AI Explanations Collapse When Supply Networks Fracture, Spectral Study Warns
· Source: bioengineer.org
In short
A new study from North-West University researchers argues that permutation-based explainability methods such as SHAP and LIME become mathematically invalid when a supply network loses algebraic connectivity, measured by the Fiedler eigenvalue of the hypergraph Laplacian. The authors introduce SpectralGuard, a framework producing probabilistic validity certificates from four spectral governance metrics, plus a λ-SHAP shrinkage estimator that reduced attribution error by up to 89.2% in simulations calibrated to South Africa's chronic medicines distribution programme. The work is simulation-based, with field validation on real procurement data planned for 2027.
Read Original Article — Read the full article on the publisher’s site.
In the issue of week 43, 2026: Amazon plans faster at scale as one customer cuts 52-week planning time 40%
Tags: explainable AI, SHAP, supply chain, pharmaceutical, hypergraph, governance