Use logits per default when explaining models like tabpfn.

[mmschlk]

Describe the workflow you want to enable

If a model has a predict_logits function then this should be explained per default.

Explanations for classifiers should happen in the logits space and not in the probabilities space per default. Currently users can pass a model's predict_logits function to shapiq and get this explained. However, novice users may pass the model, where we currently are not using the predict_logits function as default.

Describe your proposed solution

Suggested implementation:

Implementation

Prioritize the predict_logits function when parsing a model over predict and predict_proba

Test

Test weather predict_logits gets picked over predict_proba and explained for a (lightweight) model that implements both.

Describe alternatives you've considered, if relevant

No response

Additional context

This got recently discussed/addressed in tabpfn (PriorLabs/TabPFN#374). Newer tabpfn versions should expose predict_logits now.

Impact

High (Major improvement)