United diverse subgraph for graph incremental learning

Graph incremental learning has emerged as a powerful graph deep learning framework, showcasing superior performance in addressing the evolving nature of graph data. However, catastrophic forgetting, which involves forgetting previously learned knowledge and overfitting to new data for sequential graph learning tasks, has become one of the most crucial challenges for graph incremental learning. Recent research has highlighted the significance of experience replay for effective anti-forgetting. This paper proposes a novel graph incremental learning model based on United Diverse Subgraph (UDS) for experience replay. This model firstly samples diverse nodes based on the uncertainty of each node for current task. Consequently, the unsampled nodes are pooled together into a supernode to extract global features from the unsampled nodes. Moreover, the structural relationships between nodes are established to form the final diverse subgraph for experience replay. This approach can capture both rich local and global information from current graph, which significantly reduces the space complexity of storing subgraphs. Extensive experiments conducted on various graph incremental learning datasets, consistently demonstrate the superior performance of our approach compared to existing graph incremental learning in the context of node classification.