Multi-modal graph neural networks for cross-domain educational recommendation: integrating behavioral analytics and institutional context for personalized learning

Abstract

Abstract Educational recommendation systems have traditionally relied on single-dataset approaches, limiting their ability to capture the complex, multi-faceted nature of student learning. This paper introduces a novel multi-modal graph neural network framework that integrates heterogeneous educational data sources to deliver superior personalized learning recommendations. Our approach combines behavioral learning analytics from EdNet with institutional context from OULAD, creating a large-scale cross-dataset educational framework. The proposed architecture employs Graph Convolutional Networks for structural modeling, Graph Attention Networks for dynamic weighting, and hierarchical temporal components to capture learning dynamics. Novel cross-modal attention mechanisms enable knowledge transfer between behavioral patterns and contextual factors, while cognitive load-aware optimization ensures educationally appropriate recommendations. Comprehensive experimental evaluation demonstrates substantial improvements in recommendation accuracy and educational effectiveness. Individual-level assessment reveals high accuracy in predicting students’ actual learning choices, with superior success rates for recommended learning activities. Cross-dataset transfer learning achieves excellent performance, showing significant improvements over traditional domain adaptation approaches. Beyond performance metrics, our framework delivers tangible educational benefits including substantial reduction in learning time while maintaining high engagement levels through adaptive optimization. The system demonstrates its capability in learning gap identification and targeted remediation, with strong correlations to educational psychology indicators validating pedagogical authenticity.