Condensation-Concatenation Framework for Dynamic Graph Continual Learning

1 Condensation-Concatenation Framework for Dynamic Graph Continual Learning 1st Tingxu Yan College of Computer and Information Science Southwest University ChongQing, China 929549561@qq.com 2nd Ye Yuan* College of Computer and Information Science Southwest University ChongQing, China *yuanyekl@swu.edu.cn Abstract Dynamic graphs are prevalent in real-world scenarios, where continuous structural changes induce catastrophic forgetting in graph neural networks (GNNs). While continual learning has been extended to dynamic graphs, existing methods overlook the effects of topological changes on existing nodes. To address it, we propose a novel framework for continual learning on dynamic graphs, named Condensation-Concatenation-based Continual Learning (CCC). Specifically, CCC first condenses historical graph snapshots into compact semantic representations while aiming to preserve the original label distribution and topological properties. Then it concatenates these historical embeddings with current graph representations selectively. Moreover, we refine the forgetting measure (FM) to better adapt to dynamic graph scenarios by quantifying the predictive performance degradation of existing nodes caused by structural updates. CCC demonstrates superior performance over state-of-the-art baselines across four real-world datasets in extensive experiments. Index Terms Continual Learning, Dynamic Graphs, Catastrophic Forgetting I. INTRODUCTION Graphs serve as fundamental structures for modeling relational data in domains like social networks. Graph Neural Networks (GNNs) have become the standard framework for graph representation learning with significant success. However, real-world graphs are dynamic, continually evolving through the addition and deletion of nodes and edges. This temporal dimension challenges conventional GNNs designed for static graphs, leading to catastrophic forgetting where learning new patterns overwrites previously acquired knowledge. Although continual learning research has expanded to dynamic graphs, evaluation frameworks lack graph-specific adaptations. Metrics like forgetting rate, borrowed from static domains, fail to capture structural cascading effects. Existing methods employ broad preservation strategies while overlooking that topological changes influence representations of numerous existing nodes. To tackle these limitations, we propose the Condensation-Concatenation Framework for Dynamic Graph Continual Learning (CCC) framework. CCC compresses historical graph snapshots into compact semantic representations and integrates them with current embeddings through concatenation. This approach aims to capture structural change propagation while maintaining representation stability. Our contributions are: • Identification of limitations in existing continual learning evaluation metrics for dynamic graphs. • Proposal of the CCC framework combining graph condensation with feature concatenation. • Comprehensive experimental validation demonstrating superior performance in balancing knowledge retention with new information integration. II. RELATED WORK Graph Neural Networks. Graph Neural Networks (GNNs) [1] aim to apply deep learning to graph-structured data. The core of their approach is to aggregate information from neighboring nodes for learning node representations. Graph Convolutional Network (GCN) [2] established a spectral graph convolution framework through first-order neighborhood approximation. GraphSAGE [3] proposed a framework based on sampling and aggregation instead of using all neighboring nodes.In addition, numerous other GNN studies [4]–[80] have also made significant contributions and proven valuable in various graph learning tasks. arXiv:2512.11317v1 [cs.LG] 12 Dec 2025 2 Fig. 1: An overview of CCC. The historical graph sequence is first condensed to capture historical information. Subsequently, by detecting the k-hop structural change regions triggered by node and edge additions/deletions, the embeddings extracted from the condensed historical graph are selectively concatenated with the current embeddings of the affected nodes. Continual Learning. Continual learning allows models to learn from sequentially arriving data while avoiding catastrophic forgetting of previously acquired knowledge. Existing approaches can be categorized into three groups: Regularization-based methods protect learned knowledge by incorporating constraints into the loss function. TWP [81] applies regularization based on parameter sensitivity to topological structures. DyGRAIN identifies affected nodes from a receptive field perspective for selective updates. Memory replay-based methods consolidate knowledge by storing and replaying historical data. ER-GNN [82] adopts multiple strategies to select replay nodes. SSM uses sparsified subgraphs as memory units to retain topological information. DSLR [83] selects replay nodes based on coverage and trains a link prediction module. PUMA condenses original grap

This content is AI-processed based on open access ArXiv data.