Research map / Graph learning / Knowledge and temporal graphs
Analytical Construction on Geometric Architectures: Transitioning from Static to Temporal 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 a cross-geometric (Euclidean + hyperbolic) dynamic graph framework for temporal link prediction, combining per-ego-graph hyperbolicity-based geometry selection, a temporal state aggregator, and a link evolution loss.
Problem
Temporal link prediction on dynamic graphs where local structures have heterogeneous geometry (hierarchical vs. non-hierarchical) that changes over time, which single-geometry dynamic GNNs fail to capture.
Approach
Selects Euclidean or hyperbolic embeddings per k-hop ego-graph based on Gromov hyperbolicity at each timestamp, aggregates past hidden states via attention-based temporal state aggregation, and adds a link evolution loss that pulls new edges together and pushes lost edges apart.
Key results
- On temporal link prediction, the method reports AUC 94.13 on DBLP and 97.19 on HepPh, outperforming the best baseline HGWaveNet (89.96 and 92.37).
- On temporal new link prediction, AUC improves by 20.49% on Ia-Enron over the suboptimal baseline.
- Ablation on LFB shows removing LEL drops new-link AUC by 4.85%, and removing DGM drops link-prediction AUC by 2.42%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the full cross-geometric method instead of the strongest baseline | Temporal link prediction · Five datasets: DBLP, Ia-Enron, LFB, HepPh, UNVote | AUC and AP: + 4.17% AUC and + 0.40% AP on DBLP; + 5.25% AUC and + 3.65% AP on Ia-Enron; + 3.80% AUC and + 2.70% AP on LFB; + 4.82% A | Helped |
| Use the full cross-geometric method instead of the strongest baseline | Temporal new link prediction · Five datasets: DBLP, Ia-Enron, LFB, HepPh, UNVote | AUC and AP: + 9.69% AUC on DBLP; + 20.49% AUC and + 16.67% AP on Ia-Enron; + 4.66% AUC and + 3.54% AP on LFB; + 5.73% AUC and + 7.46 | Helped |
| Add dynamic geometric modeling (DGM) (ablation) | Full method on HepPh · Temporal link prediction | AUC: - 2.42% | Helped |
| Add temporal state aggregation (TSA) (ablation) | Full method on HepPh · Temporal link prediction | AUC: - 1.21% | Helped |
| Add link evolution loss (LEL) (ablation) | Full method on HepPh · Temporal new link prediction | AUC: - 4.85% | Helped |
| Vary k-hop ego-graph radius from 1 to 6; use k of 4 or greater (ablation) | DGM on HepPh and LFB · Temporal link and new link prediction | AUC and AP | Mixed |
| Vary embedding dimension from 2 to 16 (ablation) | Full method on HepPh and LFB · Temporal link prediction | AUC | Helped |
| Replace the hyperbolic component with spherical geometry (ablation) | Full method on HepPh and LFB · Link prediction and new link prediction | AUC and AP | Hurt |
Surprising findings
- Spherical geometry embedding did not improve over Euclidean or the hybrid hyperbolic-Euclidean approach on HepPh and LFB.
- The method's performance on UNVote is nearly identical to the baselines in the main link prediction table (AUC 61.68 vs. 61.40 for HTGN-B).
Limitations
- Fixed hyperparameter k=4 for ego-graph hyperbolicity, and large k collapses toward Euclidean-only behavior.
- Only Euclidean and hyperbolic (plus one spherical extension) geometries are considered.
Experimental setup
Datasets: DBLP HepPh Ia-Enron LFB UNVote
Benchmarks: DBLP HepPh Ia-Enron LFB UNVote
Metrics: AUC AP
Baselines: EdgeBank GAE GRUGCN EvolveGCN DySAT HGCN HAT HTGN-B HTGN-L HGWaveNet
Builds on: HGCN GRU GCN HTGN HGWaveNet
Closest papers by meaning
- TMetaNet: Topological Meta-Learning Framework for Dynamic Link Prediction (ICML 2025 · similarity 0.88)
- Future Link Prediction Without Memory or Aggregation (NeurIPS 2025 · similarity 0.87)
- DG-Mamba: Robust and Efficient Dynamic Graph Structure Learning with Selective State Space Models (AAAI 2025 · similarity 0.86)
- Geometry Awakening: Cross-Geometry Learning Exhibits Superiority over Individual Structures (NeurIPS 2024 · similarity 0.86)
- A Deep Probabilistic Framework for Continuous Time Dynamic Graph Generation (AAAI 2025 · similarity 0.86)
- When Graph Neural Networks Meet Dynamic Mode Decomposition (ICLR 2025 · similarity 0.85)
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.