Developers' perception matters: machine learning to detect developer-sensitive smells

Daniel Oliveira

doi:10.1007/s10664-022-10234-2

2022-10-12

PRE

Daniel Oliveira *

Wesley K. G. Assunção

Alessandro Garcia

Baldoíno Fonseca

Márcio Ribeiro

DOI：10.1007/s10664-022-10234-2

原文链接

原文求助

摘要

En 中文

Code smells are symptoms of poor design that hamper software evolution and maintenance. Hence, code smells should be detected as early as possible to avoid software quality degradation. However, the notion of whether a design and/or implementation choice is smelly is subjective, varying for different projects and developers. In practice, developers may have different perceptions about the presence (or not) of a smell, which we call developer-sensitive smell detection. Although Machine Learning (ML) techniques are promising to detect smells, there is little knowledge regarding the accuracy of these techniques to detect developer-sensitive smells. Besides, companies may change developers frequently, and the models should adapt quickly to the preferences of new developers, i.e., using few training instances. Based on that, we present an investigation of the behavior of ML techniques in detecting developer-sensitive smells. We evaluated seven popular ML techniques based on their accuracy and efficiency for identifying 10 smell types according to individual perceptions of 63 developers, with some divergent agreement on the presence of smells. The results showed that five out of seven techniques had statistically similar behavior, being able to properly detect smells. However, the accuracy of all ML techniques was affected by developers' opinion agreement and smell types. We also observed that the detection rules generated for developers individually have more metrics than in related studies. We can conclude that code smells detection tools should consider the individual perception of each developer to reach higher accuracy. However, untrained developers or developers with high disagreement can introduce bias in the smell detection, which can be risky for overall software quality. Moreover, our findings shed light on improving the state of the art and practice for the detection of code smells, contributing to multiple stakeholders.

Keyword:

Software quality

Machine learning

Code smell detection

Code smell