QM4ER: Qualitative Prediction Models for Evaluating the Energy Efficiency of Code Refactoring

Code refactoring can impact the energy consumption of an Android application by restructuring its code while preserving its external behavior, thus providing a useful method for developing highly energy-efficient Android applications. However, two problems persist when utilizing existing methods to evaluate the energy efficiency of code refactoring. One is their relatively weak capability across diverse Android smartphone platforms. The other is that for refactorings causing minor code changes, it becomes easy to misjudge whether energy consumption has improved after applying the refactorings due to inaccuracies in evaluating the energy consumption of Android applications before and after the refactorings. To alleviate these problems, this paper proposes the QM4ER, Qualitative prediction Models for evaluating the Energy efficiency of code Refactoring. Unlike previous methods, QM4ER employ the differences in the values of these features before and after refactoring as input to mitigate the impact of irrelevant features on prediction results, thereby improving prediction accuracy. Furthermore, they utilize output labels of 1 and -1 to qualitatively evaluate whether code refactoring improves energy efficiency, providing a certain level of tolerance for errors in energy consumption measurements and enhancing cross-platform capability. Additionally, this paper provides detailed explanations of constructing the training dataset, selecting training algorithms, and training models. The experimental results across 10 Android applications, 11 refactoring patterns, and 3 different Android smartphone platforms, demonstrate that the proposed qualitative prediction model, constructed using Random Forest, achieves an accuracy rate that is 43.5% higher than the current optimal model at best. Moreover, the metric for measuring cross-platform capability shows a 30% improvement.