FITTER: Vocabulary-Agnostic Cross-Domain Inference on Temporal Knowledge Graphs

FITTER: Vocabulary-Agnostic Cross-Domain Inference on Temporal Knowledge Graphs

Abstract

Temporal knowledge graphs are central to many uses of the Semantic Web, but existing completion methods assume the entities, relation names, and timestamps to be reasoned about are already known at training time, restricting each model to a single graph and vocabulary. We propose FITTER, the first fully-inductive structural model for temporal knowledge graph link prediction that supports cross-domain transfer: the inference graph may contain entirely unseen entities, relation names, and timestamps drawn from a different domain. FITTER represents each predicate by its interaction patterns with others and time through encodings of relative rather than absolute ordering; message-passing fuses local and global temporal context to produce vocabulary-agnostic embeddings. We prove the temporal encoding is time-shift invariant and evaluate FITTER on cross-domain, cross-graph transfer over six temporal knowledge graph benchmarks of diverse domains, granularities, and time spans. FITTER consistently outperforms inductive baselines without retraining, indicating that vocabulary-agnostic structural learning is a viable foundation for inference over the heterogeneous knowledge graphs of the Semantic Web.

1. Introduction

Temporal Knowledge Graphs (TKGs) are a central object of study in the Semantic Web: large openly published graphs such as Wikidata, DBpedia, and YAGO carry millions of facts that hold during specific intervals or at specific points in time. A TKG fact is commonly treated as a time-stamped quadruple (s, p, o, τ), where s, p, o, and τ denote the subject entity, predicate, object entity, and timestamp. Temporal link prediction, i.e., predicting missing entities in queries of the form (s, p, ?, τ) or (?, p, o, τ), is one of the central problems in temporal knowledge graph research.

Existing Temporal Knowledge Graph Embedding (TKGE) models are primarily developed for temporal interpolation or extrapolation settings, where the entity and relation vocabularies are shared between training and inference. In practice, however, temporal knowledge graphs often differ substantially across domains. For example, diplomatic event graphs such as ICEWS contain dense daily interactions, while encyclopedic graphs such as YAGO describe sparse long-term facts spanning decades or centuries. These datasets also differ in temporal granularity, time span, and relational structure.

This heterogeneity raises a natural transfer challenge: can a model trained on one TKG generalize to a different TKG without retraining? Such a capability is particularly useful in cold-start scenarios, where a target TKG is newly constructed or contains only limited training data. Most existing TKGE models rely on dataset-specific entity, relation, and timestamp embeddings. Consequently, their learned representations cannot be directly applied to graphs containing unseen entities, relations, or timestamps.

In this work, we study cross-domain transfer for temporal knowledge graphs under a fully-inductive setting, where the training and inference graphs contain disjoint entity, relation, and timestamp sets. We propose FITTER (Fully Inductive Time-aware Transferable Representation), a structure-driven TKGE framework designed for cross-domain transfer across heterogeneous temporal knowledge graphs.

FITTER addresses two key challenges in transferable temporal reasoning. First, temporal knowledge graphs may operate at different granularities and over different time spans, making dataset-specific timestamp embeddings difficult to transfer across domains. To address this, FITTER represents temporal information using sinusoidal positional encodings over snapshot indices, enabling the model to capture relative temporal ordering independently of absolute timestamps. Second, to transfer structural knowledge across unseen vocabularies, FITTER constructs universal relation interaction graphs based on vocabulary-agnostic interaction types and learns adaptive entity and relation representations through temporal-aware message passing. In addition, FITTER combines local and global temporal contexts to capture both short-term event dependencies and long-range temporal patterns.

The contributions are summarized as follows:

• We formulate cross-domain transfer for temporal knowledge graphs as a fully-inductive link prediction problem.

• We propose FITTER, a transferable TKGE framework that avoids dataset-specific entity, relation, and timestamp embeddings.

• We introduce transferable temporal representations based on relative temporal ordering together with vocabulary-agnostic relational structure learning.

• We conduct extensive experiments across heterogeneous temporal knowledge graphs and demonstrate strong cross-domain transfer performance across domains and temporal granularities.

