Research map / Graph learning / Knowledge and temporal graphs
TMetaNet: Topological Meta-Learning Framework for Dynamic Link Prediction
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 Dowker Zigzag Persistence (DZP), a landmark-based zigzag persistence representation for dynamic graphs, and TMetaNet, a meta-learning model that uses DZP persistence images to learn per-step learning rates for dynamic GNN weight updates, evaluated on dynamic link prediction.
Problem
Fixed weight-update parameters in meta-learning-based dynamic GNNs ignore high-order topological evolution of discrete-time dynamic graphs, and zigzag persistence is computationally expensive for dynamic link prediction.
Approach
Replace fixed meta-learning step sizes with a learned scalar learning rate predicted by a CNN+FC adaptor from the difference of zigzag persistence images computed on Dowker complexes built from epsilon-net landmarks of adjacent snapshots.
Key results
- In the ROLAND live-update setting, TMetaNet reaches 86.84% ACC on Bitcoin-Alpha vs 83.18% for the best baseline ROLAND, and 17.68 MRR vs 15.23.
- In the WinGNN setting, TMetaNet averages about 3.27% ACC and 24.08% MRR improvement over baselines, with 98.26% ACC on Reddit-Body and 99.63% on Reddit-Title.
- Ablation shows replacing DZP images with random images lowers Alpha ACC in the WinGNN setting from 89.92% to 82.14%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DZP persistence images instead of random images (ablation) | TMetaNet · Alpha (ROLAND live update) | ACC: 83.07 $\pm$ 2.44 -> *86.84 $\pm$ 1.02 | Helped |
| Use DZP persistence images instead of random images (ablation) | TMetaNet · Alpha (ROLAND live update) | MRR: 15.85 $\pm$ 0.63 -> 17.68 $\pm$ 0.55 | Helped |
| Use DZP persistence images instead of statistically matched images (ablation) | TMetaNet · Alpha (ROLAND live update) | ACC: 84.24 $\pm$ 1.42 -> *86.84 $\pm$ 1.02 | Helped |
| Use DZP persistence images instead of statistically matched images (ablation) | TMetaNet · Alpha (ROLAND live update) | MRR: 15.55 $\pm$ 0.72 -> 17.68 $\pm$ 0.55 | Helped |
| Use DZP persistence images instead of random images (ablation) | TMetaNet · UCI (ROLAND live update) | ACC: 80.04 $\pm$ 1.26 -> *80.88 $\pm$ 0.08 | Helped |
| Use DZP persistence images instead of random images (ablation) | TMetaNet · UCI (ROLAND live update) | MRR: 10.32 $\pm$ 0.36 -> *10.99 $\pm$ 0.92 | Helped |
| Use DZP persistence images instead of statistically matched images (ablation) | TMetaNet · UCI (ROLAND live update) | ACC: 79.96 $\pm$ 0.94 -> *80.88 $\pm$ 0.08 | Helped |
| Use DZP persistence images instead of statistically matched images (ablation) | TMetaNet · UCI (ROLAND live update) | MRR: 10.88 $\pm$ 0.87 -> *10.99 $\pm$ 0.92 | Helped |
| Use DZP persistence images instead of random images (ablation) | TMetaNet · Alpha (WinGNN setting) | ACC: 82.14 $\pm$ 1.76 -> *89.92 $\pm$ 1.84 | Helped |
| Use DZP persistence images instead of random images (ablation) | TMetaNet · Alpha (WinGNN setting) | MRR: 33.98 $\pm$ 3.12 -> *38.93 $\pm$ 3.06 | Helped |
| Use DZP persistence images instead of statistically matched images (ablation) | TMetaNet · Alpha (WinGNN setting) | ACC: 88.63 $\pm$ 1.23 -> *89.92 $\pm$ 1.84 | Helped |
| Use DZP persistence images instead of statistically matched images (ablation) | TMetaNet · Alpha (WinGNN setting) | MRR: 38.02 $\pm$ 2.92 -> *38.93 $\pm$ 3.06 | Helped |
Surprising findings
- In the ROLAND setting on Reddit-Body, ROLAND (91.63% ACC) outperforms TMetaNet (89.59%), so TMetaNet is not best there.
- Distribution-matched random replacement (Dist) performs close to Topo on some settings, suggesting part of the gain comes from distributional properties of the input.
Limitations
- Gains on some datasets in the ROLAND setting are small (e.g., under 1.5% ACC on OTC and Title) and TMetaNet loses to ROLAND on Reddit-Body.
- Reported improvement percentages are not consistently derived from the tables, and the stability theorem's proof sketch is informal, with a dimension-dependent constant left loosely defined.
Experimental setup
Datasets: Bitcoin-OTC Bitcoin-Alpha Reddit-Body Reddit-Title UCI-Message ETH-Yocoin
Benchmarks: Bitcoin-OTC Bitcoin-Alpha Reddit-Body Reddit-Title UCI-Message ETH-Yocoin
Metrics: ACC MRR AUC Recall@1 Recall@3 Recall@10
Baselines: GCN-L GCN-G EvolveGCN WinGNN ROLAND DeGNN
Builds on: ROLAND WinGNN zigzag persistence witness complex model-agnostic meta-learning
Closest papers by meaning
- Link Prediction with Persistent Homology: An Interactive View (ICML 2021 · similarity 0.90)
- DG-Mamba: Robust and Efficient Dynamic Graph Structure Learning with Selective State Space Models (AAAI 2025 · similarity 0.88)
- A Deep Probabilistic Framework for Continuous Time Dynamic Graph Generation (AAAI 2025 · similarity 0.88)
- Analytical Construction on Geometric Architectures: Transitioning from Static to Temporal Link Prediction (ICML 2025 · similarity 0.88)
- DyG-Mamba: Continuous State Space Modeling on Dynamic Graphs (NeurIPS 2025 · similarity 0.87)
- TopER: Topological Embeddings in Graph Representation Learning (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.