Research map / Continual learning / Class-incremental
A Selective Learning Method for Temporal Graph Continual 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 temporal graph continual learning (TGCL) and LTF, a selective learning method that picks and learns representative old-class subsets via a domain-adaptation error bound, using MMD-based selection and a stop-gradient distribution regularizer.
Problem
Node classification on temporal graphs where new classes emerge over time: fine-tuning on new classes causes forgetting of old classes, while full retraining is costly and replay subsets go stale as old-class data evolves.
Approach
Derives an upper bound on old-class classification error when replacing full old-class data with a subset, converts it into a submodular subset-selection objective (classification error plus MMD on embeddings from the previous model) solved greedily with random partitioning, and aligns the learned subset's embedding distribution with a selected similar-distribution set via a stop-gradient MMD loss
Key results
- On TGAT, LTF reaches AP 0.0682 on Yelp, 0.0871 on Reddit and 0.1110 on Amazon, the best among compared methods, with AF 0.0195, 0.0474 and 0.0165.
- On DyGFormer, LTF achieves AP 0.0681 / 0.1134 / 0.1253 and AF 0.0096 / 0.0081 / 0.0383 on Yelp / Reddit / Amazon, outperforming OTGNet-style baselines where applicable.
- LTF is faster than OTGNet on TGAT (e.g., 39.16s vs 49.42s on Reddit; 72.94s vs 709.49s on Amazon) though slower than cheap baselines such as EWC and Finetune.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use LTF instead of the strongest listed baseline | TGAT · Yelp, Reddit, and Amazon | AP: 0.0648 -> 0.0682; 0.0868 -> 0.0871; 0.1031 -> 0.1110 | Helped |
| use LTF instead of the strongest listed baseline | DyGFormer · Yelp, Reddit, and Amazon | AP: 0.0560 -> 0.0681; 0.0917 -> 0.1134; 0.1089 -> 0.1253 | Helped |
| add distribution-based subset selection alongside error-based selection (ablation) | LTF · Yelp, Reddit, and Amazon | AP: 0.0438 -> 0.0654; 0.0579 -> 0.0866; 0.1063 -> 0.1004 | Helped |
| add error-based subset selection alongside distribution-based selection (ablation) | LTF · Yelp and Reddit | AP: 0.0565 -> 0.0654; 0.0640 -> 0.0866 | Helped |
| add error-based subset selection alongside distribution-based selection (ablation) | LTF · Amazon | AP: 0.0592 -> 0.1004 | Hurt |
| add distribution-alignment learning loss (ablation) | LTF with both selection terms · Yelp, Reddit, and Amazon | AP: 0.0654 -> 0.0682; 0.0866 -> 0.0871; 0.1004 -> 0.1110 | Helped |
| replace random partitioning with k-Means partitioning (ablation) | LTF · Yelp, Reddit, and Amazon | AP: 0.0682 -> 0.0598; 0.0871 -> 0.0754; 0.1110 -> 0.0929 | Hurt |
| replace random partitioning with Hierarchical partitioning (ablation) | LTF · Yelp, Reddit, and Amazon | AP: 0.0682 -> 0.0613; 0.0871 -> 0.0792; 0.1110 -> 0.0897 | Hurt |
| increase memory size m (ablation) | LTF · Yelp, Reddit, and Amazon | performance | Helped |
| increase the number of partitions (use smaller partitions) (ablation) | LTF with DyGFormer on Amazon · Amazon | AP: 0.1253; 0.1147; 0.1027 | Hurt |
Surprising findings
- Random partitioning outperforms k-means and hierarchical clustering for partitioning old-class data, since clustering distorts each partition's distribution.
Limitations
- Selection requires computing MMD over partitions, which is costlier than simple baselines like EWC and Finetune.
- Experiments are limited to three datasets and two backbones (TGAT, DyGFormer); OTGNet is not compatible with DyGFormer and is excluded there.
- Theoretical bound relies on approximations (previous model used in place of the future subset model, and empirical MMD/error estimates).
Experimental setup
Datasets: Yelp Reddit Amazon
Benchmarks: Yelp Reddit Amazon
Metrics: average precision (AP) average forgetting (AF) training time per epoch
Baselines: Joint Finetune LwF EWC iCaRL ER SSM OTGNet URCL
Builds on: OTGNet DyGLib EWC LwF
Closest papers by meaning
- A Topology-aware Graph Coarsening Framework for Continual Graph Learning (NeurIPS 2024 · similarity 0.89)
- Exploring Rationale Learning for Continual Graph Learning (AAAI 2025 · similarity 0.89)
- Dynamic Sub-graph Distillation for Robust Semi-supervised Continual Learning (AAAI 2024 · similarity 0.89)
- What Matters in Graph Class Incremental Learning? An Information Preservation Perspective (NeurIPS 2024 · similarity 0.88)
- Efficient Graph Continual Learning via Lightweight Graph Neural Tangent Kernels-based Dataset Distillation (ICML 2025 · similarity 0.87)
- GraphKeeper: Graph Domain-Incremental Learning via Knowledge Disentanglement and Preservation (NeurIPS 2025 · 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.