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

Yukun Cao, Lisheng Wang, Luobin HuangAAAI 2025 · DOI

Abstract

Temporal Knowledge Graph (TKG) reasoning aims to predict future facts based on historical interactions. However, existing methods often struggle with sparse data for new entities and the inability to distinguish between similar periodic events. We present DPCL-Diff, a framework that integrates a Graph Node Diffusion (GNDiff) model to augment sparse historical data and a Dual-Domain Periodic Contrastive Learning (DPCL) module to map events into specialized geometric spaces. Experimental results demonstrate that DPCL-Diff achieves state-of-the-art performance, notably improving Hits@1 by over 30% on the ICEWS14 dataset.

Introduction

Temporal Knowledge Graphs (TKGs) store facts as quadruples $(s, r, o, t)$ representing a subject, relation, object, and timestamp. Predicting future events in these graphs is challenging due to two primary hurdles:

  1. The Sparse Data Problem: Approximately 40% of events in common datasets involve entities with minimal historical interactions. Traditional models relying on high-frequency patterns fail to generalize to these "cold start" scenarios.
  2. The Similarity Trap: Many periodic events share the same subject and relation but involve different objects. Standard Euclidean representations often fail to capture the nuances required to distinguish these similar-looking events effectively.

The DPCL-Diff Framework

To address these challenges, DPCL-Diff employs a "divide and conquer" strategy through two main components: GNDiff and DPCL.

The DPCL-Diff architecture featuring the Graph Node Diffusion (GNDiff) process and the Dual-Domain Periodic Contrastive Learning (DPCL) module.

Graph Node Diffusion (GNDiff)

GNDiff adapts generative diffusion models to graph nodes to handle sparse data. By injecting noise into sparse historical sequences and learning to denoise them, the model effectively simulates realistic new event data. This process acts as a sophisticated data augmentation technique, allowing the model to reason about entities with limited history by learning from the underlying distribution of the graph's evolution.

Dual-Domain Periodic Contrastive Learning (DPCL)

For recurring events, the model utilizes two distinct geometric domains:

  • Poincaré Space (Hyperbolic): This space is used for periodic events. Hyperbolic geometry is inherently suited for capturing hierarchical and complex relational structures, making it easier to separate entities that appear similar in standard spaces.
  • Euclidean Space: This space is reserved for non-periodic events.

A supervised contrastive loss ensures that representations of similar event classes are clustered together while distinct events are pushed apart in their respective domains.

Key Insights and Results

Geometry and Separation

A critical insight of this work is that Euclidean space is not always optimal for TKG reasoning. By mapping periodic events to Poincaré space, the model leverages hyperbolic curvature to better separate entities sharing the same head and relation but occurring at different frequencies.

2D visualization of entity embeddings showing the separation of related entities (Germany, Berlin, Munich) in the learned space.

Performance Benchmarks

DPCL-Diff was evaluated on four standard benchmarks: ICEWS18, ICEWS14, WIKI, and YAGO. It consistently outperformed 12 baseline models, including CENET and RE-Net. Notable achievements include:

  • ICEWS14: A 30.37% improvement in Hits@1 over the previous state-of-the-art.
  • YAGO: A 6.04% improvement in Mean Reciprocal Rank (MRR).

Ablation studies confirmed that both the generative power of GNDiff and the geometric separation provided by DPCL are essential for these performance gains.

Conclusion

DPCL-Diff represents a significant advancement in TKG reasoning by combining generative diffusion with non-Euclidean geometry. By simulating sparse data and utilizing Poincaré embeddings for periodic patterns, the model overcomes the limitations of traditional embedding-based approaches. Future work may explore dynamic curvature adjustments and scaling to larger batch sizes to further enhance relational pattern capture.

Summary updated

BibTeX
@inproceedings{cao2025dpcl,
  author = {Yukun Cao and Lisheng Wang and Luobin Huang},
  title = {{DPCL-Diff:Temporal Knowledge Graph Reasoning Based on Graph Node Diffusion Model with Dual-Domain Periodic Contrastive Learning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i14.33623},
}
Figure