Automatic Determination of Size for Feature Selection

摘要

In pattern recognition, feature selection is an important technique to improve the performance of classifiers designed from a limited number of training samples. So far, many algorithms have been proposed, but one of the problems is that it is hard to determine the size of a feature subset to be selected. In this paper, we discuss this issue and propose a novel method to determine the size automatically. This method is applicable to large-scale problems with features over fifty. Through some experiments with synthetic data the approach is shown to work almost optimally.