Temporal Knowledge Graph Extrapolation via Causal Subhistory Identification

Kai Chen, Ye Wang, Xin Song, Siwei Chen, Han Yu, Aiping Li, Chunming Hu, Xianxian Li, Xingcheng Fu

Abstract

Predicting future events in Temporal Knowledge Graphs (TKGs) is a fundamental task with applications ranging from geopolitics to epidemiology. However, existing models often suffer from "spurious correlations"—statistical coincidences in historical data that lack causal power. This paper introduces Causal Subhistory Identification (CSI), a novel framework designed to disentangle genuine causal factors from misleading shortcuts. By extracting query-relevant subhistories and employing a dual-view disentangled representation learning approach, CSI isolates causal features and applies causal intervention via backdoor adjustment. Experimental results across four major TKG benchmarks demonstrate that CSI consistently outperforms state-of-the-art models and exhibits exceptional robustness against data sparsity and noise.

Introduction

Temporal Knowledge Graphs (TKGs) store facts as quadruples: (subject, relation, object, timestamp). The task of extrapolation—predicting future facts based on historical sequences—is challenging because models often latch onto non-causal links, known as spurious correlations or "shortcuts." For instance, if a virus spreads while phone call frequency increases, a standard model might incorrectly learn that phone calls cause the spread, whereas physical contact is the true causal factor. These shortcuts make models fragile; if the correlation changes in a new environment, predictive accuracy collapses. CSI addresses this by identifying and isolating the true causal subhistory behind an event.

Methodology: The CSI Framework

The CSI strategy follows a three-step process to isolate the "why" behind an event:

  1. Query-Relevant Subhistory Extraction: Instead of processing the entire graph, the model extracts a localized subgraph centered around the entities and relations of the specific query.
  2. Disentangled Representation Learning: Using an attention mechanism, the subhistory is split into two parts: the Causal Subhistory ($G_c$), containing information that truly influences the outcome, and the Shortcut Subhistory ($G_s$), containing non-causal features. A KL-divergence loss is applied to ensure the shortcut representation remains uninformative.
  3. Backdoor Adjustment via Causal Intervention: Inspired by Judea Pearl’s causal theory, CSI implements a "backdoor adjustment." During training, shortcut representations are randomly permuted. This forces the model to rely solely on stable causal features, effectively cutting off the influence of confounding shortcuts.
The CSI framework architecture, illustrating the transition from raw TKG data to disentangled causal and shortcut representations.

Experimental Evaluation

The researchers evaluated CSI on four standard TKG benchmarks: GDELT, ICEWS14*, ICEWS05-15, and WIKI. These datasets cover diverse domains from 15-minute news cycles to annual historical facts.

Key Results

CSI achieved state-of-the-art performance, outperforming 11 top-tier baseline models (including RE-NET, TiRGN, and RPC) across all datasets. For example, on the GDELT dataset, it achieved a Mean Reciprocal Rank (MRR) of 25.6%, a significant improvement over previous methods.

Performance comparison of CSI against baseline models across four TKG benchmarks. CSI (ours) consistently achieves the highest scores in MRR and Hits@N metrics.

Robustness and Explainability

CSI demonstrated extreme robustness in stress tests involving data sparsity and noise. Even when 75% of training facts were removed or replaced with noise, CSI maintained high performance (MRR > 45 on ICEWS14*), while baseline models collapsed. Furthermore, the model provides explainability; path scores for causal clues were significantly higher (0.22) than for spurious clues (0.04), allowing users to understand the rationale behind forecasts.

Conclusion

Causal Subhistory Identification represents a significant step forward in making TKG forecasting more robust and interpretable. By moving beyond simple statistical correlation and embracing causal inference, CSI ensures that models learn the underlying mechanisms of events rather than coincidental patterns. Future work will explore the computational efficiency of dual GNN encoders and the evolution of causal relationships over long time horizons.

Summary updated

BibTeX
@inproceedings{chen2024temporal,
  author = {Kai Chen and Ye Wang and Xin Song and Siwei Chen and Han Yu and Aiping Li and Chunming Hu and Xianxian Li and Xingcheng Fu},
  title = {{Temporal Knowledge Graph Extrapolation via Causal Subhistory Identification}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
  url = {https://openreview.net/forum?id=q2Jp5mUXic},
}
Figure