Software Change Prediction with Homogeneous Ensemble Learners on Large Scale Open-Source Systems

Abstract

Customizability, extensive community support and ease of availability have led to the popularity of Open-Source Software (OSS) systems. However, maintenance of these systems is a challenge especially as they become considerably large and complex with time. One possible method of ensuring effective quality in large scale OSS is the adoption of software change prediction models. These models aid in identifying change-prone parts in the early stages of software development, which can then be effectively managed by software practitioners. This study extensively evaluates eight Homogeneous Ensemble Learners (HEL) for developing software change prediction models on five large scale OSS datasets. HEL, which integrate the outputs of several learners of the same type are known to generate improved results than other non-ensemble classifiers. The study also statistically compares the results of the models developed by HEL with ten nonensemble classifiers. We further assess the change in performance of HEL for developing software change prediction models by substituting their default base learners with other classifiers. The results of the study support the use of HEL for developing software change prediction models and indicate Random Forest as the best HEL for the purpose.

Machine Learning for Change-Prone Class Prediction: A History-Based Approach

“…The first one is based on the number of Source Lines of Code (SLOC). It compares the SLOC between current and previous releases and considers that a class changed if the values are different (Malhotra and Khanna, 2019;Arisholm et al, 2004;Kaur and Jain, 2017;Khanna et al, 2021;Koru and Tian, 2005;Lu et al, 2012;Malhotra and Khanna, 2021;Martins et al, 2020;Zhou et al, 2009).…”

An assessment of heterogenous ensemble classifiers for analyzing change‐proneness in open‐source software systems

“…Contrary to this, a heterogenous ensemble uses the entire original training dataset for each base model 15 . Although, homogeneous ensembles can easily aggregate a large number of base models, 16,17 they can exploit the potential of only one classification algorithm. Thus, we may have to evaluate and choose the most appropriate base learning classifier for creating a successful homogeneous ensemble 17 .…”

“…Several other studies 3,35,36 have validated the successful use of ensemble learning algorithms such as Bagging, AdaBoost, LogitBoost, Logit Model Trees and Random Forests. A recent study by Khanna et al 16 assessed the favorability of eight homogeneous ensemble learners by comparing them with 10 learners belonging to other diverse families on five large OSS systems. However, all the studies have only exploited the use of homogeneous ensembles for the task of SCP.…”