Principle:Interpretml Interpret Score Tensor Harmonization
| Field | Value |
|---|---|
| Sources | Repo: InterpretML |
| Domains | Federated_Learning, Numerical_Methods |
| Updated | 2026-02-07 |
Overview
A numerical remapping procedure that transforms per-term score tensors from one bin definition to another through interpolation and redistribution.
Description
When merging EBMs with different bin boundaries, each model's score tensors must be remapped to a unified set of bins. Score Tensor Harmonization handles this by interpolating continuous feature scores between old and new bin boundaries, redistributing categorical bin scores when category dictionaries differ, and handling missing value bins. For multi-dimensional interaction terms, the remapping is applied across all dimensions.
Usage
Use this principle internally during EBM model merging when score tensors from different models need to be expressed in a common bin space.
Theoretical Basis
For continuous features, scores are interpolated proportionally when an old bin spans multiple new bins. For categorical features, scores are redistributed based on evidence weights.