EXPLAINABLE SUPPLIER RISK PREDICTION FOR CRITICAL TELECOMMUNICATIONS INFRASTRUCTURE PROCUREMENT: AN EMPIRICAL XGBOOST–SHAP FRAMEWORK
Keywords:
Supplier risk prediction, Explainable artificial intelligence, XGBoost, SHAP, Monotonicity constraints, Critical infrastructure procurement, Telecommunications supply chain, Cost-sensitive learningAbstract
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.References
[1] Radu R, Amon C. The governance of 5G infrastructure: between path dependency and risk-based approaches. Journal of Cybersecurity, 2021, 7(1): tyab017.
[2] European Union Agency for Cybersecurity (ENISA). ENISA Threat Landscape for 5G Networks. Athens: ENISA, 2020.
[3] Enduring Security Framework. Potential Threat Vectors to 5G Infrastructure. NSA, CISA and ODNI, 2021.
[4] Ho W, Zheng T, Yildiz H, et al. Supply chain risk management: a literature review. International Journal of Production Research, 2015, 53(16): 5031-5069.
[5] Baryannis G, Validi S, Dani S, et al. Supply chain risk management and artificial intelligence: state of the art and future research directions. International Journal of Production Research, 2019, 57(7): 2179-2202.
[6] Baryannis G, Dani S, Antoniou G. Predicting supply chain risks using machine learning: the trade-off between performance and interpretability. Future Generation Computer Systems, 2019, 101: 993-1004.
[7] Brintrup A, Pak J, Ratiney D, et al. Supply chain data analytics for predicting supplier disruptions: a case study in complex asset manufacturing. International Journal of Production Research, 2020, 58(11): 3330-3341.
[8] Kosasih E E, Brintrup A. A machine learning approach for predicting hidden links in supply chain with graph neural networks. International Journal of Production Research, 2022, 60(17): 5380-5393.
[9] Chen T, Guestrin C. XGBoost: a scalable tree boosting system. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2016: 785-794.
[10] Friedman J H. Greedy function approximation: a gradient boosting machine. Annals of Statistics, 2001, 29(5): 1189-1232.
[11] Grinsztajn L, Oyallon E, Varoquaux G. Why do tree-based models still outperform deep learning on typical tabular data?. In: Advances in Neural Information Processing Systems, 2022, 35: 507-520.
[12] Shwartz-Ziv R, Armon A. Tabular data: deep learning is not all you need. Information Fusion, 2022, 81: 84-90.
[13] Lundberg S M, Lee S-I. A unified approach to interpreting model predictions. In: Advances in Neural Information Processing Systems, 2017, 30: 4765-4774.
[14] Lundberg S M, Erion G, Chen H, et al. From local explanations to global understanding with explainable AI for trees. Nature Machine Intelligence, 2020, 2(1): 56-67.
[15] Cano J-R, Gutiérrez P A, Krawczyk B, et al. Monotonic classification: an overview on algorithms, performance measures and data sets. Neurocomputing, 2019, 341: 168-182.
[16] Chen C-C, Li S-T. Credit rating with a monotonicity-constrained support vector machine model. Expert Systems with Applications, 2014, 41(16): 7235-7247.
[17] Chen D, Ye W. Monotonic neural additive models: pursuing regulated machine learning models for credit scoring. In: Proceedings of the 3rd ACM International Conference AI in Finance, 2022: 70-78.
[18] Slack D, Hilgard S, Jia E, et al. Fooling LIME and SHAP: adversarial attacks on post hoc explanation methods. In: Proceedings of AAAI/ACM Conference AI, Ethics, and Society, 2020: 180-186.
[19] Aas K, Jullum M, Løland A. Explaining individual predictions when features are dependent: more accurate approximations to Shapley values. Artificial Intelligence, 2021, 298: 103502.
[20] Constante F, Silva F, Pereira A. DataCo SMART SUPPLY CHAIN FOR BIG DATA ANALYSIS. Mendeley Data, V5, 2019. DOI: 10.17632/8gx2fvg2k6.5.
[21] Kaufman S, Rosset S, Perlich C, et al. Leakage in data mining: formulation, detection, and avoidance. ACM Transactions on Knowledge Discovery from Data, 2012, 6(4): Art. 15.
[22] Elkan C. The foundations of cost-sensitive learning. In: Proceedings of the 17th International Joint Conference on Artificial Intelligence, 2001: 973-978.
[23] Drummond C, Holte R C. Cost curves: an improved method for visualizing classifier performance. Machine Learning, 2006, 65(1): 95-130.
[24] Brintrup A, Wichmann P, Woodall P, et al. Predicting hidden links in supply networks. Complexity, 2018, 2018: Art. ID 9847925.
[25] Handfield R B, Sun H, Rothenberg L. Assessing supply chain risk for apparel production in low cost countries using newsfeed analysis. Supply Chain Management: An International Journal, 2020, 25(6): 803-821.
[26] Burstein G, Zuckerman I. Deconstructing risk factors for predicting risk assessment in supply chains using machine learning. Journal of Risk and Financial Management, 2023, 16(2): Art. 97.
[27] Altman E I. Financial ratios, discriminant analysis and the prediction of corporate bankruptcy. Journal of Finance, 1968, 23(4): 589-609.
[28] Ohlson J A. Financial ratios and the probabilistic prediction of bankruptcy. Journal of Accounting Research, 1980, 18(1): 109-131.
[29] Barboza F, Kimura H, Altman E. Machine learning models and bankruptcy prediction. Expert Systems with Applications, 2017, 83: 405-417.
[30] Liang D, Lu C-C, Tsai C-F, et al. Financial ratios and corporate governance indicators in bankruptcy prediction: a comprehensive study. European Journal of Operational Research, 2016, 252(2): 561-572.
[31] Ribeiro M T, Singh S, Guestrin C. 'Why should I trust you?' Explaining the predictions of any classifier. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2016: 1135-1144.
[32] Rudin C. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 2019, 1(5): 206-215.
[33] Barredo Arrieta A, Díaz-Rodríguez N, Del Ser J, et al. Explainable artificial intelligence (XAI): concepts, taxonomies, opportunities and challenges toward responsible AI. Information Fusion, 2020, 58: 82-115.
[34] Molnar C. Interpretable Machine Learning: A Guide for Making Black Box Models Explainable. 2nd ed. 2022.
[35] Saito T, Rehmsmeier M. The precision-recall plot is more informative than the ROC plot when evaluating binary classifiers on imbalanced datasets. PLoS ONE, 2015, 10(3): e0118432.