Improved Normalized Mutual Information Feature Selection Method in Software Cost Estimation

Abstract

In this paper we conduct a comparisonstudyamongimproved normalized mutual information feature selection (INMIFS), normalized mutual information feature selection (NMIFS)and minimal-redundancy-maximal-relevance feature selection (mRMRFS)based on CBR approach in software cost estimation. In feature selection, relevance is a measurement of dependence between two features while redundancy is a measurement of the duplicated information among features in the same set. The INMIFS method applies normalized mutual information to calculate both the relevance and redundancy, and it is able to select an optimal subset of features to establish estimation model from the original set of features. We conduct the experiments with the International Software Benchmarking Standard Group (ISBSG) Release 8 data set and the Desharnais data set. The experiment results demonstrate that INMIFS method can obtain 0.49 and 0.28 in PRED (0.25) value in Desharnais and ISBSG R8 data set respectively, which means12.5% and 10.89% improvement over the NMIFS method and mRMRFS method.