3DA-NTC: 3D Channel Attention Aided Neural Tensor Completion for Crowdsensing Data Inference

Mobile crowdsensing is a promising scheme for performing large-scale urban monitoring, but it always faces the issue of unstable spatiotemporal coverage, which results in the incompletion of data collection. The common solutions for tackling this issue, are to use the existing subset of measurements for inferring the remaining unsensed data by leveraging the latent data correlations. However, existing data inference techniques, both for matrix/tensor factorization based methods and deep learning based methods, cannot well capture the high-order and dynamic data correlations simultaneously under the mobile crowdsensing scheme. In this paper, we propose a novel 3Dimensional (3D) channel attention aided neural tensor completion method, called "3DA-NTC", for more accurate crowdsensing data inference, through leveraging both the multi-dimensional data structure mining ability of tensor factorization as well as the high-order, dynamic correlation learning ability of deep neural network. Specifically, to capture the spatiotemporal and multitype data correlations, we first use a 3D tensor to model the 3Order interaction among crowdsensing data. Then, we combine the traditional inner product based tensor factorization with outer product computing to enhance the modeling of nonlinear data correlations and form an interaction tensor, based on which, we apply a 3D channel attention aided convolutional neural network to further extract the features of high-order and dynamic data interactions for missing value inference. Extensive experiments on two real-world urban sensing datasets, including U-Air and SensorScope, are conducted to evaluate the performance of 3DANTC, and the results demonstrate the superiority of our method compared with the state-of-the-art (SOTA) baselines in missing data recovery.