Effects of extended features on text-based classifier for corrective maintenance

Software maintenance (SM) is a complex process and is composed of various tasks that are supported by software maintainer. Classification of maintenance request (MR) is one of the tasks in large software system, yet it is often not well classified. Classification of the MRs depends on their types, which are corrective, adaptive, perfective or preventive, which are also known as maintenance type (MT). The MTs are important in keeping the quality factors of the software system. Especially, corrective maintenances are the most requests which are released in bug tracking system (BTS) in comparison to other MTs. Corrective maintenance indicates the modification of software product after its delivery in order to correct the discovered faults, and non-corrective indicated other types of maintenance. However, classification of MT is difficult in nature and this affect maintainability of the system. A number of researches in this area are dedicated to automate methods and processes in SM in order to aid MT classification. Thus, there is a need for tools that support the maintainers in doing their daily maintenance activities more effectively. The tools should be able to classify MRs automatically without human intervention in performing MT classification. MR is textual information that can be categorized according to various features by using machine learning (ML) techniques. Title, description, error encountered, and source of request are four features for the datasets which are used to train the classifiers. The two recent features, error encountered and source of request are considerate as new features in this study. These new features are added to two previous features (title and description) which were used by Antoniol et al., (2008). The goal of this research is to increase the classification accuracy of MRs into MT by using these features and present the effect of each feature in determining MT and show the best feature among the two new features. Next, the textual information of the reported bugs in the BTS will be also considered to determine whether it is sufficient to classify the MRs into corrective or non-corrective MTs. This research also presents the results of applied combining new features in the MR classification, which affect the maintainability and other quality factors of the software system. Three different classification techniques, namely Decision Tree, Naïve Bayesian, and Logistic Regression are used as the classifier. The dataset used in the experiment are from three BTS, which are Mozilla, Eclipse and JBOSS. The dataset comprises of 1800 MRs with the corresponding features. The experimental results show that the proposed MRMT model is able to achieve higher classification accuracy. The MRMT model, which is employed two more features, namely source of request and error encountered, has also outperformed the previous works

Preview PDF FSKTM 2011 22R.pdf Download (818kB) | Preview

Actions (login required)

View Item