Knowledge Graph-Augmented Large Language Models for Explainable Supply Chain Risk Intelligence

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The quality of large language model responses is affected by the inherent opaqueness of their architectures, weakening trustworthiness and limiting their applicability in domains where every recommendation must be justifiable [1]. In supply chain risk management, this is not a nicety but a requirement: predictive analytics that cannot be interpreted cannot be integrated into risk-related decision processes [2]. This paper argues that knowledge graphs are the missing substrate that converts LLMs from opaque reasoners into explainable risk-intelligence engines: the graph supplies the structure, the LLM supplies the narrative, and evidence grounding supplies the audit trail. We propose KGX-SC, a knowledge graph-augmented framework coupling (i) automatic KG construction from unstructured public data, (ii) graph-based retrieval, (iii) graph-neural-network risk reasoning with model-internal evidence extraction, (iv) evidence-grounded natural-language explanation generation, and (v) verification and governance. The framework consolidates the reported evidence envelope: a graph-based LLM framework raising ROUGE-1 F1 from 0.42 to 0.67 — a 59% improvement over the best comparative methods — across entity classification, link prediction, and reasoning [1]; a temporal-graph-attention predictor coupled with structured LLM reasoning that achieves test AUC 0.761, AP 0.344, and recall 0.504 under a strict chronological split while producing early-warning explanations with 99.6% directional consistency against underlying statistical evidence [3]; knowledge-graph-augmented RAG outperforming traditional RAG across an eight-function supply chain benchmark, with smaller open-weight models achieving notable gains that reduce the performance gap with state-of-the-art frontier models [4]; and RAG-plus-KG decision support improving decision accuracy, reasoning transparency, and context relevance over either technology alone, particularly for cross-domain reasoning and ambiguous queries [5]. The paper argues that explainability in supply chain risk intelligence is achieved not by simplifying models but by grounding them: graph structure-plus-evidence turns every risk narrative into an auditable claim.