MIRROR: multi-objective refactoring recommendation via correlation analysis

Abstract

Refactoring is a critical but complex process to improve code quality by altering software structure without changing the observable behavior. Search-based approaches have been proposed to recommend refactoring solutions. However, existing works tend to leverage all the sub-attributes in an objective and ignore the relationship between the sub-attributes. Furthermore, the types of refactoring operations in the existing works can be further augmented. To this end, this paper proposes a novel approach, called MIRROR, to recommend refactoring by employing a multi-objective optimization across three objectives: (i) improving quality, (ii) removing code smell, and (iii) maximizing the similarity to refactoring history. Unlike previous works, MIRROR provides a way to further optimize attributes in each objective. To be more specific, given an objective, MIRROR investigates the possible correlations among attributes and selects those attributes with low correlations as the representation of this objective. MIRROR is evaluated on 6 real-world projects by answering 6 research questions. The experimental results demonstrate that MIRROR recommends an average of 43 solutions for each project. Furthermore, we compare MIRROR against existing tools JMove and QMove, and show that the F1 of MIRROR is 5.63% and 3.75% higher than that of JMove and QMove, demonstrating the effectiveness of MIRROR.