Teaching Learning-Based Optimization With Evolutionary Binarization Schemes for Tackling Feature Selection Problems

Machine learning techniques heavily rely on available training data in a data set. Certain features in the data can interfere with the learning process, so it is required to remove irrelevant and redundant features to build a robust training model. As such, several feature selection techniques are usually applied in a pre-processing phase to obtain the most appropriate set of features and improve the overall learning process. In this paper, a new feature selection approach is proposed based on a modified Teaching-Learning-based Optimization (TLBO) combined with four new binarization methods: the Elitist, the Elitist Roulette, the Elitist Tournament, and the Rank-based method. The influence of these binarization methods is studied and compared to other state-of-the-art techniques. The experimental results such as Shapiro-Wilk normality and Wilcoxon ranksum test show that both transfer functions and binarization approaches have a significant influence on the effectiveness of the binary TLBO. The experiments show that choosing a fitting transfer function along with a suitable binarization method has a substantial impact on the exploratory and exploitative potentials of the feature selection technique.