Experience report: Evaluating the effectiveness of decision trees for detecting code smells

Developers continuously maintain software systems to adapt to new requirements and to fix bugs. Due to the complexity of maintenance tasks and the time-to-market, developers make poor implementation choices, also known as code smells. Studies indicate that code smells hinder comprehensibility, and possibly increase change- and fault-proneness. Therefore, they must be identified to enable the application of corrections. The challenge is that the inaccurate definitions of code smells make developers disagree whether a piece of code is a smell or not, consequently, making difficult creation of a universal detection solution able to recognize smells in different software projects. Several works have been proposed to identify code smells but they still report inaccurate results and use techniques that do not present to developers a comprehensive explanation how these results have been obtained. In this experimental report we study the effectiveness of the Decision Tree algorithm to recognize code smells. For this, it was applied in a dataset containing 4 open source projects and the results were compared with the manual oracle, with existing detection approaches and with other machine learning algorithms. The results showed that the approach was able to effectively learn rules for the detection of the code smells studied. The results were even better when genetic algorithms are used to pre-select the metrics to use.