GreenAudit-GNN: Sector-Stratified Benchmarking and Dual-Method Explainability Validation for ESG Anomaly Detection

Abstract

Mandatory ESG disclosure regimes such as India's Business Responsibility and Sustainability Reporting mandate produce a disclosure corpus that manual audit cannot review at scale. Two earlier papers in this programme reviewed the AI-for-ESG literature and proposed GreenAudit-GNN, a seven-stage framework combining NLP and a graph neural network for explainable ESG anomaly detection, piloted on 58 audited claim-KPI pairs from 40 companies. That pilot trained a classifier and benchmarked it against four unsupervised baselines, but left two things only partly addressed: cross-sector performance comparison, and validating the framework's explanations for auditor and regulator use. This paper closes both gaps on the same audited dataset of 38 companies. We retain every model's per-company prediction and group the 23 disclosed sectors into ten broader industry buckets, giving a descriptive cross-sector comparison of weak-label prevalence and model flag rates. We then implement and cross-check two independent local-explanation methods for the trained classifier: an exact closed-form linear-SHAP decomposition and an independently reimplemented LIME-style local surrogate. The two agree on direction for every company and correlate strongly in magnitude across the two model features. Finally, we introduce a gradient-by-input attribution that splits each company's anomaly score into its own reported trend versus its sector-peer average, the programme's first auditor-readable, evidence-linked explanation. Every result follows the same honesty discipline as the earlier papers: sector buckets of one to seven companies support no inferential claim, no human auditor or regulator has yet reviewed these explanations, and every number is a real, seeded, reproducible computation on audited disclosure data.