An Application of Feature Selection Methods to Compare the Performances of Classification Algorithms

Abstract

In this study, it is aimed to determine fewer and significant variables with the help of feature selection methods among a large number of variables in the data discussed. Feature selection methods are effective methods that have great importance in statistics in recent years and provide great convenience to researchers. Depending on the technique used in the method, different numbers of variables are included in the model, but the correct classification rates may vary. In this context, being able to express the variables in a data set with a large number of variables of interest with a high classification percentage and fewer new variables makes positive contributions to issues such as time and cost. The variables in the data set discussed in this study were firstly analyzed with different feature selection methods and new data sets were created. Afterwards, these new data sets containing different numbers of variables were analyzed with different machine learning techniques and the best machine learning technique was determined. In this study, chronic kidney disease data were handled and the variables in the data set were classified with different feature selection methods. When the results of the study are examined, the highest classification rate with 99.75% was obtained from the correlation-based feature selection method, which includes the random forest and multilayer perceptron technique, and the filter method, which includes the k-nearest neighbor technique, with the same rate. The results of the study show that the percentage of correct classification obtained from this study is higher than that of other studies, when compared with other studies using the same dataset.