Data Sub-sampling Method for Developing Personalized Human Activity Model Based on Incremental Learning

Abstract

This paper proposes a method to conduct data sub-sampling for developing personalized human activity recognition (HAR) model based on incremental learning. The quality of training dataset required to execute incremental learning is a critical factor affecting its performance. To generate a high-quality training dataset, this paper performs two-phase sub-sampling. The first phase discriminates the activity data into the borderline data with a high risk of mislabeling and the non-borderline data with a low risk of mislabeling. To do so, entropy-based distances between the decision boundaries and the activity data are calculated. The second phase separates the correctly labeled borderline data from the entire borderline data. This paper generates the KDE-based probability density function to discover the correctly labeled borderline data and performs binary clustering. Finally, incremental learning is performed to obtain the personalized HAR model using the non-borderline and correctly labeled borderline data. To show the superiority of the proposed method, we conduct experiments on two benchmark datasets named the HAPT and WISDM datasets.