Research map / Graph learning / Knowledge and temporal graphs
A Deep Probabilistic Framework for Continuous Time Dynamic Graph Generation
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 DG-Gen, an encoder-decoder framework that generates continuous-time dynamic graphs by autoregressively sampling temporal interactions from a factorized joint probability, using TGN as encoder and a deep probabilistic decoder.
Problem
Generating synthetic continuous-time dynamic graphs (CTDGs) that are statistically faithful, original, inductive, scalable to long time horizons, and able to include edge features, which prior static-snapshot-based generators handle poorly.
Approach
Directly models temporal interactions as a joint probability factorized into source, destination given source, and time/edge features given both; a TGN encoder produces node embeddings and a deep probabilistic decoder with categorical, exponential and GMM heads samples new interactions autoregressively, avoiding static adjacency matrices.
Key results
- DG-Gen outperforms TIGGER-I on most topology metrics across datasets, e.g., Wikipedia closeness centrality error 0.656 vs 0.992 and LastFM 0.681 vs 0.849.
- DG-Gen achieves the best inductive link prediction AP on all datasets except Reddit, where it matches CAWN.
- Mean Jensen-Shannon distance between real and synthetic single edge features ranges from 0.107 (Bikeshare) to 0.208 (Wikipedia).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DG-Gen instead of TIGGER-I for graph generation | Inductive CTDG generation · Wikipedia, Bikeshare, MOOC, Reddit, and LastFM | Median absolute error of graph statistics | Helped |
| Add graph attention layers (memory/attention module) (ablation) | DG-Gen · All five datasets | Closeness centrality (CC) median absolute error: 0.723 -> 0.656; 0.305 -> 0.298; 0.486 -> 0.430; 0.133 -> 0.066; 0.760 -> 0.681 | Helped |
| Add stabilizing random noise during training (ablation) | DG-Gen Time+MSG module · Wikipedia, Bikeshare, MOOC, and Reddit | Closeness centrality (CC) median absolute error: 0.701 -> 0.656; 0.297 -> 0.298; 0.449 -> 0.430; 0.089 -> 0.066 | Mixed |
| Use DG-Gen rather than TGN alone as the link prediction model | Inductive link prediction · All datasets except Reddit | Average Precision (AP) | Helped |
Surprising findings
- Both DG-Gen and TIGGER-I show no edge overlap with source graphs, so originality metric was excluded.
- DG-Gen outperforms TGN even though TGN is used as its encoder, suggesting the decoder improves representation learning.
Limitations
- Only TIGGER-I is available as an inductive generative baseline, so comparisons are limited.
- Edge feature generation cannot be directly compared to existing baselines.
- Numerical edge features require noise injection for training stability.
Experimental setup
Datasets: Reddit Wikipedia MOOC LastFM Capital Bikeshare
Benchmarks: Reddit Wikipedia MOOC LastFM Capital Bikeshare
Metrics: Jensen-Shannon distance median absolute error of closeness centrality, mean degree, average precision AUROC
Baselines: TIGGER-I Jodie DyRep TGAT TGN CAWN
Builds on: TGN TIGGER-I GRU GraphSAGE
Closest papers by meaning
- Quality Measures for Dynamic Graph Generative Models (ICLR 2025 · similarity 0.90)
- DyG-Mamba: Continuous State Space Modeling on Dynamic Graphs (NeurIPS 2025 · similarity 0.89)
- State Space Models on Temporal Graphs: A First-Principles Study (NeurIPS 2024 · similarity 0.88)
- Node-Time Conditional Prompt Learning in Dynamic Graphs (ICLR 2025 · similarity 0.88)
- Continuous-Time Graph Representation with Sequential Survival Process (AAAI 2024 · similarity 0.88)
- Discrete-state Continuous-time Diffusion for Graph Generation (NeurIPS 2024 · similarity 0.88)
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.