Partial Dependence Beyond the Data: Support Diagnostics and Common-Reference Feature Effects (research archive)
Year: 2026
Type: software
Abstract: Research code, simulation pipeline, and manuscript sources accompanying Partial Dependence Beyond the Data: Support Diagnostics and Common-Reference Feature Effects. Partial-dependence plots may evaluate a fitted model at feature combinations that the data represent poorly. This archive turns that problem into an explicit support audit and characterizes when a single reference population remains valid over a displayed grid. It contains in-support-coverage (ISC), hard common-support partial-dependence (CSPD), and relaxed common-reference (RCPD) estimators, an exact empirical solver for interval validity, ALE, conditional-curve and subgroup-PDP baselines, and an independent audit split that certifies simultaneous query validity with explicit abstention when no candidate reference can be certified. The archive includes 34 simulation studies, a targets pipeline, 229 unit tests, and the Quarto manuscript sources. Stored target objects and figure panels are included, so both PDFs can be re-rendered without rerunning the pipeline. The R environment is pinned in renv.lock (R 4.5.3, 79 packages). No raw data is redistributed. data/README.md provides download commands, licences, and SHA-256 checksums for the Wine Quality, Bank Marketing, Online Shoppers, and California Housing datasets.
Author: Bücker, Michael
Institution: FH Münster University of Applied Sciences
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7214005477