Feature selection based on a data quality measure

摘要: In this paper, we present two new procedures for feature selection using a data quality measure. The first procedure is a filter method and the second is a hybrid method that combines the former method with a sequential forward selection (SFS), which is a wrapper method. Three classifiers; LDA, KNN and RPART, are used along with the wrapper method. Comparisons with the RELIEF and the usual SFS method are carried out on twelve well-known Machine Learning datasets. The experimental results show that our filter method outperforms the RELIEF, regarding both the misclassification error rate and the running time. Our hybrid method is faster than the SFS and it gives misclassification error rates quite similar to those given by the SFS.