Dynamic Glocal XAI Framework for Machine Learning-Based Prediction of Education ROI with Hierarchical Explainability

Abstract

We propose a novel Dynamic Glocal XAI framework for predicting education return on investment (ROI) that simultaneously addresses predictive accuracy and interpretability for diverse stakeholders. The framework integrates gradient-boosted trees with dynamic explainable AI (XAI) mechanisms, enabling both global trends and local instance-level insights through hierarchical clustering of Shapley-based explanations. The model optimizes a regularized objective function to minimize prediction errors while maintaining computational efficiency through TreeSHAP approximations. Moreover, the system introduces a hierarchical clustering module that groups similar explanations using a dual-term metric combining Euclidean distance and distributional divergence, thereby capturing both magnitude and pattern similarities in feature contributions. A gating network with attention mechanisms dynamically routes explanation requests between granular local insights and aggregated global summaries, adapting to stakeholder intent. Furthermore, the framework incorporates domain adaptation via an education-specific feature ontology, translating technical model outputs into pedagogically meaningful concepts through a bipartite graph neural network. The proposed method interfaces with existing education ROI systems through standardized APIs, replacing static regression coefficients with interactive SHAP-based visualizations for counterfactual analysis. Experimental validation demonstrates that the framework not only achieves competitive predictive performance but also provides actionable insights for policymakers and institutional administrators. The integration of dynamic glocal explainability with domain-aware feature interpretation represents a significant advancement over conventional black-box predictive models in education analytics