EMPIRICAL ASSESSMENT OF CODE SMELLS FOR PREDICTING SOFTWARE CHANGE PRONENESS

Abstract

Poor design choices called anti-patterns manifest themselves in the source code as code smells. Code smell is a synonym for bad implementation and is assumed to make maintenance tasks difficult to perform. In our previous study we validated the fact that it is possible to determine the degree of changeproneness for a class on the basis of certain code smells in an object-oriented system. The data used for the assessment was source code of Quartz, an open source job scheduler, from two versions 1.5.2 and 1.6.6. A total of 79 classes were examined and the results suggested a clear relationship between code smells and change proneness of a class. The dataset we used was very small to reach a strong conclusion so we extended our previous work by examining a dataset consisting of 4120 classes spanning 14 software systems. The dataset is created by preprocessing the class files that included removal of classes not common to both versions of the systems used. This was followed by assessment of code smells which was done on the basis of metrics. The dataset finally derived was then analyzed using Machine Learning Methods and the results suggest that code smells can classify a change prone class with a probability of .7 or more and a not change prone class with a probability of .67 or more using Multilayer Perceptron model.