ExplainerPFN: Zero-shot feature importance for tabular data
A new method called ExplainerPFN computes feature importance in supervised classification tasks without accessing the underlying model. Built on TabPFN, it learns a posterior mean attribution from synthetic structural causal datasets, targeting attributions that are 'true to the data' rather than 'true to the model'. This addresses the non-identifiability of Shapley values when only input data is available.
Key facts
- ExplainerPFN is a tabular foundation model
- It computes feature importance in a zero-shot setting
- It requires only input data distribution, not model evaluations
- It learns a posterior mean attribution under a meta-training prior
- Built on TabPFN
- Pretrained on synthetic structural causal datasets
- Targets attributions 'true to the data'
- Addresses non-identifiability of Shapley values
Entities
—