EcoStack-Pro: an adaptive federated learning framework for interpretable ESG auditing across heterogeneous industrial...

Area of Science:

• Sustainable Finance
• Artificial Intelligence
• Data Privacy

Background:

• The increasing importance of environmental, social, and governance (ESG) metrics requires advanced auditing systems.
• Traditional centralized artificial intelligence (AI) models face challenges with data privacy regulations.
• Existing federated learning methods struggle with the statistical heterogeneity and data imbalance common in diverse industries.

Purpose of the Study:

• To develop a decentralized auditing framework that balances high-precision forecasting with data sovereignty for ESG metrics.
• To address the limitations of current AI and federated learning approaches in the context of sustainable finance.

Main Methods:

• Proposes EcoStack-Pro, a decentralized framework using a stacked ensemble of LightGBM, XGBoost, and Gradient Boosting regressors.
• Employs the Fed-GenAdaptive algorithm with a soft-gating mechanism for dynamic client contribution weighting.
• Utilizes Bayesian ridge meta-learning for optimization.

Main Results:

• Achieved a test-set R2 of 0.9614 on a stratified dataset of 21,400 firm-year observations across 10 industrial clients.
• Retained 98.2% of the predictive power of a centralized model while ensuring corporate privacy.
• Integrated Shapley additive explanations (SHAP) and local interpretable model-agnostic explanations (LIME) for enhanced model interpretability.

Conclusions:

• Adaptive, diverse ensemble strategies in federated learning can overcome limitations of single-model baselines.
• EcoStack-Pro provides a robust framework for secure, cross-sector sustainable finance auditing.
• The framework enhances interpretability of non-linear drivers in governance ratings.