DEVELOPMENT OF SOFTWARE PREDICTION MODELS USING VARIOUS MACHINE LEARNING TECHNIQUES

Abstract

In current world, all information systems development follows well defined development process based on requirements & business needs. Every stake holder‟s expectation is to get best product with zero or minimum defects also cost effective. So, In order to achieve best product in given time, it is necessary to uncover most defects as early as possible in development life cycle. In spite of thorough planning, well documentation and proper process control during software development, occurrences of certain defects are inevitable. These defects may leads to poor software quality which may hamper the brand image & lead to business failure. In today‟s competitive world it‟s necessary to make world class product with minimum defects, high in quality & cost effective. Cost of defects finding is directly proportional to time of its finding in development cycle. Later are defects found, more is the cost of fixing them leads higher development cost. Hence it‟s necessary to identify defective classes in early phase of software development to reduce the testing cost. This may guide the product planning team for efficient resource planning for testing. Software metrics have been used to describe the complexity of the program and, to estimate software development time. Software metrics can be used in simultaneity with defect data to develop models for predicting defective classes. The development of predictive models to predict fault classes can help & guide the stakeholders in predicting faulty classes in early phase of software development. Hence, it is vital to analyse and compare the predictive accuracy of machine learning classifiers. Various Machine Learning Techniques were used to understand & analyse the core relationships of classes and fetching useful information from problems. The objective of this thesis is to evaluate the performance & comparison of Machine Learning Techniques over unpopular data sets. The evaluation is performed with an intention to identify which algorithm suits best for prediction of defect prone classes in software based on software quality metrics. Chidamber & Kemerer Java matrices [4] were generated over 4 subsequent releases of Android „Contact‟ Module. Jelly Bean to KitKat to Lollypop to Marshmallow. 7 Machine Learning techniques were vi | P a g e compared to evaluate the relationship of Chidamber & Kemerer Java matrices on defective classes. The result shows the predictive capability of Machine Learning Techniques & suggested model. The results of work were based on data sets obtained from popular open source mobile software Android “Contacts” module.