Ensemble Machine Learning Model for Software Defect Prediction

Abstract

Emmanuel Gbenga Dada, David Opeoluwa Oyewola, Stephen Bassi Joseph, Ali Baba Dauda

Software defect prediction is a significant activity in every software firm. It helps in producing quality software by reliable defect prediction, defect elimination, and prediction of modules that are susceptible to defect. Several researchers have proposed different software prediction approaches in the past. However, these conventional software defect predictions are prone to low classification accuracy, time-consuming, and tasking. This paper aims to develop a novel multi-model ensemble machine-learning for software defect prediction. The ensemble technique can reduce inconsistency among training and test datasets and eliminate bias in the training and testing phase of the model, thereby overcoming the downsides that have characterized the existing techniques used for the prediction of a software defect. To address these shortcomings, this paper proposes a new ensemble machine-learning model for software defect prediction using k Nearest Neighbour (kNN), Generalized Linear Model with Elastic Net Regularization (GLMNet), and Linear Discriminant Analysis (LDA) with Random Forest as base learner. Experiments were conducted using the proposed model on CM1, JM1, KC3, and PC3 datasets from the NASA PROMISE repository using the RStudio simulation tool. The ensemble technique achieved 87.69% for CM1 dataset, 81.11% for JM1 dataset, 90.70% for PC3 dataset, and 94.74% for KC3 dataset. The performance of the proposed system was compared with that of other existing techniques in literature in terms of AUC. The ensemble technique achieved 87%, which is better than the other seven state-of-the-art techniques under consideration. On average, the proposed model achieved an overall prediction accuracy of 88.56% for all datasets used for experiments. The results demonstrated that the ensemble model succeeded in effectively predicting the defects in PROMISE datasets that are notorious for their noisy features and high dimensions. This shows that ensemble machine learning is promising and the future of software defect prediction.

PDF