DPCL-Diff: The Temporal Knowledge Graph Reasoning Based on Graph Node Diffusion Model with Dual-Domain Periodic Contrastive Learning

TL;DR

DPCL-Diff tackles the sparsity of new events in Temporal Knowledge Graphs by combining Graph Node Diffusion (GNDiff), which generates high-quality samples for unseen events, with Dual-Domain Periodic Contrastive Learning (DPCL) that distinguishes periodic from non-periodic events by mapping entities into Poincaré and Euclidean spaces. GNDiff implements a forward diffusion over discrete graph nodes with an absorbing inactive state and a learnable reverse denoising process, optimizing via a variational lower bound $\mathcal{L}_{\mathrm{diff}}$, while DPCL uses a dual-space contrastive objective combining $\mathcal{L}_{\mathrm{ce}}$ and $\mathcal{L}_{\mathrm{sup}}$ to refine periodic relationships. The two components are fused in a joint objective $\mathcal{L}_{\mathrm{dpcl}}=\alpha\mathcal{L}_{\mathrm{diff}}+(1-\alpha)(\mathcal{L}_{\mathrm{ce}}+\mathcal{L}_{\mathrm{sup}})$, and inference averages the diffusion-based probability $P_{\mathrm{diff}}$ with the DPCL-based probability $P_{\mathrm{dpcl}}$ to predict tail entities. Experiments on ICEWS14/ICEWS18/WIKI/YAGO show that DPCL-Diff consistently surpasses 12 baselines in MRR and Hits@k, with notable gains on new-event and periodic-event scenarios, demonstrating the effectiveness of combining graph diffusion for data generation with dual-domain geometric reasoning for discrimination. The work advances practical TKGR by enabling robust reasoning over both unseen and recurring events, with implications for forecasting and decision-making in dynamic knowledge systems.

Abstract

Temporal knowledge graph (TKG) reasoning that infers future missing facts is an essential and challenging task. Predicting future events typically relies on closely related historical facts, yielding more accurate results for repetitive or periodic events. However, for future events with sparse historical interactions, the effectiveness of this method, which focuses on leveraging high-frequency historical information, diminishes. Recently, the capabilities of diffusion models in image generation have opened new opportunities for TKG reasoning. Therefore, we propose a graph node diffusion model with dual-domain periodic contrastive learning (DPCL-Diff). Graph node diffusion model (GNDiff) introduces noise into sparsely related events to simulate new events, generating high-quality data that better conforms to the actual distribution. This generative mechanism significantly enhances the model's ability to reason about new events. Additionally, the dual-domain periodic contrastive learning (DPCL) maps periodic and non-periodic event entities to Poincaré and Euclidean spaces, leveraging their characteristics to distinguish similar periodic events effectively. Experimental results on four public datasets demonstrate that DPCL-Diff significantly outperforms state-of-the-art TKG models in event prediction, demonstrating our approach's effectiveness. This study also investigates the combined effectiveness of GNDiff and DPCL in TKG tasks.

Figures (3)

• Figure 1: Overall architecture of DPCL-Diff. The first part is GNDiff, and the second part is DPCL.
• Figure 2: Illustration of 2D entity embedding learned by Euc (left) and Hyp (right) on YAGO.
• Figure 3: Results of hyper-parameters $\alpha$ and $\lambda$ of DPCL-Diff on ICEWS14 and YAGO.