2. Task Formulation

A temporal knowledge graph is a tuple G = (V, R, T, Q), where V is a set of entities, R a set of relations, T an ordered set of timestamps, and Q ⊆ V × R × V × T a set of temporal facts. The temporal link prediction task asks: given a query (s, p, ?, τ) or (?, p, o, τ), predict the missing entity.

Inference settings are characterized along two orthogonal axes: a vocabulary axis and a temporal axis. The vocabulary axis describes whether the entity and relation vocabularies are shared between the training and inference graphs. The temporal axis describes whether inference timestamps are evaluated within an observed temporal range (interpolation) or on future timestamps beyond the observed range (extrapolation).

In transductive settings, the training and inference graphs share the same entity and relation vocabularies: Vtrain = Vinf, Rtrain = Rinf. In fully-inductive settings, the training and inference graphs contain disjoint entity, relation, and timestamp sets: Vtrain ∩ Vinf = ∅, Rtrain ∩ Rinf = ∅, Ttrain ∩ Tinf = ∅. A practically important special case is cross-domain transfer, where the training and inference graphs originate from different source TKGs.

In interpolation settings, message, validation, and test facts share the same timestamp range: Tmsg = Tvalid = Ttest. In extrapolation settings, the inference graph is split chronologically such that: max(Tmsg) < min(Tvalid) < min(Ttest).

This paper studies fully-inductive inference under cross-domain transfer, where a model trained on one source TKG is evaluated directly on a different target TKG without retraining, under both interpolation and extrapolation conditions.

3. Related Work

Transductive temporal interpolation models (e.g., TTransE, TA-DistMult, TComplEx, TNTComplEx, TLTKGE, HGE) rely on dataset-specific entity, relation, and timestamp embeddings, which prevents direct application to graphs with unseen vocabularies. Transductive temporal extrapolation models (e.g., RE-NET, CyGNet, TiRGN, TLogic, xERTE) also assume vocabulary overlap between training and inference.

Structure-based fully-inductive learning methods for static KGs (e.g., NBFNet, Grail, INDIGO, Morse, INGRAM, ULTRA, TRIX, GraphOracle) enable transfer across entirely unseen entity and relation sets by constructing vocabulary-agnostic relation interaction graphs or relation dependency structures. However, all of these approaches are designed for static knowledge graphs and therefore do not model temporal dynamics, temporal ordering, or heterogeneous temporal granularities required for temporal knowledge graphs.

Text-based and LLM-based transfer for TKGs (e.g., ICL, zRLLM, GenTKG, Chain-of-History) leverage semantic information from entity descriptions, relation names, or in-context temporal reasoning to support transfer across unseen temporal facts. However, their transferability primarily arises from textual semantics rather than structural temporal representations. In contrast, FITTER is purely structure-driven: it learns transferable temporal and relational patterns directly from graph structure without requiring textual annotations or external language models.

4. NBFNet-Style Fully-Inductive KG Reasoning

FITTER builds on ULTRA, a fully-inductive KG reasoning framework derived from NBFNet-style message passing. ULTRA is particularly suitable for cross-domain transfer because it avoids dataset-specific entity and relation embeddings, enabling inference on entirely unseen graphs. Given a query (s, r, ?), these models perform query-conditioned reasoning by propagating information over the graph. The source entity s is initialized using the query relation r, and message passing iteratively aggregates relational information from neighboring entities. ULTRA constructs relation representations from a vocabulary-agnostic relation interaction graph, where relations are treated as nodes and connected through structural interaction types: head-to-head, head-to-tail, tail-to-head, and tail-to-tail interactions.

5. FITTER

FITTER introduces three temporal components on top of the relational backbone: (1) transferable temporal encodings that capture relative ordering independently of absolute timestamps; (2) temporal-aware message passing that incorporates time into fact propagation; and (3) local/global context fusion that integrates short-term and long-range temporal dependencies.

