Principle:Interpretml Interpret Model Finalization After Merge
| Field | Value |
|---|---|
| Sources | Repo: InterpretML |
| Domains | Federated_Learning, Machine_Learning |
| Updated | 2026-02-07 |
Overview
A postprocessing procedure that finalizes a merged EBM by aggregating harmonized scores, cleaning unused bins, and generating presentation-ready term metadata.
Description
After score tensors have been harmonized and averaged across models during merging, the merged model must be finalized. This involves calling process_terms to compute weighted averages and standard deviations, removing unused bin levels, generating human-readable term names, and zeroing out tensor entries that have no evidence. This is the same aggregation principle used in standard EBM training but applied to merged model outputs.
Usage
Use this principle as the final step of EBM model merging, after bin harmonization and score averaging.
Theoretical Basis
Same as Model_Aggregation_And_Postprocessing but applied to merged (rather than bagged) model outputs. The averaging and standard deviation formulas are identical.