Developing an Efficient Method for Automatic Threshold Detection Based on Hybrid Feature Selection Approach

Abstract

Dimensionality reduction is an interesting area of research in data mining. An effective way to reduce dimensions is feature selection that removes irrelevant information meanwhile helping to understand the learning model better and improving prediction accuracy. In this paper, we face a challenge of filter methods to determine number of significant features that achieves better performance since filters don’t evaluate performance based on accuracy but use certain criteria to rank features based on some scores. To handle this challenge, we proposed an effective hybrid approach for feature selection that is a filter-based method inspired by concepts of chi-square, Relief-F, and mutual information. It provides a score for each feature then specifies threshold value automatically based on dataset in use to select important subset of features used to build model, which reduces required execution time and amount of memory. Our proposed approach was analyzed empirically and theoretically to demonstrate its efficiency.