Title: Conditional PED-ANOVA: Assessing Hyperparameter Importance Within Hierarchical and Dynamic Search Spaces

Abstract:

This paper introduces conditional PED-ANOVA (condPED-ANOVA), a robust framework designed to estimate hyperparameter importance (HPI) specifically within conditional search spaces. In such environments, the availability or range of a hyperparameter is contingent upon the values of other hyperparameters. While the foundational PED-ANOVA method offers a swift and efficient mechanism for calculating HPI in the highest-performing areas of a search space, its reliance on a static, unconditional structure renders it unsuitable for handling conditional dependencies. To overcome this limitation, we define a conditional HPI metric tailored to top-performing regions and develop a closed-form estimator that precisely captures shifts in conditional activation and parameter domains. Our experimental results demonstrate that straightforward adaptations of current HPI estimators often produce misleading or unintelligible results in conditional contexts. In contrast, condPED-ANOVA consistently delivers interpretable importance scores that accurately mirror the underlying conditional architecture. The source code for this method is publicly accessible at https://github.com/kAIto47802/condPED-ANOVA.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC