USTformer: Transformer Based on Multigranularity Attention for Knowledge Completion Over Uncertain Spatiotemporal Knowledge Graph

The constructed knowledge graphs (KGs) are often incomplete and the completeness of KGs directly impacts the performance of various applications and methods based on these KGs. Knowledge completion, which involves inferring and completing missing knowledge in KGs, is considered a key technology in KG research. Meanwhile, to represent and manage spatiotemporal data with uncertainty, uncertain spatiotemporal knowledge graphs (USTKGs) have received considerable attention. Unfortunately, there has been limited research on knowledge completion for USTKGs. To solve this problem, first, we investigate the USTKG embedding based on box vector, embedding USTKGs into vector spaces to obtain the vector matrix. Subsequently, we design an uncertain spatiotemporal attention mechanism capable of jointly extracting multigranularity features from the vector matrix. Finally, we propose an uncertain spatiotemporal transformer based on uncertain spatiotemporal knowledge embedding and uncertain spatiotemporal attention, to achieve uncertain spatiotemporal knowledge completion. We conducted extensive experiments on two uncertain spatiotemporal datasets, and the results demonstrate the effectiveness and scientific validity of the proposed method.