EXPLAINABLE SUPPLIER RISK PREDICTION FOR CRITICAL TELECOMMUNICATIONS INFRASTRUCTURE PROCUREMENT: AN EMPIRICAL XGBOOST–SHAP FRAMEWORK
Social Science and Management · 2024
DOI: 10.61784/ssm4094 ↗Kurz gefasst
Fachartikel in der Zeitschrift Social Science and Management, 2024. Verfasst von MengLin Bian und 2 weiteren Autor:innen. Die Arbeit wurde bisher 3-mal zitiert (Crossref). Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Telecommunications networks are critical national infrastructure, yet their equipment supply base is highly concentrated, so a single supplier failure can propagate into a national service outage. Machine-learning supplier-risk models are increasingly proposed for procurement screening, but they are almost always evaluated on discrimination alone: nobody asks whether the resulting explanations are directionally sensible, whether they survive a model refit, or whether the model actually reduces procurement cost. We present CTI-SRP, a two-tier explainable framework that decomposes supplier risk into financial-viability risk and operational delivery risk, couples a monotonicity-constrained XGBoost learner to exact TreeSHAP attribution, and adds an audit layer with three new explanation-quality metrics — Explanation Coherence Rate (ECR), Local Coherence Rate (LCR) and Monotonicity Violation Rate (MVR) — plus an Explanation Stability Index (ESI) and a cost-sensitive decision layer. Experiments use two real datasets (6,819 listed firms with 95 financial ratios; 180,519 order lines) and a transparently specified synthetic telecom cohort in which the true drivers, signs and interactions are known by construction. Encoding 46 accounting-theory sign priors as monotone constraints raises ECR from 0.870 to 0.978 and drives MVR from 0.345 to exactly zero at no cost in AUC (0.9547→0.9573). TreeSHAP nearly doubles top-5 driver stability over gain importance (Jaccard 0.446 vs 0.247). A leakage audit shows that the near-perfect delivery-risk accuracy reported in prior work collapses from AUC 0.992 to 0.733 once post-shipment features are removed. On the synthetic cohort the pipeline recovers all 16 true drivers with zero false drivers, all three true interactions in the top three, and a 0.953 correlation between true coefficient magnitude and mean |SHAP|. A criticality-weighted composite index captures 62.7% of realised loss within a 10% audit budget, versus 48.2% for probability ranking alone.
Zitieren
APA
Bian, M., Wang, Y., & Lin, Y. (2024). EXPLAINABLE SUPPLIER RISK PREDICTION FOR CRITICAL TELECOMMUNICATIONS INFRASTRUCTURE PROCUREMENT: AN EMPIRICAL XGBOOST–SHAP FRAMEWORK. Social Science and Management. https://doi.org/10.61784/ssm4094
Vancouver
Bian M, Wang Y, Lin Y. EXPLAINABLE SUPPLIER RISK PREDICTION FOR CRITICAL TELECOMMUNICATIONS INFRASTRUCTURE PROCUREMENT: AN EMPIRICAL XGBOOST–SHAP FRAMEWORK. Social Science and Management. 2024. doi:10.61784/ssm4094
BibTeX
@article{Bian2024,
author = {Bian, MengLin and Wang, Ying and Lin, YiZhen},
title = {EXPLAINABLE SUPPLIER RISK PREDICTION FOR CRITICAL TELECOMMUNICATIONS INFRASTRUCTURE PROCUREMENT: AN EMPIRICAL XGBOOST–SHAP FRAMEWORK},
journal = {Social Science and Management},
year = {2024},
doi = {10.61784/ssm4094}
}
EUVIMED