TiRGN: Time-Guided Recurrent Graph Network with Local-Global Historical Patterns for Temporal Knowledge Graph Reasoning

Yujia Li, Shiliang Sun, Jing ZhaoIJCAI 2022 · DOI

Introduction

Knowledge Graphs (KGs) are essential for structured information representation, yet static KGs often fail to capture the temporal dynamics of real-world events. Temporal Knowledge Graphs (TKGs) address this by representing facts as quadruples: (subject, relation, object, timestamp). A critical challenge in TKG research is the extrapolation problem—predicting future events based on historical data. Existing models often struggle to integrate the diverse "laws of historical development," such as sequential, repetitive, and cyclical patterns.

Methodology

TiRGN (Time-Guided Recurrent Graph Network) addresses these challenges through a three-pronged architecture designed to capture complex historical patterns simultaneously.

The TiRGN architecture featuring Local Recurrent Encoder, Global History Encoder, and Time-Guided Decoder.

Local Recurrent Encoder

This module focuses on capturing sequential patterns in recent history. It utilizes a 1D convolution-based Graph Convolutional Network (GCN) to aggregate neighborhood information at each timestamp. A unique double recurrent mechanism employing two separate Gated Recurrent Units (GRUs) evolves both entity and relation embeddings over time. This allows the model to track the co-evolution of actors and their interaction types.

Global History Encoder

To account for repetitive patterns without the bias inherent in frequency-based models, TiRGN incorporates a Global History Encoder. This module tracks whether a specific fact has ever occurred in the past, creating a candidate set that constrains the search space. This "global memory" ensures that rare but historically repeated facts are not overlooked.

Time-Guided Decoder

To model cyclical patterns (e.g., periodic political events), the framework integrates model-independent time vectors inspired by Time2Vec. By combining sine-based periodic vectors and linear non-periodic vectors, the decoder can learn the "rhythm" of history, guiding the final prediction toward events likely to occur at specific intervals.

Experimental Setup

The model was evaluated on six standard TKG benchmark datasets:

  • ICEWS18, ICEWS14, and ICEWS05-15: Political event data from the Integrated Crisis Early Warning System.
  • WIKI and YAGO: Temporal knowledge bases derived from Wikipedia.
  • GDELT: A massive and noisy dataset from the Global Database of Events, Language, and Tone.

Key Results

TiRGN consistently outperformed state-of-the-art baselines, including RE-GCN, CyGNet, and TITer, across all evaluated datasets.

Performance comparison of TiRGN against state-of-the-art baselines on ICEWS datasets.

On the WIKI dataset, TiRGN achieved an MRR (Mean Reciprocal Rank) of 81.65, a notable improvement over RE-GCN's 78.53. It also demonstrated robustness on the large-scale GDELT dataset, maintaining a performance lead despite the high volume and noise of the event streams. Ablation studies confirmed that while the local recurrent encoder has the most significant impact, the global constraints and periodic time vectors are essential for achieving state-of-the-art results.

Discussion and Insights

The success of TiRGN highlights several key insights:

  • History as a Constraint: Using global history to narrow the prediction scope effectively reduces noise and improves accuracy.
  • Occurrence over Frequency: Balancing recent frequency (via the local encoder) with historical possibility (via the global encoder) prevents the model from being overly biased toward "popular" but irrelevant facts.
  • Periodicity Guidance: Incorporating sinusoidal time vectors allows the model to capture the inherent cycles in political and social events.

Conclusion and Future Work

TiRGN represents a significant step forward in TKG reasoning by integrating local and global historical patterns. Future work may explore multi-hop global history to further refine prediction constraints and investigate more flexible periodic functions to better model non-sinusoidal real-world cycles. While the current architecture is computationally intensive, its performance gains justify the complexity for high-stakes forecasting tasks.

Summary updated

BibTeX
@inproceedings{li2022tirgn,
  author = {Yujia Li and Shiliang Sun and Jing Zhao},
  title = {{TiRGN: Time-Guided Recurrent Graph Network with Local-Global Historical Patterns for Temporal Knowledge Graph Reasoning}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.24963/ijcai.2022/299},
}
Figure