Work

Rough-set regions as a local reliability signal for conformal prediction: when does Pawlak boundary membership help?

Year: 2026
Type: preprint
Abstract: Conformal prediction is a widely used framework for distribution-free uncertainty quantification. It guarantees that prediction sets contain the true label with a prescribed probability on average, but the average guarantee may hide groups of inputs on which the prediction sets fail. Rough set theory divides the input space into information granules without training any model. Granules whose training labels agree form the positive region, granules with conflicting labels form the boundary region, and attribute combinations absent from training form an unseen region. In this paper, a region-aware framework is proposed to examine whether this three-way partition is a useful reliability signal for conformal prediction. The framework uses the regions in two ways: as predictors of conformal miscoverage, and as a Mondrian taxonomy for group-wise calibration. Experiments are conducted on eight OpenML classification datasets, three levels of granulation and ten random splits. Pooled over all settings, region indicators reduced the held-out log loss of a miscoverage model by 0.0035 (95\% interval -0.00002 to 0.0081) and raised its area under the ROC curve (AUC) by 0.0054 (0.0006 to 0.0116). The effect was concentrated on Nursery and Car (AUC gains 0.023 and 0.014) and weakened as the share of unseen inputs grew (Spearman correlation -0.60). Region calibration restored coverage on boundary and unseen inputs at a few percent larger sets, and region-by-class calibration used 18.3\% smaller sets than confidence-bin-by-class calibration at 8.3 points lower worst confidence-bin coverage. The results indicate that rough-set regions are a local, granulation-dependent reliability signal rather than a general one.
Author: Balaraju M
Language: English
Open Access status: green
OpenAlex ID: W7235686935
Cites: 17
Cited by:
Related to: 10