Joint Neural Matrix Completion for Multi-Attribute Mobile Crowd Sensing

Mobile Crowd Sensing (MCS) has recently emerged as a new paradigm for urban sensing systems. However, the incomplete sensory data prevent the large-scale deployment of MCS and compromise the accuracy of sensing systems. To obtain the complete sensory data, recent studies propose matrix completion-based methods to estimate the missing entries based on a small portion of observed sensory data. Although promising, as MCS applications usually require collecting multi-attribute sensory data, simply applying the matrix completion to each attribute individually will result in a low estimation accuracy. We aim to study a novel problem of accurately estimating multi-attribute sensory data even though some attributes have a very low sampling ratio. Firstly, we design a neural spatial-temporal matrix completion (NSTMC) model to exploit both nonlinear spatial and temporal features in sensory data to improve the estimation accuracy of a single attribute. Then, based on NSTMC, to well exploit the correlations among multiple attributes and preserve the specific knowledge within each attribute, we further propose a joint spatial-temporal matrix completion (JSTMC) model, which includes a low-rank representation based temporal feature sharing strategy and a Laplacian regularizer based spatial feature sharing strategy. The extensive experimental results demonstrate that, by exploiting the correlations among multiple attributes, our JSTMC can accurately estimate the missing entries in sensory data even when the sampling ratio is 10%.