EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Leakage-resistant graph learning for risk propagation and risk-constrained control in construction supply networks

Qin Fu, Yao Wei, Yong Xiang

Frontiers in Physics · 04.11.2026 · 14

DOI: 10.3389/fphy.2026.1946795 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Frontiers in Physics, online erschienen am 04.11.2026. Verfasst von Qin Fu und 2 weiteren Autor:innen. Die Arbeit nennt 63 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Introduction

Construction supply network decisions require both predecision warning from information available before operations unfold and topology-aware diagnosis after the disruption context becomes visible. We introduce an auditable framework that separates these regimes through a strict start-of-week information contract and connects risk scores to safety-gated resource allocation.

Methods

We combined public SUCCESS construction-logistics records with a frozen queue-flow mechanism producing paired counterfactual outcomes on a heterogeneous network with 302 nodes and 1,277 edges. Models were trained on pilots A and B, selected on pilot C, and evaluated on pilot D. The strict contract excluded all week-t outcomes, service levels, and shock-source indicators. Complementary evaluations comprised context-aware diagnosis, a construction consolidation center (CCC) outage panel, leave-one-project-out evaluation, and an independent SupplyGraph boundary test. Six resource policies were compared through validation-only selection and a held-out Pareto audit.

Results

Under the strict contract, the graph convolutional network achieved the highest held-out area under the precision–recall curve (AUPRC) of 0.2397 and area under the receiver operating characteristic curve (AUROC) of 0.8161 for 4,252 node-weeks with 60 positives. In the context-aware regime, relational GraphSAGE achieved an AUPRC of 0.5702, exceeding the matched multilayer perceptron (MLP) by 0.0803. Across eight positive pairs involving CCC outages, GraphSAGE achieved an AUPRC of 0.5845 versus 0.4994 for the MLP, with a positive pair-specific advantage in each case. The calibrated-risk guardrail identified an aggregate service improvement of 0.827% but screened the allocation out because the affected sites and failure durations increased. The held-out frontier also identified a deficit-only allocation that improved unserved demand, affected sites, and failure durations without overriding the predeclared deployment rule.

Discussion

The framework combines leakage-resistant graph evaluation, context-aware relational diagnosis, project-level applicability testing, and risk-gated network control. Its explicit safety gate prevents aggregate efficiency gains from concealing propagation-risk transfer.

Zitieren

APA

Fu, Q., Wei, Y., & Xiang, Y. (2026). Leakage-resistant graph learning for risk propagation and risk-constrained control in construction supply networks. Frontiers in Physics, 14. https://doi.org/10.3389/fphy.2026.1946795

Vancouver

Fu Q, Wei Y, Xiang Y. Leakage-resistant graph learning for risk propagation and risk-constrained control in construction supply networks. Frontiers in Physics. 2026;14. doi:10.3389/fphy.2026.1946795

BibTeX
@article{Fu2026,
  author = {Fu, Qin and Wei, Yao and Xiang, Yong},
  title = {Leakage-resistant graph learning for risk propagation and risk-constrained control in construction supply networks},
  journal = {Frontiers in Physics},
  year = {2026},
  volume = {14},
  doi = {10.3389/fphy.2026.1946795}
}
Kostenlos starten →