Wearable Big Data Pertinence Learning with Deep Spatiotemporal co-Mining

Abstract

Wearable Computers are greatly advancing big data practices, by levering their capabilities of ubiquitous big data capturing and streaming. However, one critical challenge is the amount of data to be transmitted, which consumes too much energy of the battery-constrained wearable devices. Targeting this obstacle, we propose a novel big data pertinence learning approach, which can learn and extract pertinent patterns in wearable big data for redundancy reduction. More specifically, a hybrid deep learning approach based on both Convolutional Autoencoder and Long Short-term Memory is proposed, which can mine both spatial and temporal patterns in the data for key pattern extraction. The achieved spatiotemporal co-mining ability when evaluated on a real- world motion dynamics big data application, demonstrates the attractive potential of pertinence extraction and redundancy minimization. This study is expected to greatly advance wearable big data practices.