FSTRE: Fuzzy Spatiotemporal RDF Knowledge Graph Embedding Using Uncertain Dynamic Vector Projection and Rotation

Knowledge graphs (KGs) use resource description framework (RDF) triples to model various crisp and static resources in the world. Meanwhile, knowledge embedded into vector space can imply more meanings. Much real-world information, however, is often uncertain and dynamic. Existing KG embedding (KGE) models are insufficient to deal with uncertain dynamic knowledge in vector spaces. To overcome this drawback, this article concentrates on an embedding module for the distributed representation of uncertain dynamic knowledge and proposes a strongly adaptive fuzzy spatiotemporal RDF embedding model (FSTRE). Specifically, we first propose a fine-grained fuzzy spatiotemporal RDF model, which provides the underlying representation framework for FSTRE. Then, within the complex vector space, spatial and temporal information is embedded by projection and rotation, respectively. Fine-grained fuzziness penetrates each element of the spatiotemporal five-tuples by a modal length of the anisotropic vectors. By using geometric operations as its transformation operator, FSTRE can capture the rich interaction between crisp and static knowledge and fuzzy spatiotemporal knowledge. We performed an experimental evaluation of FSTRE based on the built fuzzy spatiotemporal KG. It was shown that our FSTRE model is superior to state-of-the-art methods and can handle complex fuzzy spatiotemporal knowledge.