Data Quality Network for Spatiotemporal Forecasting
Abstract
The importance of learning from spatiotemporal data has been growing with the increasing number of data sources distributed over space. While numerous studies have been done for analyzing the spatiotemporal signals, existing models have assumed that the heterogeneous data sources in spatial domain are equally reliable over time. In this paper, we propose the novel method that infers the time-varying data quality level based on the local variations of spatiotemporal signals without explicitly assigned labels. Furthermore, we extend the formulation of the quality level by combining with graph convolutional networks to exploit the efficient architecture. Finally, we evaluate the proposed method by simulating a forecasting task with real-world climate data.
Sungyong Seo
