Robust Bayesian Dynamic Graph Embedding via Contrastive Learning

doi:10.1109/tbdata.2026.3695337

Robust Bayesian Dynamic Graph Embedding via Contrastive Learning

2026-05-20

0

PRE

AI

DOI：10.1109/tbdata.2026.3695337

Original

Original request for help

Save

Abstract

En 中文

Dynamic network representation learning seeks to create low-dimensional node embeddings that capture both the structural and temporal evolution patterns of networks. While Bayesian deep learning-based models for dynamic network embedding have shown promising performance, they often face challenges with the noisiness of real-world data, resulting in less accurate and robust results. Contrastive learning in graph representation learning has made progress in improving the robustness and accuracy of these models by generating node representations from different perspectives. To tackle the aforementioned challenges, this study introduces a novel approach that combines Bayesian deep learning with contrastive learning. Our method unifies a history-conditioned variational graph autoencoder with topology-preserving local–global contrast and node-level cross-view InfoNCE, aligning generative and discriminative embeddings while retaining a conditional prior for forecasting. This design addresses limitations of existing dynamic graph models by bringing uncertainty-aware generation to contrastive methods and adding discriminative regularization and noise robustness to Bayesian models, enabling next-snapshot prediction and improved resilience on noisy graphs. Extensive testing on diverse real-world networks demonstrates that our model significantly improves dynamic graph representation learning, offering a more effective solution for analyzing network dynamics.

Keywords:

Dynamic networks

network embedding

contrastive learning

Bayesian deep learning

Robust Bayesian Dynamic Graph Embedding via Contrastive Learning

Abstract

Journal

Organization

Cited Papers