Cost-Effective Active Sparse Urban Sensing: Adversarial Autoencoder Approach

The ever-expanding applications of mobile crowdsensing have made scalable environment sensing possible by exploiting the power of ubiquitous smart devices. Nevertheless, the implementation of sensing applications in urban scale meets serious challenges in terms of sensing costs and quality. For example, data sparsity will arise due to sensing ability and cost, and the measurements could contain noise or errors. Therefore, missing data inference with low-quality measurements is critical. To tackle these challenges, we design a low-cost crowdsensing system by missing data inference incorporating with active sensing grids selection. Specifically, an adversarial autoencoder (AAE)-based scheme is proposed for missing data inference. This model applies VAE to learn latent variables and generates full data and further utilizes the adversarial nets to play a min-max game with the autoencoder. Furthermore, an active learning-based method is designed to iteratively select sensing grids to further reduce the cost. The proposed scheme can handle large missing rate, both random and block missing patterns, and is robust against measurement noise. Comprehensive experiments based on three data sets are conducted to evaluate the effectiveness of the proposed system.