A noise handling method for hyper surface classification

Abstract

Hyper surface classification (HSC) based on Jordan Curve Theorem is proven to be a simple and effective method to classify large datasets. Like most of classification algorithms, noise could also impact its accuracy even if the HSC algorithm limits the influence of noise in a local small region. In this paper, we propose a method that intuitively captures the primary goal of improving the accuracy of HSC when trained on noisy training datasets. The proposed method uses a separate pruning set to test whether the hyper surfaces covering few samples are assigned wrong labels due to the existence of noise. And then reassigns them appropriate labels if necessary. We compare the performance of HSC with and without the noise handling method. The promising experimental results indicate that the noise handling method can improve the accuracy of HSC when trained on noisy datasets, while keeping good performance when it is applied to datasets without noise. At the same time, it also reduces the model complexity of HSC to some extent.