Bounding Retraining Equivalence and the Deletion Floor in Materials Machine Unlearning
Abstract
In materials machine learning, closely related retained structures can sustain accurate property predictions even after removing a specific record, rendering post-deletion prediction error an ambiguous metric for machine unlearning. To resolve this ambiguity, we define the deletion floor as the expected target loss under a specified retraining procedure at the deleted request. Standard indistinguishability constraints yield a sharp interval bounding an update's target loss around this baseline reference. Theoretically, a conditional neighbor bound links a low deletion floor directly to retained fit, prediction regularity, and local label agreement, while an exact ridge identity isolates residual fit from the prediction change induced by record deletion. Empirically, controlled redundancy sweeps show an ≈ 8× drop in median normalized retraining loss when one retained relative remains after deletion. Across two distinct fitting regimes in a paired Materials Project study, the lower-floor regime also exhibits a larger prediction change on more than 50% of the shared requests. Systematic comparisons against approximate updates and the original model decouple deliberate target suppression from preserved overall model utility. Consequently, request-level unlearning evaluations should report reference loss, prediction change, and retained utility together, interpreting post-deletion accuracy against what retraining itself leaves behind.
Related
- The equivalence of quantum deletion and insertion errors on permutation-invariant codes
- From random patterns to nanophotonic structures: Electromagnetic field prediction without retraining
- Reliable Uncertainty Estimation for Machine-Learned Multigroup Cross Sections Without Retraining
- Federated Unlearning of Quantum Data via Classical Shadows
- Bounding the entanglement of a state from its spectrum
BibTeX
@misc{polat2026bounding,
title = {{Bounding Retraining Equivalence and the Deletion Floor in Materials Machine Unlearning}},
author = {Can Polat and Mustafa Kurban and Erchin Serpedin and Hasan Kurban},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.35635},
eprint = {2609.35635},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.35635},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Bounding Retraining Equivalence and the Deletion Floor in Materials Machine Unlearning
AU - Can Polat
AU - Mustafa Kurban
AU - Erchin Serpedin
AU - Hasan Kurban
PY - 2026
DA - 2026/09/28
PB - arXiv
DO - 10.48550/arxiv.2609.35635
UR - https://arxiv.org/abs/2609.35635
AB - In materials machine learning, closely related retained structures can sustain accurate property predictions even after removing a specific record, rendering post-deletion prediction error an ambiguous metric for machine unlearning. To resolve this ambiguity, we define the deletion floor as the expected target loss under a specified retraining procedure at the deleted request. Standard indistinguishability constraints yield a sharp interval bounding an update's target loss around this baseline reference. Theoretically, a conditional neighbor bound links a low deletion floor directly to retained fit, prediction regularity, and local label agreement, while an exact ridge identity isolates residual fit from the prediction change induced by record deletion. Empirically, controlled redundancy sweeps show an ≈ 8× drop in median normalized retraining loss when one retained relative remains after deletion. Across two distinct fitting regimes in a paired Materials Project study, the lower-floor regime also exhibits a larger prediction change on more than 50% of the shared requests. Systematic comparisons against approximate updates and the original model decouple deliberate target suppression from preserved overall model utility. Consequently, request-level unlearning evaluations should report reference loss, prediction change, and retained utility together, interpreting post-deletion accuracy against what retraining itself leaves behind.
ER -