ME-SFP: A Mixture-of-Experts-Based Approach for Software Fault Prediction
Aman Omer, Santosh Singh Rathore, Sandeep Kumar
Abstract
In the last two decades, many machine-learning-based works have been presented to build software fault prediction (SFP) models. The assessment of all these works showed that none of the machine learning classifiers could be generalized as the best-performing classifier in all prediction contexts. However, these techniques showed complementary behavior among them, which suggests combining their learning for improved performance. Various ensemble models explored in SFP have a major concern with the static weights assigned to the base learners for combining their decisions. In this article, we present a Mixture-of-Experts (MoE)-based approach named ME-SFP that uses the experts generated using a learning technique and a Gaussian mixture model as a gating function and a data partition technique. For the experimentation using the presented approach, we have shown the use of decision trees (DTs) as well as multilayer perceptrons (MLPs) as experts. We call them ME-SFP[DT] and ME-SFP[MLP]. We conduct a set of experiments on 35 publicly available software project datasets (PROMISE, JIRA, AEEEM, and Eclipse datasets) for SFP and measure the performance of built fault prediction models by using different measures such as F1-score, precision, recall, area under ROC curve (AUC), probability of false alarm, Mathews correlation coefficient, and G-means. Additionally, we perform Friedman's test and the Wilcoxon signed-rank sum test between the presented models, ensemble methods, baseline method, and individual learning techniques (DT and MLP). Results showed that ME-SFP[DT] and ME-SFP[MLP] produced improved results compared to individual techniques and ensemble methods.