Work

Experience replay with cohesive-subgraph awareness for continual graph learning in IoT

Year: 2025

Type: article

Abstract: In Internet-of-Things (IoT) scenarios, Continual Graph Learning (CGL) has become a key technique for capturing the complex relational structures underlying diverse applications including sensor networks, road systems and biomedical networks. However, the structural changes in these evolving graphs introduce instability, making catastrophic forgetting a primary challenge for CGL. Experience replay is currently a promising method, as it strikes a balance between new and old knowledge. It also provides CGL models with a human-like memory capability. However, prior work rarely leverages the graph's intrinsic properties to proactively identify the critical patterns hiding in the evolving graphs. To this end, we propose a unified framework that integrates cohesion-subgraph awareness into existing CGL mechanisms. We propose a novel cohesive structure-aware experience replay framework that leverages intrinsic graph properties, such as k -core and k -truss metrics, to guide the selection of representative historical nodes for replay. Unlike conventional replay strategies that rely on random sampling or task-driven node selection, our approach systematically identifies structurally significant nodes that encapsulate the evolving patterns of streaming graphs. By integrating these cohesive subgraph properties into the experience replay process, our method effectively preserves critical historical knowledge while adapting to new graph structures with minimal computational overhead. The experimental results demonstrate that our method consistently outperforms existing replay strategies in mitigating catastrophic forgetting and maintaining classification performance. On the PubMed dataset, our k -core-based replay strategy improves the F 1 score by 3.7% compared to random sampling, while reducing training time by up to 85% compared to full retraining. Similarly, on the Cora dataset, our approach achieves a 98.3% F 1 score, surpassing baseline methods by 4.5%.

Source: Digital Communications and Networks

DOI: 10.1016/j.dcan.2025.08.010

Authors: Zhenzhen Xie, Qi Luo, Yan Huang, Yongqi Yin, Jiaqi Zhang, Junjie Pang

Institutions: Shandong University, Kennesaw State University, Changchun University of Technology, Qingdao University

Language: English

Topic: Advanced Graph Neural Networks

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: gold

Estimated APC paid: $800

Funders: National Natural Science Foundation of China, Natural Science Foundation of Shandong Province, Young Scientists Fund