Hyper-Relational Temporal KGs

• Hyper-Relational Temporal Knowledge Graphs (HTKGs) are advanced data structures that represent evolving, time-stamped facts enriched with qualifier pairs for refined semantic context.
• They enable robust temporal reasoning by jointly embedding entities, relations, time points, and qualifiers, supporting tasks like link prediction and forecasting.
• Empirical evaluations on benchmark datasets show that integrating temporal and qualifier dimensions significantly improves inference accuracy in complex, dynamic domains.

A hyper-relational temporal knowledge graph (HTKG) is a data structure that encodes evolving world knowledge by modeling time-stamped, n-ary relations between entities, augmented with qualifiers that refine the semantics of each fact. HTKGs unify the expressivity of hyper-relational knowledge graphs—which allow key-value qualifiers per fact—with explicit temporal anchoring, enabling precise representation and inference over temporally contextualized, richly qualified facts. Recent work has established formal foundations for HTKGs, developed benchmark datasets, and introduced models that jointly embed entities, relations, time, and qualifiers for tasks such as link prediction, temporal reasoning, and forecasting (Ahrabian et al., 1 Jan 2026, Ding et al., 2023, Un et al., 17 May 2025).

1. Formal Structure and Foundations

Let $\mathcal{E}$ denote the set of entities, $\mathcal{R}$ the set of relations, and $\mathcal{T}$ the set of timestamps. A standard temporal knowledge graph (TKG) consists of quadruples $(s, r, o, t)$ where $s, o \in \mathcal{E}$, $r \in \mathcal{R}$, and $t \in \mathcal{T}$.

HTKGs extend this representation by attaching to each quadruple an auxiliary set of qualifiers $Q = \{(k_1, v_1), \ldots, (k_m, v_m)\}$, where $k_i$ are qualifier keys (relations) and $v_i$ are qualifier values (entities or literals): $$\bigl(s, r, o,\, Q,\, t\bigr), \qquad s,o \in \mathcal{E},\, r \in \mathcal{R},\, t \in \mathcal{T},\, Q \subseteq \mathcal{R}\times(\mathcal{E}\cup\mathcal{L})$$ Alternatively, one can distinguish between non-temporal qualifiers $Q^{(\mathrm{non\text{-}time})}$ and time qualifiers $Q^{(\mathrm{time})}$, with temporal validity defined via explicit intervals $S = (\tau_s, \tau_e)$, where $\tau_s, \tau_e \in \mathbb{R} \cup \{\pm\infty\}$, to capture "since," "until," "period," and "time-invariant" modalities (Un et al., 17 May 2025).

Comparison of Notations: | Model | Entities | Relations | Time/Timestamps | Qualifiers | |------------|-------------|---------------------|------------------------------|-----------------------------------------------| | TKG | $s,o$ | $r$ | $t$ | None | | HTKG [1] | $s,o$ | $r$ | $t$ | $Q\subseteq \mathcal{R}\times(\mathcal{E})$ | | HTKG [2] | $h,t$ | $r$ | $Q^{(\mathrm{time})}$ | $Q^{(\mathrm{non\text{-}time})}$ |

This formalism supports temporal reasoning while retaining the modeling flexibility required for real-world facts with additional context (location, role, instrumentality).

2. Expressiveness and Limitations

HTKGs can represent time-stamped binary relations with arbitrary qualifiers. However, they are fundamentally limited to encoding facts involving exactly two "primary" entities per relation. Two categories of complex facts frequently arising in domains such as geopolitics are not naturally supported:

• Group-type ("clique") facts: Multiple entities concurrently share the same relation at a given time (e.g., a multi-country summit).
• Set2Set-type ("biclique") facts: A group of actors acts on a group of recipients (e.g., a coalition imposing sanctions on another coalition).

To encode such facts within HTKGs, it is necessary to decompose the group relation into multiple binary facts or engineer auxiliary group entities, which introduces redundancy, sparsity, and inference complexity. This motivates extensions to generalized hypergraph constructions (Ahrabian et al., 1 Jan 2026).

