Community-Aware Variational Autoencoder for Continuous Dynamic Networks

Junwei Cheng, Chaobo He, Pengxing Feng, Weixiong Liu, Kunlin Han, Yong TangAAAI 2025 · DOI

Real-world networks, such as social media interactions, neural activity, and citation graphs, are inherently continuous and dynamic. Traditional community detection methods often simplify these networks into static snapshots or discrete-time sequences, which "squashes" temporal information and loses the causal relationships between events. This paper introduces CT-VAE and its scalable variant CT-CAVAE, the first Variational Autoencoder (VAE) frameworks designed specifically for continuous-time dynamic networks.

Addressing the Dynamic Network Challenge

The authors identify two primary limitations in existing VAE-based community detection:

  1. Structural Loss from Discretization: Relying on adjacency matrices for dynamic graphs ignores the precise timing and chronological order of interactions.
  2. The Non-Adjacent Challenge: In evolving networks, nodes within the same community may not have interacted yet. Standard VAEs, which focus on reconstructing existing edges, struggle to group these potentially related but currently non-adjacent nodes.

Methodology: Hawkes Processes and Robust Priors

To overcome these gaps, the proposed framework replaces traditional Graph Convolutional Networks (GCNs) with the Hawkes process. The Hawkes process is a self-exciting point process where past events increase the probability of future ones, making it ideal for capturing the temporal causality of node interactions.

Overview of the CT-VAE framework, illustrating the transition from continuous-time event streams to Hawkes-based node embeddings and subsequent inference and generation processes.

Key components of the architecture include:

  • Hawkes Process Inference: This captures "bursty" activity patterns as signals for community membership.
  • Compact Constraint & Pseudo-Labeling: During the generation phase, the model uses pseudo-labels to enforce a "compact constraint." This ensures that nodes predicted to be in the same community have similar latent representations, even if they lack a direct edge.
  • Student’s t-Distribution Prior: CT-CAVAE utilizes a Student’s t-distribution instead of a standard Gaussian prior. This provides greater robustness against "outlier" nodes—those with sparse or irregular interaction patterns.
  • End-to-End Clustering: By incorporating a Gaussian Mixture Model (GMM), CT-CAVAE can learn node representations and community assignments simultaneously.

Experimental Evaluation

The models were tested on six diverse real-world datasets: DBLP, arXivCS, arXivAI (academic), Brain (biological), Patent (technical), and School (social). These datasets represent varying levels of "homogeneity"—the likelihood that connected nodes share a community. Performance was measured using four standard metrics: Accuracy (ACC), Normalized Mutual Information (NMI), Adjusted Rand Index (ARI), and F1-score.

Key Results and Insights

The experimental results demonstrate that CT-VAE and CT-CAVAE consistently outperform state-of-the-art baselines like node2vec, TGAT, and TGN.

Comparison of CT-VAE and CT-CAVAE against various baseline models across six datasets. The proposed models show superior performance in most metrics, particularly on sparse or high-resolution temporal data.

Significant findings include:

  • Superiority in Sparse Networks: On the Patent dataset, which is characterized by high sparsity, CT-CAVAE improved Accuracy by 10.15% over the best baseline. This highlights the effectiveness of the Student’s t-distribution and compact constraint in handling isolated nodes.
  • High Precision in Social Structures: The model achieved near-perfect (100%) scores on the School dataset, proving its ability to accurately map social structures when high-resolution temporal data is available.
  • Domain Sensitivity: While highly effective for event-driven networks (like Patents or student interactions), the models faced more challenges with academic networks (DBLP, arXiv). The authors attribute this to the asynchronous nature of publication cycles, which can obscure clear causal links.

Conclusion

CT-VAE and CT-CAVAE represent a significant step forward in community detection for continuous dynamic networks. By leveraging Hawkes processes and robust statistical priors, these models capture the vital "when" and "why" of node interactions. Future research will focus on better integrating node attributes and handling the asynchronous "noise" found in specific domains like academic publishing.

Summary updated

BibTeX
@inproceedings{cheng2025community,
  author = {Junwei Cheng and Chaobo He and Pengxing Feng and Weixiong Liu and Kunlin Han and Yong Tang},
  title = {{Community-Aware Variational Autoencoder for Continuous Dynamic Networks}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i15.33757},
}
Figure