5.1 Transferable Temporal Encoding

FITTER represents temporal information via relative snapshot ordering rather than absolute timestamps, enabling transfer across TKGs with different granularities and time spans. Given a temporal knowledge graph G, we represent it as an ordered sequence of temporal snapshots G1, G2,..., GT, where each snapshot Gi contains all temporal facts associated with the i-th timestamp in chronological order. We encode the temporal position of snapshot Gi using sinusoidal positional encodings:

[TE(i)]2n = sin(ωn i), [TE(i)]2n+1 = cos(ωn i), ωn = β(−2n/d)

where d denotes the temporal embedding dimension and β controls the frequency scale. Unlike learned timestamp embeddings, this representation is independent of dataset-specific timestamp identities and depends only on relative temporal ordering. As a result, it naturally supports transfer across temporal knowledge graphs with different timestamp vocabularies, temporal granularities, and time spans. Furthermore, the multi-frequency sinusoidal representation enables the model to capture both short-term and long-range temporal dependencies.

5.2 Temporal-Aware Quadruple Propagation

Temporal relations exhibit substantially different dynamics: some evolve slowly over long time spans (e.g., diplomatic alliances), while others are highly localized and depend primarily on recent events (e.g., daily conflict responses). FITTER therefore injects temporal information via two complementary graphs: a global entity graph over the full TKG for slow-changing relations, and a local entity graph restricted to a window around the query timestamp for fast-changing ones.

The global entity graph performs message passing over the entire temporal knowledge graph, capturing long-range temporal dependencies. For a temporal query (s, p, ?, τi), the initial entity representation is: e0vs = 1v=s · rp, where rp is the query-conditioned relation representation from the relational backbone. At each message-passing layer, FITTER injects temporal information into propagation:

el+1vs = AGG(T-MSG(elw, rq, g l+1(TEτj)) (ew, q, v, τj) ∈ G, w ∈ Nq(v), q ∈ R)

where TEτj is the temporal encoding of timestamp τj and g l+1(·) is a learnable linear transformation per layer. T-MSG(·) denotes the temporal message function: it computes the message propagated along an edge by applying a temporal scoring function, such as T(NT)ComplEx, to the neighbor representation elw, the query-relation representation rq, and the injected temporal encoding.

The local entity graph restricts propagation to a temporal window centered around the query timestamp: Glocal = Gi−k,..., Gi,..., Gi+k, where k controls the window size. The final entity representation combines local and global contexts:

Vs,p,o = α eos,τi,local + (1 − α) eos

where α balances short-term and long-range temporal information. The quadruple score is then computed as: S(s, p, o, τi) = fθ(Vs,p,o, TEτi), where fθ is a multilayer perceptron.

5.3 Loss Function

FITTER trains the model using a binary cross-entropy objective with negative sampling. For each positive quadruple (s, p, o, τ), negative samples are generated by corrupting the head or tail entity. The training objective is:

L = −Σ [y log σ(S(s, p, o, τ)) + (1 − y) log(1 − σ(S(s, p, o, τ)))]

where y ∈ 0, 1 denotes the label of the quadruple and σ(·) is the sigmoid function.

Theoretical analysis

Theorem 1 (Time-Shift Invariance in Sinusoidal Positional Embeddings): For any four time-points indices τ1, τ2, τ1′, τ2′ ∈ N, if ∆τ = τ2 − τ1 = ∆τ′ = τ2′ − τ1′, then TE(τ2) − TE(τ1) = TE(τ2′) − TE(τ1′). That is, the Euclidean distance between two sinusoidal embeddings depends only on the time difference, not on the absolute timestamps.

Theorem 2 (Multi-Frequency Affine Scorer): FITTER's scorer can capture and express various frequencies even if fθ is simply a linear function (with m linear nodes for representing m different frequencies). The theorem establishes conditions under which the scorer can be Pi-periodic, non-constant, and interpolate on the basic window.

Theorem 3 (Universal Compatibility under harmonically-aligned frequencies): When frequencies are chosen as ωn = 2πkn/L where L = lcm(P1,..., Pm), the compatibility system is always solvable, regardless of the choice of any target sequences gi.

6. Experimental Setup

Datasets: Six widely-used TKG benchmark datasets that differ substantially in domain, temporal granularity, and time span: ICEWS14, ICEWS05-15, GDELT, ICEWS18, YAGO, and WIKI. These datasets cover diverse domains (news events, human behavior events, commonsense facts), temporal granularities (15-minute intervals, daily, yearly), and time spans (1 month to 189 years).

