Evaluation of instance-based feature subset selection algorithm for maintainability prediction

An essential attribute of the software quality is maintainability which incurs almost 60-70% of total project cost. Since software maintainability prediction is a complicated process; estimating maintainability in the prior phases of software development lifecycle (SDLC) is advantageous. Further, it helps in building economical software and improving resource planning well in advance. Software metrics are strongly correlated with software maintainability as they help in examining the structural quality and characteristics of a software. Feature subset selection (FSS) is an important data preprocessing technique used in data mining. It involves determining a subset of notable features for building a prediction model. All software metrics are not equally relevant; hence, using all of them for predicting maintainability will significantly increase time, budget and effort. Thus, to achieve best maintainability prediction results with a particular learning algorithm, it is critical to select the most relevant features that manifest the characteristics of the software in hand which is this study are two open source software - Apache Jackrabbit and Light Weight Java Game Library (LWJGL). Our main focus has been to reduce the number of metrics using an instance-based FSS technique Relief and then use these relevant metrics to predict maintainability. It was observed that Linear Regression algorithm showed maximum increase of approximately 11% in accuracy levels with Relief FSS algorithm for both the open source software datasets. This paper will enable software developers to improve design and coding and identify the most relevant software metrics that affect software maintainability.