An illustrative prediction separates into an unpenalized linear trend and a curved kernel residual. Conditional KRR allows a chosen feature space to remain unpenalized while controlling the complexity of the residual component. Research source
About this paper
Combines unpenalized features with kernel ridge regression, giving a statistical framework for learning the residual component.
A residual kernel links conditional KRR to ordinary KRR and yields a test-risk comparison. Principal eigenfunctions and random features provide two related ways to choose the unpenalized component.
Citation
BibTeX
@inproceedings{takhanov2026conditionalkrr,
title = {Conditional KRR: Injecting Unpenalized Features into Kernel Methods with Applications to Kernel Thresholding},
author = {Takhanov, Rustem and Assylbekov, Zhenisbek},
year = {2026},
booktitle = {ICML 2026},
eprint = {2605.26067},
archivePrefix = {arXiv},
url = {https://arxiv.org/pdf/2605.26067}
}