On Diffusion-based Multiplex Dynamic Attributed Network Generator

Jui-Yi Tsai, Ya-Wen Teng, Lydia Y. Chen, De-Nian Yang

Keywords: Generative diffusion model, multiplex temporal graphs

Abstract: Multiplex dynamic attributed networks are essential for modeling complex systems, such as social platforms and telecommunication networks, where each layer represents distinct interaction types and attribute dynamics. However, existing generative models fall short in capturing their structural-semantic coupling, temporal evolution, and inter-layer dependencies, failing to reproduce network-level emergent behaviors like explosive synchronization and hysteresis. We introduce MulDyDiff, a diffusion-based generative framework that incorporates attribute-aware dynamic transition-based denoising, cross-layer correlation-aware denoising, and behavior-aware guidance. These components are unified through a novel Behavioral-guided Attributed Cross-layer Temporal (BACT) loss. Evaluations of three real-world datasets demonstrate that MulDyDiff consistently outperforms state-of-the-art dynamic graph generators, achieving 6%-9% improvement in terms of temporal metrics, offering a comprehensive solution for realistic multiplex dynamic attributed network synthesis.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 4988