Sequence Diffusion Model for Temporal Link Prediction in Continuous-Time Dynamic Graph

TL;DR

SDG tackles temporal link prediction in continuous-time dynamic graphs by reframing the task as sequence-level diffusion: noise is injected into the entire historical interaction sequence and a cross-attention denoising network reconstructs the destination sequence conditioned on history. This approach models uncertainty and long-range temporal dependencies, producing richer destination distributions than pointwise discriminative methods. Empirical results across diverse benchmarks show state-of-the-art performance and robustness to noise, with favorable scalability on large datasets. The method offers a principled generative perspective for dynamic graphs and suggests future work in efficiency, inductive settings, and broader dynamic graph tasks.

Abstract

Temporal link prediction in dynamic graphs is a fundamental problem in many real-world systems. Existing temporal graph neural networks mainly focus on learning representations of historical interactions. Despite their strong performance, these models are still purely discriminative, producing point estimates for future links and lacking an explicit mechanism to capture the uncertainty and sequential structure of future temporal interactions. In this paper, we propose SDG, a novel sequence-level diffusion framework that unifies dynamic graph learning with generative denoising. Specifically, SDG injects noise into the entire historical interaction sequence and jointly reconstructs all interaction embeddings through a conditional denoising process, thereby enabling the model to capture more comprehensive interaction distributions. To align the generative process with temporal link prediction, we employ a cross-attention denoising decoder to guide the reconstruction of the destination sequence and optimize the model in an end-to-end manner. Extensive experiments on various temporal graph benchmarks show that SDG consistently achieves state-of-the-art performance in the temporal link prediction task.

Figures (7)

• Figure 1: The overall architecture of SDG, which injects noise into both historical neighbors and the future destination. SDG reconstructs the destination sequence through a cross-attention denoising decoder conditioning on the encoded history.
• Figure 2: (Top) Performance of SDG with different hyperparameter settings, including (a) Diffusion Steps, (b) Diffusion Loss Ratio, and (c) Intermediate Loss Ratio. (Bottom) T-SNE visualization of learned node embeddings on GoogleLocal.
• Figure 3: Training and Inference time comparison on two large-scale datasets Flickr and ML-20M.
• Figure 4: Inserting noisy edges from 10% to 60%.
• Figure 5: Performance of SDG with different noise schedules on GoogleLocal and YouTube.