AL-GNN: Privacy-Preserving and Replay-Free Continual Graph Learning via Analytic Learning
Overview
Problem & Motivation
Class-incremental graph learning suffers from catastrophic forgetting. Existing replay-based methods store historical graph data, introducing storage overhead, privacy concerns, and repeated optimization cost.
Our Approach
We propose AL-GNN, a replay-free analytic framework. A graph encoder is trained in the base stage and then frozen. An analytic classifier is updated recursively in closed form using only current-session data and a regularized feature autocorrelation matrix, eliminating replay buffers.
Results
AL-GNN improves average performance on Cora from 73.50% (ERGNN) to 75.86%, reduces forgetting on Reddit from 16.95% (GEM) to 1.37%, and cuts training time on Reddit from 4580.42s (GEM) to 28.45s.
Authors and Affiliations
- Xuling Zhang — The Hong Kong University of Science and Technology (Guangzhou), China
- Jindong Li — The Hong Kong University of Science and Technology (Guangzhou), China
- Yifei Zhang — Nanyang Technological University (NTU), Singapore
- Mingqi Yang — South China University of Technology (SCUT), China
- Menglin Yang — The Hong Kong University of Science and Technology (Guangzhou), China
Research Area
- Continual Graph Learning
- Analytic Learning
- Replay-Free Learning
Key Contributions
- Introduced a replay-free analytic learning framework for class-incremental graph learning.
- Removed the need for replay buffers, reducing privacy risk and storage overhead.
- Designed recursive closed-form classifier updates using current-session data only.
- Achieved strong gains in accuracy, forgetting reduction, and training efficiency.
Publication Details
- Venue: IEEE Transactions on Pattern Analysis and Machine Intelligence Submission 2026 (TPAMI 2026)
- Presentation Type: Submission
Download
| Download Paper | Code Repository |