Robustness
  • [] Localisation
  • [] Complexity
  • [] Randomisation

Description of a solution

• Define minimal examples for perturbation given different dtypes
• Provide data_availability desc for all metrics if cannot support
• Warnings

Minimum acceptance criteria

• New tests added
• All tests

[aaarrti]

Randomisation and Complexity metrics already have support for tabular data, it is demonstrated in Tutorial_Getting_Started_with_Tabular_Data.ipynb.

[abarbosa94]

Hi @annahedstroem, @dilyabareeva, @aaarrti

Are there any updates about this task? I was doing some investigation using Tabular data as a proxy measure before going to NLP and I could benefit from this

[boltzmann-brain]

Can you confirm whether Randomization and Complexity are the only metrics that support tabular data, or are there any others that also do?

[aaarrti]

Can you confirm whether Randomization and Complexity are the only metrics that support tabular data, or are there any others that also do?

Hi @boltzmann-brain,

you can check which data domain metric supports with:

import quantus

quantus.metrics.<MetricClass>.data_applicability

If the return value contains <DataType.TABULAR: 'tabular'>, the metric was verified to support tabular data.