Improvement for Boundary-Uncertainty-Based Classifier Parameter Status Selection Method
David Ha, Yuya Tomotoshi, Masahiro Senda, Hideyuki Watanabe, Shigeru Katagiri, Miho Ohsaki · 2019
We propose an improved version of our boundary-uncertainty-based method for selecting the optimal classifier parameter status that corresponds to the optimal Bayes boundary. Our original method could accurately estimate the optimal status on various real-life tasks. However, several tasks showed improvement room for the estimation accuracy, time complexity, and stopping criterion of the method. This proposal reformalizes our original method to address these three issues. Experiments for selecting the optimal parameter status of an SVM classifier over 15 datasets show that our improved method can achieve even higher selection reliability, with a reduction of time complexity by a factor exceeding 102to 103over the presented datasets.