Spatial and temporal twin-guided pattern recurrent graph network for implementing reasoning of spatiotemporal knowledge graph

The extrapolation of knowledge graphs (KGs) has been the subject of numerous studies. However, real world data often has complex spatial attributes, which makes reasoning on spatiotemporal knowledge graphs (STKGs) challenging. In response, we propose a model that captures both temporal and spatial patterns to address the challenge of predicting future facts in STKGs. The proposed spatial and temporal twin-guided pattern recurrent graph network (STTP-RGN) utilizes temporal and spatial sequences to identify cyclic and repetitive patterns in data. It performs spatiotemporal-twin encoding and temporal and spatial sequence encoding respectively, and inputs the encoded three results into three corresponding decoders to determine the evolution of entity and predicate representations in time and space. We used the YAGO10K, Wikidata40K, Opensky18K and DY-NB21K for tests on entity and predicate prediction. On YAGO10K, the model's entity prediction performance outperforms the best temporal extrapolation model RETIA by 20 %. The predicate and entity predictions on Wikidata40K have improved by 3 % and 20 %, respectively. Results for entity prediction on Opensky18K have increased by 30 %, while results for predicate prediction have improved by 1 %. The experimental results demonstrate that the model fills the gap in knowledge extrapolation on STKG.