Principle:Interpretml Interpret Interaction Detection
| Metadata | |
|---|---|
| Sources | FAST, InterpretML |
| Domains | Feature_Engineering, Interpretability |
| Last Updated | 2026-02-07 12:00 GMT |
Overview
An algorithm that identifies pairs of features whose joint effect on the target exceeds the sum of their individual effects.
Description
Interaction Detection finds feature pairs with significant interaction effects, meaning the combined effect of two features cannot be explained by summing their individual contributions. The FAST (Feature Alternating Statistical Testing) algorithm measures interaction strength by fitting a 2D model to each candidate pair and comparing it to the sum of 1D models. This is crucial for EBMs because they can include pairwise interaction terms to capture non-additive relationships while remaining interpretable.
Usage
Use this principle after initial model fitting (main effects) to discover which feature pairs should be added as interaction terms. This step is optional but significantly improves model accuracy when true interactions exist in the data.
Theoretical Basis
Interaction strength for features (i,j) is measured as:
where Gain measures the reduction in loss from fitting a model to the given feature(s). The FAST algorithm efficiently computes this for all pairs by leveraging greedy cyclic boosting on 2D histograms.