Spatial Aggregation and Temporal Convolution Networks for Real-time Kriging

• DOI: 10.48550/arxiv.2109.12144
• Publication Date: 2021-01-01
• Authors: Yuankai Wu; Dingyi Zhuang; Mengying Lei; Aurélie Labbé; Lijun Sun
• Keywords: Convolution (computer science); Variogram

Spatiotemporal kriging is an important application in spatiotemporal data analysis, aiming to recover/interpolate signals for unsampled/unobserved locations based on observed signals. The principle challenge how effectively model and leverage the dependencies within data. Recently, graph neural networks (GNNs) have shown great promise tasks. However, standard GNNs often require a carefully designed adjacency matrix specific aggregation functions, which are inflexible general applications/problems. To address this issue, we present SATCN -- Spatial Aggregation Temporal Convolution Networks universal flexible framework perform various datasets without need specification. Specifically, propose novel spatial network (SAN) inspired by Principal Neighborhood Aggregation, uses multiple functions help one node gather diverse information from its neighbors. exclude unsampled nodes, masking strategy that prevents sensors sending messages their neighborhood introduced SAN. We capture temporal convolutional networks, allows our cope with of sizes. make generalizable unseen nodes even structures, employ inductive train SATCN. conduct extensive experiments three real-world datasets, including traffic speed climate recordings. Our results demonstrate superiority over traditional GNN-based models.