MTSTI: A Multi-task Learning Framework for Spatiotemporal Imputation

Yakun Chen, Kaize Shi, Xianzhi Wang, Guandong Xu

Abstract: Spatiotemporal data analysis is crucial for various fields of applications, such as transportation, healthcare, and meteorology. Spatiotemporal data collected in the real world often contain missing values due to sensor failures or transmission loss. Therefore, spatiotemporal imputation aims to fill in the missing values by leveraging the underlying spatial and temporal dependencies in the partially observed data. Previous models for spatiotemporal imputation focus solely on the imputation task as a preparatory step for solving the downstream tasks. Instead, we aim to use downstream tasks to reinforce spatiotemporal imputation and further propose a multi-task learning framework, MTSTI, for spatiotemporal imputation. Our proposed framework utilizes a graph neural network to learn spatiotemporal representations via message-passing. The multi-task learning structure, combining spatiotemporal imputation with the forecasting task, provides additional insights that enhance the model’s performance and generality. Our empirical results demonstrate that our proposed framework outperforms state-of-the-art methods in the imputation task on various real-world datasets across different fields.

External IDs:doi:10.1007/978-3-031-46677-9_13