Calibrated cardinality estimation under workload drift — confidence intervals that hold without distributional assumptions, with provable coverage. Our estimator wraps any black-box learned cardinality model with a calibration layer based on conformal prediction, giving distribution-free guarantees that survive arbitrary shifts in the query workload.
Paper ICML 2026 DBMS
Learned Cardinality with Distribution-Free Confidence
arXiv:2604.18221
Cite this paper
Learned Cardinality with Distribution-Free Confidence
@inproceedings{park2026card,
title = {Learned Cardinality with Distribution-Free Confidence},
author = {Jihoon Park and Aylin Demir and Sasha Volkov},
booktitle = {ICML 2026},
year = {2026},
url = {https://arxiv.org/abs/2604.18221}
}