An explainable graph learning framework for carbon footprint prediction: Integrating SHAP-based policy insights and country similarity networks

Accurate prediction of carbon footprint dynamics is essential for climate policy design and sustainable development planning, yet complex nonlinear interactions among energy systems, environmental factors, and socioeconomic activities make estimation challenging. To address this gap, this study proposes an integrated framework combining machine learning, explainable artificial intelligence, and graph neural networks to predict carbon footprints across 27 European countries from 2015 to 2023. Evaluating several algorithms revealed that tree-based models substantially outperformed linear approaches, with LightGBM achieving the highest machine learning performance. SHAP analyses identified fossil fuel dependency, municipal waste generation, environmental impacts, protection capacity, and climate initiatives as the primary predictive determinants. Furthermore, incorporating a country similarity network into Graph Neural Network architectures improved predictive accuracy, with the Graph Attention Network (GAT) delivering the overall best predictive performance. Robustness checks confirmed model stability across various configurations, demonstrating that carbon footprint formation is driven by both national indicators and structural cross-country relationships. Ultimately, this framework bridges a methodological gap in environmental forecasting, offering robust predictive tools for policymakers and actionable, feature-level insights for targeted climate mitigation strategies.