Application of mutual information-based sequential feature selection to ISBSG mixed data

摘要

There is still little research work focused on feature selection (FS) techniques including both categorical and continuous features in Software Development Effort Estimation (SDEE) literature. This paper addresses the problem of selecting the most relevant features from ISBSG (International Software Benchmarking Standards Group) dataset to be used in SDEE. The aim is to show the usefulness of splitting the ranked list of features provided by a mutual information-based sequential FS approach in two, regarding categorical and continuous features. These lists are later recombined according to the accuracy of a case-based reasoning model. Thus, four FS algorithms are compared using a complete dataset with 621 projects and 12 features from ISBSG. On the one hand, two algorithms just consider the relevance, while the remaining two follow the criterion of maximizing relevance and also minimizing redundancy between any independent feature and the already selected features. On the other hand, the algorithms that do not discriminate between continuous and categorical features consider just one list, whereas those that differentiate them use two lists that are later combined. As a result, the algorithms that use two lists present better performance than those algorithms that use one list. Thus, it is meaningful to consider two different lists of features so that the categorical features may be selected more frequently. We also suggest promoting the usage of Application Group, Project Elapsed Time, and First Data Base System features with preference over the more frequently used Development Type, Language Type, and Development Platform.