Research map / Probabilistic and scientific modeling / Molecules and proteins
Community-Aware Variational Autoencoder for Continuous Dynamic Networks
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.
Extends variational autoencoder community detection to continuous-time dynamic networks by using a Hawkes-process inference component, pseudo-label compact reconstruction, and a Gaussian-mixture variant (CT-CAVAE) for end-to-end clustering.
Problem
Community detection in continuous dynamic networks is hard for VAE methods because their GCN-based inference and adjacency-matrix reconstruction cannot capture timeliness, causality, and non-adjacent interactions.
Approach
Treats interactions as event streams modeled with a Hawkes process for inference, reconstructs edges via pseudo-labels from community membership with a compact constraint, and uses a Student's t prior plus a Gaussian-mixture community constraint (CT-CAVAE) for end-to-end detection.
Key results
- CT-VAE reaches 48.40% ACC on DBLP and 60.51% ACC for CT-CAVAE on Brain (Table 1 entries; CT-VAE gets 46.13% on Brain and 38.21% NMI on Patent for CT-CAVAE).
- On Patent, CT-VAE improves average ACC by 4.82%, NMI by 14.68%, ARI by 14.69% and F1 by 8.48% over baselines (as reported in the paper).
- CT-VAE attains 100% on School across metrics, matching the top baseline level.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace compact pseudo-label reconstruction in Eq. (5) with relaxed homogeneity reconstruction in Eq. (4) (ablation) | CT-VAE · Patent | NMI: 34.21 -> 36.30 | Helped |
| Replace compact pseudo-label reconstruction in Eq. (5) with relaxed homogeneity reconstruction in Eq. (4) (ablation) | CT-CAVAE · Patent | ACC: 54.72 -> 60.51 | Helped |
| Use K-means for community detection in place of end-to-end community detection (ablation) | CT-VAE · Patent | ACC: 46.13 -> 53.89 | Helped |
| Use K-means for community detection in place of end-to-end community detection (ablation) | CT-VAE · DBLP | ARI: 22.69 -> 22.88 | Helped |
| Use K-means for community detection in place of end-to-end community detection (ablation) | CT-VAE · arXivCS | NMI: 45.75 -> 45.56 | Hurt |
| Use K-means for community detection in place of end-to-end community detection (ablation) | CT-VAE · Brain | NMI: 50.42 -> 50.36 | No clear effect |
| Increase negative-sampling count (ablation) | CT-VAE · six datasets | community detection metrics | Mixed |
| Use CT-VAE instead of the strongest baseline | continuous dynamic community detection · Patent | NMI: +11.26% | Helped |
| Use CT-VAE instead of the strongest baseline | continuous dynamic community detection · School | ACC | Helped |
| Use CT-CAVAE instead of the strongest baseline | continuous dynamic community detection · Patent | ACC: +10.15% | Helped |
Surprising findings
- CT-CAVAE underperforms CT-VAE on academic networks (DBLP, arXivAI, arXivCS), which the authors attribute to asynchronous publication-driven interactions and multi-interest hub nodes.
Limitations
- CT-CAVAE gives weaker results on academic networks with asynchronous, bursty interactions
- Relies on node2vec initialization and K-means for cluster centers in CT-VAE
- Datasets are only from prior work; no new dataset released
Experimental setup
Datasets: DBLP arXivCS arXivAI Brain Patent School
Benchmarks: DBLP arXivCS arXivAI Brain Patent School
Metrics: ACC NMI ARI F1-macro
Baselines: node2vec HTNE JODIE TGAT TGN MTNE M2DNE TREND TGC
Builds on: VGAE HTNE TGC node2vec
Closest papers by meaning
- Learning Persistent Community Structures in Dynamic Networks via Topological Data Analysis (AAAI 2024 · similarity 0.87)
- Graph Representation Learning via Ladder Gamma Variational Autoencoders (AAAI 2020 · similarity 0.87)
- Semi-Implicit Graph Variational Auto-Encoders (NeurIPS 2019 · similarity 0.86)
- vGraph: A Generative Model for Joint Community Detection and Node Representation Learning (NeurIPS 2019 · similarity 0.86)
- Community detection in sparse time-evolving graphs with a dynamical Bethe-Hessian (NeurIPS 2020 · similarity 0.86)
- TreeVI: Reparameterizable Tree-structured Variational Inference for Instance-level Correlation Capturing (NeurIPS 2024 · similarity 0.84)
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.