Prediction of code refactoring using class and file level software metrics

Abstract

Maintenance costs as a proportion of software development cost could be very high, especially in multi version real-time systems such as in telecommunications domain due to tight coupling of modules. The readability of the code becomes very hard with the development of complex classes, so the maintenance and the adaptation of the new developers to the project becomes a diffucult job. One way to overcome this problem is to predict what parts of the system are difficult to maintain and likely to change. Refactoring decisions are taken through a costly manual inspection of the code based on developer experience. It makes the system dependent to people rather than processes. Also, manual inspection increases the cost of the project. The managers are generally interested in projects which are completed on time and within budget rather than code quality. However, it would be preferable to make the project less costly and finish it before deadline by also preserving or enhancing the code quality and structure. In this research we aim to detect the modules that need to be refactored by analyzing the code complexity of the projects with version history. We propose a machine learning based model that prioritizes attributes to predict modules to be refactored. Our prediction results revealed that assigning weights to certain attributes considerably improves the prediction performance of the model as high as 71% of probability of detection and as low as 18% of false alarm rates on the average in class-level. Further our proposed model provides on average as high as 81% efficiency in maintenance effort, over and above the manual code inspection.