Multi-Task Multi-Attention Graph Neural Network for Mobile Crowd Sensing Data Reconstruction and Prediction

Abstract

In recent years, the number of static or mobile sensing devices in the city has increased rapidly, which offers a new sensing paradigm, Mobile crowd sensing (MCS). However, MCS data confronts the common sparseness issue due to the limitation of human mobility and sensing costs, which is difficult to meet the data quality requirements of urban sensing applications represented by smart traffic. It is important to not only reconstruct the largely and randomly missing data but also predict the future data to enable rich applications, which is still a non-trivial task due to two major challenges: 1) error propagation, and 2) complex spatio-temporal correlations. To this end, we propose a Multi-Task Multi-Attention Graph Neural Network (MTMAG) model: on the one hand, it alleviates the error propagation using a dynamic multi-task learning framework and a transform attention block; on the other hand, it models the complex spatio-temporal correlations over a graph structure using a multi-attention module. Extensive experiments on two real-world datasets demonstrate the advantages of MTMAG over multiple state-of-the-art baselines for both data reconstruction and prediction.