Effective threshold estimation for filter-based feature selection
Past Pramokchon, Punpiti Piamsa-nga · 2016
For data classification, a feature subset is selected from all features by prior knowledge or determined by empirical experiments; however, it varies to contents, feature measures, and classifiers. This paper presents a filter based algorithm to select a subset of features by using outlier cut-offs of relevance between features and targeted categories. This algorithm uses the statistical techniques to effectively specify the cut-offs. Unlike traditional filter-based feature selection, our proposed method has more theoretical supports; it does not require prior knowledge. It is also a fast feature selection method because it does not require iterative empirical experiments; it does not depend on types of learning machines as well. The classification performance of the proposed feature selection algorithm on the Reuters-21578 benchmark dataset is compared with existing algorithms. It is a multi-class and highly dimensional dataset. However, the experimental results show that the proposed algorithm can select a small and effective feature subset which can be used for data classification instead of the existing algorithm.