Research map / Continual learning / Online and replay-based
A Topology-aware Graph Coarsening Framework for Continual Graph Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TACO, a continual graph learning framework that stores past tasks as a reduced graph built by RePro, an embedding-proximity graph coarsening method with node fidelity preservation, and reports gains over replay and regularization baselines on Kindle, DBLP and ACM.
Problem
Catastrophic forgetting in continual node classification on streaming, time-stamped graphs, where replay methods fail to preserve topology and inter-task correlations under inductive, generalized class-incremental settings.
Approach
Combine each new subgraph with a coarsened reduced graph of past tasks (aligning shared nodes), then coarsen again by greedily merging connected nodes with the highest GNN-embedding cosine similarity, with a replay buffer of representative nodes protected from merging and degree-based node importance weighting in the coarsened features.
Key results
- On GCN, the method reaches 82.97% F1-AP on Kindle, 84.60% on DBLP and 70.96% on ACM, with lower forgetting than the best baselines (e.g., SSM 78.99% on Kindle).
- Replacing its coarsener with RePro gives the lowest coarsening time (e.g., 2.3 s on Kindle vs 8.9 s for Alge. JC and 65.6 s for Aff. GS) with comparable accuracy.
- Theory shows that majority-vote coarsening causes minority classes to shrink in expectation, and node fidelity preservation mitigates this.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the topology-aware coarsening framework instead of SSM | GCN · Kindle | F1-AP (%): 78.99 ± 3.13 -> 82.97 ± 2.05 | Helped |
| Use the topology-aware coarsening framework instead of SSM | GCN · Kindle | F1-AF (%): 8.19 ± 3.63 -> 4.91 ± 1.90 | Helped |
| Use the topology-aware coarsening framework instead of SSM | GCN · DBLP | F1-AP (%): 82.71 ± 1.76 -> 84.60 ± 2.01 | Helped |
| Use the topology-aware coarsening framework instead of SSM | GCN · DBLP | F1-AF (%): 4.20 ± 1.26 -> 2.51 ± 1.03 | Helped |
| Use the topology-aware coarsening framework instead of SSM | GCN · ACM | F1-AP (%): 68.77 ± 2.93 -> 70.96 ± 2.68 | Helped |
| Use the topology-aware coarsening framework instead of SSM | GCN · ACM | F1-AF (%): 9.50 ± 2.47 -> 8.02 ± 2.33 | Helped |
Surprising findings
- Experience replay variations ERGNN-rs, rb, and mf do not outperform the simpler SSM and the proposed method, showing that replaying isolated nodes without structure is insufficient.
- Simple GEM and EWC variants that ignore graph structure remain competitive with some graph-aware baselines on DBLP.
Limitations
- Only considers streaming graphs with a single relation type; nodes and edges are only added, not deleted or modified.
- Evaluated only on node classification; link prediction and multi-relation graphs are left to future work.
Experimental setup
Datasets: Kindle DBLP ACM
Benchmarks: Kindle DBLP ACM
Metrics: F1-AP F1-AF BACC-AP BACC-AF runtime (seconds)
Baselines: EWC GEM TWP OTG ERGNN SSM DyGrain IncreGNN SSRM FGC
Builds on: ERGNN FGC experience replay
Closest papers by meaning
- A Selective Learning Method for Temporal Graph Continual Learning (ICML 2025 · similarity 0.89)
- Efficient Graph Continual Learning via Lightweight Graph Neural Tangent Kernels-based Dataset Distillation (ICML 2025 · similarity 0.88)
- Exploring Rationale Learning for Continual Graph Learning (AAAI 2025 · similarity 0.88)
- Towards Continuous Reuse of Graph Models via Holistic Memory Diversification (ICLR 2025 · similarity 0.88)
- GraphKeeper: Graph Domain-Incremental Learning via Knowledge Disentanglement and Preservation (NeurIPS 2025 · similarity 0.87)
- What Matters in Graph Class Incremental Learning? An Information Preservation Perspective (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.