HTKGHs: Temporal Hypergraphs with Qualifiers

• HTKGHs are advanced data structures that generalize temporal and hyper-relational graphs by encoding n-ary relationships, temporal validity, and qualifier metadata.
• They support group-type and set2set-type relations, allowing compact representation of complex events like geopolitical sanctions and collaborative agreements.
• They underpin novel reasoning approaches, such as temporal inductive logic reasoning and qualifier-aware encoding, which enhance performance on rich event datasets.

Hyper-Relational Temporal Knowledge Generalized Hypergraphs (HTKGHs) are advanced data structures for representing complex, temporally evolving facts in domains where multi-entity interactions and rich semantic attributes frequently occur. They generalize temporal knowledge graph and hyper-relational graph formalisms, extending expressive power via arbitrary-arity hyperedges, interval or pointwise temporal validity, and attribute-rich qualifiers associated with events. HTKGHs unify the handling of n-ary, multi-participant relations, and per-fact temporal reasoning, accommodating intricate real-world scenarios ranging from geopolitical forecasting to procedural workflows and event analysis (Ahrabian et al., 1 Jan 2026, Yang et al., 2022, Ding et al., 2023).

1. Mathematical Definition and Core Structure

At the core of the HTKGH formalism is the representation of knowledge as labeled, time-annotated hyperedges associating sets of entities through relations, enriched with qualifier attributes. The canonical structure is:

$$\mathcal{W} = (\mathcal{E},\, \mathcal{R},\, \mathcal{T},\, \mathcal{H}_1,\, Q)$$

where

• $\mathcal{E}$: finite set of entities,
• $\mathcal{R}$: set of relation labels,
• $\mathcal{T}$: set of discrete timestamps or intervals,
• $$\mathcal{H}_1\subseteq \mathcal{P}^+(\mathcal{E}) \times \mathcal{R} \times \mathcal{P}(\mathcal{E}) \times \mathcal{T}$$: collection of first-order hyperedges encoding $(Actors, Relation, Recipients, Timestamp)$,
• $$Q: \mathcal{H}_1 \to 2^{\mathcal{R}\times\mathcal{E}}$$: mapping associating each hyperedge with its qualifiers.

HTKGHs strictly subsume classic Temporal Knowledge Graphs (TKGs) and Hyper-Relational Temporal Knowledge Graphs (HTKGs) by relaxing the binary subject-object restriction and allowing for general groupwise, group-to-group, and n-ary relations with attached qualifiers at each time point (Ahrabian et al., 1 Jan 2026, Ding et al., 2023). The representation permits modeling facts such as:

$$\bigl( (\{\text{US},\text{UK}\},\, \texttt{sanction},\, \{\text{Russia},\text{Belarus}\},\, 2022),\, \{ (\texttt{cause}, \text{Russo-Ukrainian war}) \} \bigr)$$

Qualifiers may encode provenance, auxiliary modalities, and interval boundaries (e.g., “start time”, “end time”, “location”).

2. Fact Typology: Group-Type and Set2Set-Type Relations

HTKGHs support two complex fact types not handled compactly by HTKGs (Ahrabian et al., 1 Jan 2026):

• Group-Type Facts: Capture interaction among a set of actors without distinct recipients. Example: “China, Japan, South Korea jointly sign a free-trade deal in 2023 on cars, chips, and oil.” This is represented as a single hyperedge with actors as a set, empty recipients, and qualifiers as commodity tags.
• Set2Set-Type Facts: Encode relations where a set of actors acts on a set of recipients (e.g., sanctions, agreements). Example: “US and UK sanction Russia and Belarus in 2022 because of the Russo-Ukrainian war.” HTKGH avoids the combinatorial expansion required in pairwise decomposition by directly linking both sets in a single edge.

This facility precludes the need for reification (group nodes) or redundant fact decomposition, preserving jointness and minimizing sparsity (Ahrabian et al., 1 Jan 2026, Yang et al., 2022). A plausible implication is more compact and information-preserving representation for real-world event corpora.

3. Temporal Reasoning and Interval Structures

HTKGH edges are time-annotated, with temporal validity encoded either as single timestamps or as intervals $[t_s, t_e]$ (Yang et al., 2022). The edge set $\mathcal{H}_1$ thus generalizes event duration modeling—critical for procedural, event-centric, or evolving knowledge domains.

Temporal reasoning is enabled by associating each event with explicit validity and leveraging interval-relational algebra (e.g., Allen’s interval relations: BEFORE, MEETS, OVERLAPS, EQUAL, etc.), used both for fact indexing and for logic rule induction. In the TILR algorithm, temporal chain-like rules encode patterns such as:

$$P(X^0,X^n,T^n) \leftarrow P^1(X^0,X^1,T^1)\, \psi^1\, P^2(X^1,X^2,T^2)\, \ldots\, \psi^{n-1}\, P^n(X^{n-1},X^n,T^n)$$

with each $\psi^i$ an interval relation over $(T^i, T^{i+1})$ (Yang et al., 2022). This suggests logic-based queries and predictions over HTKGHs can exploit both relational structure and temporal constraint propagation.

