Bridging Efficiency and Adaptability: Continual Learning of MLPs on Class-Incremental Graphs

Compared to static graphs, class-incremental graphs place higher demands on inference latency to support timely predictions for newly emerged node classes, especially in latencysensitive applications. However, the high inference cost of Graph Neural Networks (GNNs) limits their scalability and motivates GNN-to-MLP distillation, which transfers knowledge from a GNN to a Multi-Layer Perceptron (MLP) to enable graph-free, low-latency inference. Yet, existing efforts focus on static graphs. When directly applied to class-incremental graphs, they inevitably suffer from the high computational cost of frequent GNN updates and MLP’s inability to retain knowledge of previously learned classes. To bridge efficiency and adaptability, we propose a novel framework featuring an asynchronous update paradigm between GNN and MLPs, allowing rapid adaptation to evolving data. MLPs employ a progressive expansion strategy for continual adaptation and an energy-based routing mechanism for test-time inference. During GNN updates, knowledge from MLPs trained in the current cycle is distilled back into GNN to preserve long-term knowledge. Experiments on real-world datasets demonstrate that our framework achieves superior performance on class-incremental graphs, effectively balancing adaptability to new data and inference efficiency.