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

Existing temporal knowledge graph (TKG) models are typically transductive, meaning they are restricted to the specific entities, relations, and timestamps present during training. This paper addresses the challenge of cross-domain transfer: can a model trained on one TKG generalize to a completely different graph with an entirely disjoint vocabulary and different temporal characteristics without needing to be retrained?

The authors propose FITTER (Fully Inductive Time-aware Transferable Representation), a framework designed to learn structural and temporal patterns that are independent of specific dataset identifiers. FITTER introduces three key innovations:

• Relative Temporal Encoding: Instead of using absolute timestamps, it uses sinusoidal positional encodings based on snapshot indices, allowing the model to handle varying temporal granularities and time spans.
• Vocabulary-Agnostic Relational Learning: It builds on a relational backbone that represents predicates through their interaction patterns with other relations, rather than learning fixed embeddings for each relation.
• Temporal-Aware Message Passing: The model performs message passing over both global and local temporal graphs, fusing short-term and long-range temporal context to produce robust, transferable embeddings.

FITTER was evaluated across 15 cross-dataset transfer settings using six diverse TKG benchmarks, including datasets from different domains (e.g., diplomatic events vs. encyclopedic facts) and varying temporal granularities (e.g., 15-minute intervals vs. yearly data). The results demonstrate that FITTER consistently outperforms existing inductive baselines in cross-domain transfer scenarios. It effectively captures structural and temporal regularities that persist across heterogeneous graphs, proving that vocabulary-agnostic structural learning is a viable foundation for inference in the Semantic Web.