Feature Selection Based on Modified Harmony Search Algorithm

Feature selection is the pre-processing step that is widely used, especially in the field of data mining, to simplify processes that can reduce costs and computing time. Selected features can improve the best classification accuracy. In this work, a wrapper method approach is proposed using a modified harmony search. Modification is to update memory harmony using binary encoding. The coding process is adopted from the coding process of genetic algorithms for feature selection. The process of finding a new solution is done by manipulating each variable of the decision solution based on the harmony memory consideration and pitch adjustment procedures and the non-uniform mutation procedure. Evaluate its features using a support vector machine and is called a modified HS-SVM. The results showed that the proposed method has the same genetic algorithm performance for feature selection with SVM classification (GA-SVM), but has faster access time. This performance will reduce costs and computing time, especially if applied to high dimensional data. Both of these algorithms have 96.6 percent accuracy with one feature selected, and the harmony memory size is 50, and the generation size is 100.