Collaborative learning based on centroid-distance-vector for wearable devices

Abstract

In typical application scenarios of Internet-of-Things (IoT), human activity recognition capability of a wearable device is usually limited to the body segment on which the device is worn. We believe that an efficient approach where wearable devices collaboratively learn can lead to better performance of activity recognition. In this paper, we first propose the concept of centroid distance vector to fully exploit and express the hidden knowledge of available data. Then, we introduce our centroid-distance-vector-based collaborative learning framework which consists of two key steps: data expansion with centroid distance vector and automatic model construction. In order to achieve automatic model construction, we further propose two collaborative learning methods: a centroid-distance-vector-based label propagation method (CD-LP) and a centroid-distance-vector-based knowledge distillation method (CD-KD). Experiments have shown that by using our proposed algorithms, the human activity recognition accuracy can achieve an average improvement of 5.24% compared to the state-of-the-art machine learning methods, and 5.88 % compared to the neural network with two fully-connected hidden layers.