Toward Better Temporal Structures for Geopolitical Events Forecasting

TL;DR

This paper introduces HTKGH, a generalized hypergraph formalism that extends hyper-relational temporal knowledge graphs (HTKGs) to efficiently model high-order geopolitical events involving more than two primary entities. It formalizes HTKGH, demonstrates backward compatibility with HTKGs, and presents htkgh-polecat, a POLECAT-based dataset that highlights two complex fact types (Group-Type and Set2Set-Type) with substantial real-world coverage. The authors benchmark 13 LLMs and compare them with graph-based baselines on relation-prediction tasks, showing that larger and well-aligned models excel with richer contextual information, while GNNs outperform in low-context settings but lose relative advantage as context quality improves. The work demonstrates the potential of LLMs to adapt to complex, higher-order factual structures and points to the importance of high-quality context retrieval, data anonymization, and future improvements in knowledge representation and SParQL-like querying for HTKGH-enabled systems.

Abstract

Forecasting on geopolitical temporal knowledge graphs (TKGs) through the lens of large language models (LLMs) has recently gained traction. While TKGs and their generalization, hyper-relational temporal knowledge graphs (HTKGs), offer a straightforward structure to represent simple temporal relationships, they lack the expressive power to convey complex facts efficiently. One of the critical limitations of HTKGs is a lack of support for more than two primary entities in temporal facts, which commonly occur in real-world events. To address this limitation, in this work, we study a generalization of HTKGs, Hyper-Relational Temporal Knowledge Generalized Hypergraphs (HTKGHs). We first derive a formalization for HTKGHs, demonstrating their backward compatibility while supporting two complex types of facts commonly found in geopolitical incidents. Then, utilizing this formalization, we introduce the htkgh-polecat dataset, built upon the global event database POLECAT. Finally, we benchmark and analyze popular LLMs on the relation prediction task, providing insights into their adaptability and capabilities in complex forecasting scenarios.

Figures (13)

• Figure 1: An overview of how facts are expressed in TKGs, HTKGs, and HTKGHs. HTKGHs adds efficient support for complex facts involving many (i.e., more than two) primary entities, expanding the possible expression scenarios such as multi-national treaties.
• Figure 2: (a) The original temporal fact representing a joint trade deal between $\{\text{China, Japan, and South Korea}\}$ on $\{\text{Cars, Chips, and Oil}\}$. (b) The decomposed version of the original fact designed for HTKGs. Notably, the decomposed version requires a lot of redundancy to represent, and without an additional entity representing the group of countries, it is indistinguishable from three separate trade deals between these countries.
• Figure 3: Vietnam War is an example of a complex geopolitical event that involved two coalitions of countries (i.e., two sets of entities) engaging in a war (i.e., an action) in locations such as the South China Sea and Gulf of Thailand (i.e., a set of qualifiers).
• Figure 4: Relation prediction accuracy (%) on htkgh-polecat context variations over the number of retrieved contextual samples. From each model family, we only display the best-performing member.
• Figure 5: Relation prediction accuracy (%) on htkgh-polecat for different model sizes in the (a) Gemma-3 and (b) Qwen3 families. Legend: N $\rightarrow$ non-thinking, T $\rightarrow$ thinking.