Dynamic selection of fitness function for software change prediction using Particle Swarm Optimization

Abstract

Abstract Context Over the past few years, researchers have been actively searching for an effective classifier which correctly predicts change prone classes. Though, few researchers have ascertained the predictive capability of search-based algorithms in this domain, their effectiveness is highly dependent on the selection of an optimum fitness function. The criteria for selecting one fitness function over the other is the improved predictive capability of the developed model on the entire dataset. However, it may be the case that various subsets of instances of a dataset may give best results with a different fitness function. Objective The aim of this study is to choose the best fitness function for each instance rather than the entire dataset so as to create models which correctly ascertain the change prone nature of majority of instances. Therefore, we propose a novel framework for the adaptive selection of a dynamic optimum fitness function for each instance of the dataset, which would correctly determine its change prone nature. Method The predictive models in this study are developed using seven different fitness variants of Particle Swarm Optimization (PSO) algorithm. The proposed framework predicts the best suited fitness variant amongst the seven investigated fitness variants on the basis of structural characteristics of a corresponding instance. Results The results of the study are empirically validated on fifteen datasets collected from popular open-source software. The proposed adaptive framework was found efficient in determination of change prone classes as it yielded improved results when compared with models developed using individual fitness variants and fitness-based voting ensemble classifiers. Conclusion The performance of the models developed using the proposed adaptive framework were statistically better than the models developed using individual fitness variants of PSO algorithm and competent to models developed using machine learning ensemble classifiers.