Experimental protocol: For each cross-domain scenario, the model is trained on Gtrain (the source TKG training split), with the best checkpoint selected on the source TKG validation split. The trained model is then applied in a fully-inductive manner to the target TKG Ginf, whose facts are partitioned as Qinf = Qmsg ∪ Qvalid ∪ Qtest. Qmsg corresponds to the target TKG training split and is used solely for inference-time message passing; no gradient updates are performed on it. Qvalid is used to tune inference-time hyperparameters such as the local window size k and fusion weight α. Qtest yields the final reported results.

Baselines: For fully-inductive cross-domain transfer, FITTER is compared against INGRAM and ULTRA, two strong fully-inductive reasoning models originally developed for static knowledge graphs. For transductive in-domain evaluation, FITTER is compared against TComplEx, TNTComplEx, TRIX, ULTRA, and GraphOracle.

Evaluation Metrics: Mean Reciprocal Rank (MRR), Hits@1, and Hits@10, using a time-aware filtering strategy.

7. Experimental Result

7.1 Cross-Domain Transfer Performance

FITTER consistently and substantially outperforms all inductive baselines across all 15 transfer scenarios. Key takeaways:

Takeaway 1: FITTER generalizes well to datasets with varying temporal granularities and spans. The time span ranges from just 1 month (GDELT) to 189 years (YAGO), while temporal granularities vary from 15 minutes (GDELT) to 1 year (YAGO). Despite being trained on one dataset and evaluated on another with significantly different temporal characteristics, FITTER consistently achieves strong performance across all scenarios.

Takeaway 2: FITTER generalizes well across datasets from different domains and densities. The experimental datasets cover a wide range of domains, from encyclopedic knowledge in YAGO to diplomatic event data in ICEWS. FITTER achieves impressive results even when trained on one domain and evaluated on another.

Takeaway 3: FITTER generalizes well to both temporal knowledge graph interpolation and extrapolation tasks. FITTER achieves strong results when trained on the interpolation task setting and tested on the extrapolation settings.

When compared with LLM-based temporal reasoning methods on extrapolation datasets, FITTER achieves higher MRR and Hits@10 on both ICEWS18 and YAGO despite being a purely structure-driven model with no access to textual descriptions or language model priors.

7.2 Transductive Competitiveness

FITTER outperforms both TRIX and ULTRA across all datasets on MRR in the standard supervised transductive setting, demonstrating that the temporal encoding in FITTER provides a clear benefit even in the transductive setting. FITTER recovers 89–96% of MRR and 97–99% of Hits@10 on ICEWS14 and ICEWS05-15 using a single fixed vocabulary-agnostic model (≈248–275K parameters). The gap is larger on GDELT (MRR recovery ≈71%), where fine-grained 15-minute timestamps and dense event structure give dataset-specific embeddings more room to specialize.

7.3 Temporal Analysis and Ablation Study

FITTER consistently outperforms ULTRA on both symmetric quadruples (where (s, p, o, τ1) implies (o, p, s, τ2)) and evolving quadruples (where (s, p, o, τ1) transitions to (s, p′, o, τ2)), demonstrating that incorporating relative temporal signals enhances the model's ability to learn and generalize temporal structural patterns.

The ablation study shows that combining both local and global context is essential. Adding temporal encoding to LQR+GQR yields substantial gains: +8.1 MRR and +10.8 H@1 on transductive, and +4.0 MRR and +8.9 H@1 on cross-domain evaluation, confirming that transferable temporal encoding contributes significantly beyond structural propagation alone.

8. Conclusion

FITTER is the first temporal KG embedding model for cross-domain transfer, requiring no dataset-specific entity, relation, or timestamp embeddings. FITTER uses sinusoidal positional encodings for granularity-agnostic temporal representation and relation interaction graphs for vocabulary-agnostic structural transfer. Across 15 cross-dataset transfer scenarios spanning six TKGs, FITTER consistently outperforms inductive baselines while remaining competitive with transductive models. Future directions include multi-source pre-training and scaling to more diverse TKG collections.

