The sudden growth in digital learning environments has aided the rise of student datasets that provide a great opportunity to predict and explain complex student behaviour and their performance. However, the goal of predicting academic performance remains very difficult because educational interactions are inherently very complex, multimodal, nonlinear, and imbalanced. To overcome limitations of current state-of-the-art (SOTA) models, this study proposes a Graph-KAN, a paradigm that integrates Kolmogorov- Arnold Networks (KANs) within a heterogeneous graph framework. Instead of relying on fixed activation functions, like in traditional Multi-Layer Perceptrons (MLPs), Graph-KAN takes advantage of the Kolmogorov-Arnold representation theorem to learn polynomial splines along the graph edges, allowing the model to dynamically approximate complex nonlinear education functions. We trained and validated Graph-KAN against standard baselines (GraphSAGE, GATv2, and Transformer-Conv) on the Open University Learning Analytics Dataset (OULAD). Our experimental results show that the hybrid Graph-KAN achieves a Macro F1-score of 0.536, outperforming Transformer-based architectures by about 4% and requires comparatively lower computational resources. The hybrid Graph-KAN also overcomes the limitation of class imbalance, which enhances the identification of “Distinction” students, achieving an AUC of 0.78 compared to 0.73 for Transformers. Along with high accuracy, the model offers explainability as the learned activation functions show a “diminishing returns” phenomenon in student engagement. Thus, providing educators with mathematically proven insights for personalised intervention. While basic binary prediction tasks on OULAD often yield high accuracy and F1-score, this study is based on a more complex 4-class task, which remains a challenge. This study proposes Graph-KAN as a new paradigm for such a complex setting. This work establishes KAN-based message passing as a sparingly, interpretable alternative to attention mechanisms in Educational Data Mining.