Active learning-based classifier personalization: A case of on-body device localization
Koji Sato, Kaori Fujinami · 2017
In activity recognition, the number of target classes and classification accuracy are insufficient for practical application because accuracy declines in real situations. To address this problem, we consider that acquiring personal data is effective. We propose an active learning system for the accuracy problem and discuss an effective classifier and method to select appropriate data. We demonstrate that a Random Forest classifier using a Least Confident active learning method increases the F-measure the fastest in the combination of classifiers (Naïve Bayes, Support Vector Machine, and Random Forest) and active learning methods (Least Confident, Margin Sampling, and Entropy).