Robust Bottleneck Prioritization from Event Logs Under Temporal Shift: A Wasserstein–CVaR Diagnostic Framework
- 1Faculty of Economic Sciences, Lucian Blaga University of Sibiu, Calea Dumbrăvii, no. 17, 550324 Sibiu, Romania
Abstract
Bottleneck prioritization selects a small set of process transitions for investigation when historical delay patterns may change. This study examines frequency-weighted Wasserstein–Conditional Value-at-Risk scores for capped inter-event elapsed times. The decision snapshots contain only transitions whose origin and destination events precede the start of Q3. Under one-dimensional Wasserstein-1 ambiguity on a bounded interval, worst-case CVaR has an exact capped affine expression. Its saturation threshold is the transport budget needed to place the upper tail at the cap. A normalized radius-to-threshold ratio measures the remaining severity headroom, separately from portfolio discrimination. With a 90-day cap and a 0.95 tail level, 33 of 61 BPI Challenge 2019 candidates saturate. Original DRO and empirical tail select the same five transitions, capturing 66.97% of subsequent mean burden, compared with 70.23% for mean burden. The paired original-DRO-minus-empirical-tail difference is exactly zero. In BPI Challenge 2017, three of 41 candidates saturate. Clipping the radius below the threshold removes induced saturation, but the four severity-based methods retain identical top-five membership and capture 70.237% of mean burden. Calibration, observation-maturity and resampling sensitivities delimit these comparisons. The diagnostic identifies saturation of the risk component. It does not establish that saturation necessarily changes the selected portfolio or that clipping improves future prioritization.
Related
- ML-CLAPPS: A Probabilistic Framework for Imbalanced Multilabel Learning under Class Prioritization
- A Unified Framework for Wasserstein Convergence of ULMC Methods beyond Log-Concavity: Old and New
- A Mathematical Modeling Framework for Semantic–Temporal Graph Learning in Log Anomaly Detection
- Integrating Data Center Flexibility into Power System Operations: A Framework for Temporal Redispatch
- Explainable machine learning framework with Granger-based temporal predictive validation for monitoring fiscal risk in national health insurance expenditure
BibTeX
@article{cristescu2026robust,
title = {{Robust Bottleneck Prioritization from Event Logs Under Temporal Shift: A Wasserstein–CVaR Diagnostic Framework}},
author = {Marian Pompiliu Cristescu},
journal = {Mathematics},
year = {2026},
month = oct,
doi = {10.3390/math14213947},
url = {https://www.mdpi.com/2227-7390/14/21/3947},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Robust Bottleneck Prioritization from Event Logs Under Temporal Shift: A Wasserstein–CVaR Diagnostic Framework
AU - Marian Pompiliu Cristescu
PY - 2026
DA - 2026/10/29
JO - Mathematics
DO - 10.3390/math14213947
UR - https://www.mdpi.com/2227-7390/14/21/3947
AB - Bottleneck prioritization selects a small set of process transitions for investigation when historical delay patterns may change. This study examines frequency-weighted Wasserstein–Conditional Value-at-Risk scores for capped inter-event elapsed times. The decision snapshots contain only transitions whose origin and destination events precede the start of Q3. Under one-dimensional Wasserstein-1 ambiguity on a bounded interval, worst-case CVaR has an exact capped affine expression. Its saturation threshold is the transport budget needed to place the upper tail at the cap. A normalized radius-to-threshold ratio measures the remaining severity headroom, separately from portfolio discrimination. With a 90-day cap and a 0.95 tail level, 33 of 61 BPI Challenge 2019 candidates saturate. Original DRO and empirical tail select the same five transitions, capturing 66.97% of subsequent mean burden, compared with 70.23% for mean burden. The paired original-DRO-minus-empirical-tail difference is exactly zero. In BPI Challenge 2017, three of 41 candidates saturate. Clipping the radius below the threshold removes induced saturation, but the four severity-based methods retain identical top-five membership and capture 70.237% of mean burden. Calibration, observation-maturity and resampling sensitivities delimit these comparisons. The diagnostic identifies saturation of the risk component. It does not establish that saturation necessarily changes the selected portfolio or that clipping improves future prioritization.
SN - 2227-7390
ER -