Skip to content

TabPFN explainability

The TabPFN-based filters expose a local explanation report for noisy instances.

What the report describes

For each instance, the report typically includes:

  • sample_idx: original sample index.
  • fold_idx: fold used to train the local model.
  • true_label: observed label.
  • oof_pred: out-of-fold predicted label.
  • confidence: probability associated with the predicted label.
  • top_k: the strongest SHAP contributions for the selected target class.

Default explanation target

By default, the SHAP values explain the predicted class. This means the report shows which features push the model toward the class selected by the filter.

How to read the signs

  • Positive SHAP values increase the score of the explained class.
  • Negative SHAP values decrease that score.
  • Larger absolute values indicate stronger local influence.

What the report does not say

  • It does not prove that a label is causally wrong.
  • It does not validate semantic correctness on its own.
  • It does not replace human domain review when causal interpretation is required.

Example interpretation

true_label = 2
oof_pred    = 1
confidence  = 0.51
top_k       = [feature_a, feature_b, feature_c]

This means the filter considers the sample closer to class 1 than to the observed label 2, but the decision is weak if the confidence is near 0.5.