Multi-dimensional spatial–temporal graph convolution for urban sensors imputation and enhancement

Abstract

Spatio-temporal data are critical for intelligent systems, such as smart transportation and smart cities. However, due to sensor failure or power failure, the spatiotemporal data missing tends to have a big impact on downstream tasks. Meanwhile, if sensors are scarce, some spatial positions without sensors need data enhancement through intelligent methods. Existing workarounds focus on modeling temporal information (such as time series), often ignoring spatial dependency, or modeling the spatial and temporal domain separately for imputation. In this paper, we propose a Long-term Multi-dimensional Spatial–Temporal Graph Convolution Network (LMSTGCN), which not only inductively estimates some missing data, but also achieves data augmentation of target locations. It contains a periodic temporal encoding mechanism, a gated temporal capture module, and a multi-dimensional spatial–temporal GCN module. The long-term temporal dependencies are captured by the periodic temporal encoding mechanism. The spatial and extra-short-term temporal dependencies are simultaneously modeled by the multi-dimensional GCN module, which can achieve exponential growth in the range of receptive fields. Corresponding to this module, we designed a spatiotemporal adjacency matrix construction method. It generates spatiotemporal adjacency matrices of corresponding time length as needed. The short-term dependencies in sequences are captured by the gated temporal capture module. In experimental analysis, results demonstrate that the proposed model outperforms the state-of-the-art baselines on real-world data sets.