Improvements to AI Systems Based on FITTER:

1. Cross-Domain Temporal Reasoning Without Retraining: Implement FITTER's vocabulary-agnostic architecture (sinusoidal positional encodings + relation interaction graphs) so AI systems can perform link prediction on entirely new temporal knowledge graphs—with unseen entities, relations, and timestamps—immediately after training on one source domain, eliminating the need for dataset-specific fine-tuning.

2. Granularity-Agnostic Time Handling: Use FITTER's relative temporal ordering (snapshot indices with sinusoidal encodings) instead of absolute timestamps, enabling AI systems to reason over graphs with different time granularities (e.g., 15-minute intervals vs. yearly) and time spans (1 month vs. 189 years) without recalibration.

3. Dual-Scale Temporal Context Fusion: Incorporate FITTER's local-global message-passing mechanism, where AI systems simultaneously propagate information over a full temporal graph (for slow-changing relations like alliances) and a windowed subgraph around the query timestamp (for fast-changing events like conflicts), improving prediction accuracy for both short-term dependencies and long-range patterns.

4. Time-Shift Invariant Embeddings: Leverage FITTER's proven time-shift invariance (Theorem 1) to build AI systems whose temporal representations are robust to absolute time offsets, allowing them to transfer learned temporal patterns across datasets with different starting points or calendar systems.

5. Universal Compatibility for Multi-Frequency Scoring: Apply FITTER's theoretical guarantee (Theorem 3) to design AI scorers that can express arbitrary periodic patterns when frequencies are harmonically aligned, enabling systems to model complex, multi-scale temporal dynamics (e.g., daily cycles + yearly seasonality) with a simple linear scorer.

6. Cold-Start Temporal Inference: Enable AI systems to make accurate predictions on newly constructed or sparsely labeled temporal knowledge graphs by using FITTER's structure-only approach, which requires no textual descriptions or language model priors—making it deployable in domains where semantic annotations are unavailable.

7. Extrapolation-Ready Transfer: Build AI systems that train on interpolation settings (observed time range) and directly generalize to extrapolation (future timestamps) on unseen graphs, as demonstrated by FITTER's strong performance on both tasks, supporting predictive analytics across heterogeneous temporal data sources.

8. Parameter-Efficient Transfer: Adopt FITTER's compact architecture (248–275K parameters) to create AI systems that achieve 89–96% of transductive model performance while being fully inductive, reducing computational and storage overhead for multi-graph deployments.

9. Temporal Pattern Generalization: Improve AI systems' ability to learn and transfer symmetric temporal relations (e.g., (s,p,o,τ1) → (o,p,s,τ2)) and evolving relations (e.g., (s,p,o,τ1) → (s,p′,o,τ2)) across domains, as FITTER outperforms static baselines on these structural patterns.

10. Multi-Source Pre-Training Foundation: Use FITTER as a backbone for future AI systems that pre-train on multiple heterogeneous temporal knowledge graphs simultaneously, leveraging its vocabulary-agnostic design to create a universal temporal reasoner capable of zero-shot transfer to any new domain.

Temporal knowledge graphs are central to many uses of the Semantic Web, but existing completion methods assume the entities, relation names, and timestamps to be reasoned about are already known at training time, restricting each model to a single graph and vocabulary. We propose FITTER, the first fully-inductive structural model for temporal knowledge graph link prediction that supports cross-domain transfer: the inference graph may contain entirely unseen entities, relation names, and timestamps drawn from a different domain. FITTER represents each predicate by its interaction patterns with others and time through encodings of relative rather than absolute ordering; message-passing fuses local and global temporal context to produce vocabulary-agnostic embeddings. We prove the temporal encoding is time-shift invariant and evaluate FITTER on cross-domain, cross-graph transfer over six temporal knowledge graph benchmarks of diverse domains, granularities, and time spans. FITTER consistently outperforms inductive baselines without retraining, indicating that vocabulary-agnostic structural learning is a viable foundation for inference over the heterogeneous knowledge graphs of the Semantic Web.