3. Datasets and Schema Characteristics

Several large-scale benchmark datasets have been introduced to fuel the empirical study of HTKGs:

• Wiki-hy and YAGO-hy (Ding et al., 2023): Derived from Wikidata and YAGO, these datasets include facts (train/valid/test splits), with on average 2–3.5 qualifier-key/value pairs per fact, and temporal spans ranging from centuries to the present. Qualifier keys cover event-specific information (e.g., location, instrument, rank). For example, Wiki-hy(100) contains 21,210 train facts, with 3,392 primary entities and 25 primary relations.
• WD-HT50K and JT-HT20K (Un et al., 17 May 2025): Wikidata-derived, with rich qualifier and temporal annotation. WD-HT50K contains 50K main facts (~3.2 qualifiers/fact, 30% temporal qualifiers). Distribution among temporal validity types: 24% since, 21% until, 43% period, 12% time-invariant.
• htkgh-polecat (Ahrabian et al., 1 Jan 2026): Designed for high temporal and relational complexity, this dataset (built from POLECAT, 2018–2024) encodes 556,000 facts about geopolitical events, using an ontology where entities fuse country and sector, relations couple event type and mode, and qualifiers annotate location and context. Group and Set2Set facts constitute ~23.6% of all edges.

All these datasets are designed to provide diverse, temporally anchored, and qualified facts for robust benchmarking of link prediction, relation forecasting, and qualifier inference tasks.

4. Modeling and Learning Frameworks

State-of-the-art models for HTKGs jointly embed entities, relations, temporal features, and qualifier information. Notable frameworks include:

• HypeTKG (Ding et al., 2023): Embeds all elements (entities, relations, qualifier keys/values, and timestamps), uses a multi-layer graph attention (QATGE) mechanism for qualifier aggregation, and fuses temporal information via a dedicated time encoder. The scoring function is an $L_2$ distance in the embedded space.
• VITA (Un et al., 17 May 2025): Implements a versatile time representation capturing all validity modalities via time-value and timespan embeddings. Non-temporal qualifiers are aggregated using key-value Hadamard products, and a gating mechanism composes the final time embedding for each fact. The overall fact representation sums the entity, relation, qualifier, and time embeddings, with scoring by dot product.

These approaches are trained with margin ranking losses, using negative sampling by corrupting head, tail, or time elements. Empirical ablations consistently show that modeling both temporal and qualifier dimensions is crucial for performance.

5. Empirical Results

HTKG models have been evaluated extensively on task suites including entity prediction, relation and qualifier inference, temporal (interval) prediction, and numeric literal prediction. Key findings include:

• In entity prediction on WD-HT50K, VITA achieves MRR 0.332 versus 0.305 for prior best models (+8.9% relative). For time prediction, MRR improves from 0.558 to 0.612 (+9.7%) (Un et al., 17 May 2025).
• Qualifier modeling is fundamental: ablating the qualifier aggregation from HypeTKG decreases MRR by 0.031 points; ablating both temporal and qualifier encodings drops MRR by 0.052 (Ding et al., 2023).
• LLMs benchmarked on htkgh-polecat (relation prediction) demonstrate that scaling context size and adopting chain-of-thought reasoning can yield +6–7% accuracy gains. Supervised GNNs outperform LLMs under loose context, but LLMs excel under high-relevance, sparse history, with up to 21% gains (Ahrabian et al., 1 Jan 2026).

These advances highlight that temporal reasoning, expressive time representations, and comprehensive qualifier aggregation are essential design principles for HTKG models.

6. Extensions, Generalizations, and Open Problems

The fundamental limitations of HTKGs regarding the expressivity of higher-arity and group–group relations have prompted development of Hyper-Relational Temporal Knowledge Generalized Hypergraphs (HTKGHs) (Ahrabian et al., 1 Jan 2026). HTKGHs generalize the HTKG edge model to support arbitrary sets of actors and recipients within a single hyperedge, thereby compactly encoding group-type and set2set-type facts. They retain backward compatibility with classic HTKGs via embedding of binary facts.

Challenges and potential future research directions include:

• Efficient hypergraph reasoning models capable of learning over general higher-arity structures.
• Integrating time-invariant background knowledge into the temporal reasoning pipeline.
• Qualifier representation for complex, semantically heterogeneous qualifiers beyond timestamp and location.
• Inductive generalization to unseen qualifier schemas and multi-hop, path-based inference across HTKGs.
• Leveraging pretrained LLMs to enhance qualifier and time interval encoding beyond structured fields.

Recent evidence suggests that the combination of time-contextualized embeddings, flexible qualifier aggregation, and generalized hypergraph representations will progressively enhance the fidelity and completeness of factual reasoning over dynamic, real-world domains (Ahrabian et al., 1 Jan 2026, Un et al., 17 May 2025, Ding et al., 2023).