Enhancing Software Maintainability Prediction Using Multiple Linear Regression and Predictor Importance

Abstract

Accurate maintenance effort and cost estimation are essential for effective software development. By identifying software modules with poor maintainability, Software Maintainability Prediction (SMP) plays a crucial role in managing software maintenance expenses. Previous research efforts have used multiple regression techniques to predict software maintainability, but the results regarding various accuracy and performance metrics are inconclusive. As such, developing a methodology that can recommend regression techniques for software maintainability prediction in the face of inconsistent performance or accuracy metrics is imperative. This research addresses the critical issue of software maintainability and presents a novel approach, the Software Maintainability Model (SMP) utilizing the Predictor Importance (PI) Method, Multiple Linear Regression (MLR), and five machine learning techniques. The proposed SMP integrates ten static source code metrics from object-oriented programming. MLR and PI implement feature selection, and the SMP's performance is evaluated based on accuracy and the Mean Magnitude of Relative Error (MMRE) parameters. Our findings are promising: for the User Interface Management System (UIMS) software, the proposed SMP demonstrates an impressive MMRE of 0.2441 and an accuracy of 91.91%. Similarly, for the Quality Evaluation System (QUES) software, an MMRE value of 0.2222 is achieved alongside a maximum accuracy of 80.95%. The ensemble method, when compared to other Machine Learning (ML) techniques, exhibits superior performance. These results affirm the effectiveness of our approach, contributing to the enhancement of software maintainability in object-oriented programming systems.