A Survey on Tensor-Based Spatiotemporal Data Imputation: Knowledge Engineering Perspectives and Modeling Strategies

Abstract

Missing spatiotemporal data is prevalent in domains like traffic forecasting, meteorological analysis, and urban management. Tensor decomposition has emerged as a prominent technique for spatiotemporal data imputation due to its capability to model multi-dimensional interactions. Recent research demonstrates that incorporating structured prior knowledge—such as spatiotemporal correlation, spatiotemporal heterogeneity, and geographical similarity—is key to enhancing imputation performance. However, a systematic framework for integrating these priors with data-driven models remains underdeveloped. To bridge this gap, this survey adopts a knowledge engineering perspective to systematically review strategies for modeling spatiotemporal effects in tensor-based imputation. We propose a taxonomy centered on the mechanisms for representing spatiotemporal effects, which comprises three primary modeling pathways: spatiotemporal correlation, spatiotemporal heterogeneity, and geographical similarity. We further summarize representative methods in terms of their modeling mechanisms, optimization objectives, and algorithmic implementations. Applications and resources across domains like transportation, meteorology, and environmental science are also reviewed. Finally, we discuss challenges and future research directions, focusing on issues such as model generalization, uncertainty quantification, and the integration of causal mechanisms. This survey provides a comprehensive methodological foundation for modeling spatiotemporal knowledge in tensor imputation, laying the theoretical groundwork for developing more generalizable, interpretable, and geographically adaptive intelligent imputation models.