Article

Adaptive Learning Through Multi-Source Educational Data Fusion Using Graph Neural Networks

---

Abstract

Adaptive learning systems increasingly rely on large-scale educational data, yet many existing approaches remain constrained by fragmented data sources and flat representations that inadequately capture the relational nature of learning processes. This study proposes a graph-based adaptive learning framework that fuses multi-source educational data including: Learning Management System (LMS) interaction logs, assessment records, learner profiles, and curriculum metadata into a unified heterogeneous educational graph. Graph Neural Networks (GNNs) are employed to learn contextualized representations of learners, content, concepts, and assessments through relational message passing, enabling coherent learner state inference and personalized adaptation. Empirical evaluation was conducted on a large-scale dataset comprising over 2.4 million interaction events, 315 thousand assessment records, 1,240 learner profiles, and 860 learning contents mapped to 210 curriculum concepts. Results show that the proposed GNN model achieves a test accuracy of 0.82 and an F1-score of 0.80, outperforming feature-based and sequential baseline models by margins of 5–8 percentage points. Training dynamics demonstrate stable convergence with limited overfitting, indicating robust generalization across learners with heterogeneous interaction densities. Comparative analysis further reveals superior cold-start robustness, as graph-based neighborhood inference compensates for sparse individual histories. Beyond predictive performance, the adaptive learning mechanism yields substantial pedagogical benefits. Learners exposed to graph-driven personalization exhibit higher average concept mastery gains (0.58 vs. 0.40), reduced time to reach target mastery (8.7 vs. 12.4 learning sessions), lower content repetition rates (18% vs. 32%), and decreased early drop-off rates (12% vs. 21%) compared to non-adaptive settings. These findings indicate that prerequisite-aware sequencing and cross-source inference significantly enhance learning efficiency and engagement. Overall, the study demonstrates that multi-source educational data fusion using GNNs provides a scalable, stable, and pedagogically grounded foundation for next-generation adaptive learning systems.

Keywords: Adaptive Learning, Educational Data Fusion, Graph Neural Networks, Learning Analytics

How to Cite:

Thepnuan, N., (2025) “Adaptive Learning Through Multi-Source Educational Data Fusion Using Graph Neural Networks”, Adaptive Learning 1(3), 253-273. doi: https://doi.org/10.63913/al.v1i3.113