Predicting Software Maintenance Effort by Mining Software Project Reports Using Inter-Version Validation

Abstract

Changes in the software are unavoidable due to an ever changing dynamic and active environment wherein expectations and requirements of the users tend to change rapidly. As a result, software needs to upgrade itself from its previous version to the next version in order to meet expectations of the user. The upgradation of the software is in terms of total number of Lines of Code (LOC) that might have been inserted, deleted or modified in moving from one version of software to the next. These changes are maintained in the change reports which constitute of the defect ID and defect description. Defect description describes the cause of defect which might have occurred in the previous version of the software due to which either new LOC needs to be inserted or existing LOC need to be deleted or modified. A lot of effort is required to correct the defects identified in software at the maintenance phase i.e., when software is delivered at the customers end. Thus, in this paper, we intend to predict maintenance effort by analyzing the defect reports using text mining techniques and thereafter developing the prediction models using suitable machine learning algorithms viz. Multi-Layer Perceptron (MLP), Radial-Basis Function (RBF) network and Decision Tree (DT). We have considered the changes between three successive versions of ‘MMS’ application package of Android operating system and have performed inter-version validation where the model predicted using the version ‘v’ is validated on the subsequent version i.e., ‘v+1’. The performance of the model was evaluated using Receiver Operating Characteristics (ROC) analysis. The results indicated that the model predicted on ‘MMS’ 4.0 version using MLP algorithm has shown good results when validated on ‘MMS’ 4.1 version. On the other hand, the performance of RBF and DT algorithms has been consistently average in predicting the maintenance effort.