An extensive analysis of search-based techniques for predicting defective classes

Abstract

In spite of constant planning, effective documentation and proper implementation of a software during its life cycle, many defects still occur. Various empirical studies have found that prediction models developed using software metrics can be used to predict these defects. Researchers have advocated the use of search-based techniques and their hybridized versions in literature for developing software quality prediction models. This study conducts an extensive comparison of 20 search-based techniques, 16 hybridized techniques and 17 machine-learning techniques amongst each other, to develop software defect prediction models using 17 data sets. The comparison framework used in the study is efficient as it (i) deals with the stochastic nature of the techniques (ii) provides a fair comparison amongst the techniques (iii) promotes repeatability of the study and (iv) statistically validates the results. The results of the study indicate promising ability of search-based techniques and their hybridized versions for predicting defective classes.