FITTER: New Model Enables Cross-Domain Inference on Temporal Knowledge Graphs

A team of researchers has unveiled FITTER, an innovative structural model designed for link prediction in temporal knowledge graphs, facilitating cross-domain transfer. Unlike traditional approaches that necessitate prior knowledge of entities, relation names, and timestamps during training, FITTER adeptly manages entirely novel entities, relation names, and timestamps across various domains. The model characterizes each predicate through its interaction patterns with others and time, utilizing encodings based on relative ordering instead of absolute. By employing message-passing to integrate both local and global temporal contexts, this method yields vocabulary-agnostic embeddings, rendering the model fully inductive. The researchers demonstrate that the temporal encoding remains invariant to time shifts and assess FITTER across six benchmarks in diverse domains. The paper can be found on arXiv with the identifier 2608.10668.