4. Datasets and Real-World Instantiations

Several benchmarks instantiate HTKGHs in various domains:

Dataset	Domain	# Entities	# Relations	# Hyperedges	Temporal Annotation	Key Modality
htkgh-polecat	Geopolitical/POLECAT	5,268	42	~556K	Jan 2018–Jul 2024	Countries, sectors, events
YouCook2-HG	Cooking procedures	—	—	~27k	Video clip intervals	Recipe steps, ingredients
nuScenes-HG	Autonomous driving	—	—	~7.1M	20 s window	Vehicle actions, actors
Wiki-hy, YAGO-hy	Wikipedia/YAGO	3,392–1,739	25–9	~21k–7k	~507, ~187 tspts	Entities, time qualifiers

The htkgh-polecat dataset (Ahrabian et al., 1 Jan 2026) aggregates POLECAT events, stratifying by country, sector, and qualifiers (location, context tags). In YouCook2-HG and nuScenes-HG (Yang et al., 2022), procedural and driving events are parsed into temporally delimited, n-ary event hyperedges. Wiki-hy and YAGO-hy (Ding et al., 2023) enrich facts from Wikidata/YAGO with hyper-relational temporal qualifiers. Notably, these datasets demonstrate empirical coverage of multi-actor, set2set, and qualified temporal edges.

5. Reasoning and Learning Over HTKGHs

Several machine learning and symbolic reasoning approaches operate on HTKGH structures:

• Temporal Inductive Logic Reasoning (TILR) (Yang et al., 2022): Applies multi-start random B-walks for path sampling in B-graphs (hypergraphs with $|T_e|=1$), learns chain-like logical rules with temporal interval constraints, and trains a differentiable model via cross-entropy on positive/negative queries.
• Qualifier-Aware Temporal Graph Encoder (QATGE) (Ding et al., 2023): Embeds entities, relations, qualifiers, and time features, aggregates qualifier embeddings per fact via learned MLP or attention, and fuses via time-dependent gating. Score computation follows via bilinear operations, with negative sampling for training.
• LLM-Based Relation Prediction (Ahrabian et al., 1 Jan 2026): Evaluated on htkgh-polecat, state-of-the-art LLMs (Gemma, Qwen, Llama, GPT-OSS) predict missing relation labels given event context. The pipeline samples up to 100 most recent related facts as context and assesses accuracy via Hits@k and MRR.

Empirical evidence shows that on large hypergraph-structured datasets, these models outperform graph-only or temporal-only baselines. For instance, HypeTKG reaches MRR≈0.73/0.82 and Hits@10≈0.90/0.92 on Wiki-hy/YAGO-hy, exceeding static KG models by 5–10 points (Ding et al., 2023). TILR’s full model achieves MRR=0.72 and Hits@3=76.0 on YouCook2-HG (Yang et al., 2022). Qualifier-aware and temporally gated modeling are confirmed as indispensable.

6. Empirical Insights and Model Benchmarking

LLM benchmarking on htkgh-polecat demonstrates that both non-thinking and thinking variants of large models outperform heuristic baselines across entity, location, and context filters (Ahrabian et al., 1 Jan 2026).

Model	Accuracy (@1)	MRR	Context Utilization
gemma-3-12b-it	60.6%	—	moderate
Qwen3-4B-Instruct	58.1%	—	moderate
gpt-oss-20b	63.9%	—	deep (chain-of-thought)
Frequency heuristic	49.6%	—	shallow

Context quality, particularly via location/context filters, is critical for maximizing LLM accuracy. Entity shuffling minimally impacts performance; relation shuffling leads to erratic results, indicating symbolic dependence. GNN-based models collapse context into single vectors and underperform when tight filters are imposed.

7. Challenges and Future Research Directions

The principal technical barriers identified are:

• Context retrieval and ranking: Selection of relevant historical edges for query conditioning remains unsolved.
• Model robustness: LLM misformatting occurs under heavy context loads.
• Graph neural representational bottlenecks: Existing architectures lose fact granularity under windowed compression.
• Expressive extensions: Nesting of hyperedges, modeling of multiple actor/recipient groups, and more flexible interval semantics.

Future research directions include:

• End-to-end retrieval and reasoning pipelines combining context selection with logic inference.
• Dynamic hypergraph neural nets that preserve per-fact structure across time windows.
• Extension of SPARQL-like query languages to HTKGHs for symbolic and subgraph-based reasoning.
• Deployment in domains such as biological pathway analysis and supply-chain event tracking, where n-ary/modal relations predominate.

HTKGHs establish a unified substrate for encoding, reasoning, and learning in domains requiring high-order relational modeling and explicit temporal semantics, validated across symbolic, neural, and language-based techniques (Ding et al., 2023, Yang et al., 2022, Ahrabian et al., 1 Jan 2026).