Authors:
Cristiano Sousa Melo
;
Matheus Mayron Lima da Cruz
;
Antônio Diogo Forte Martins
;
José Maria da Silva Monteiro Filho
and
Javam de Castro Machado
Affiliation:
Department of Computing, Federal University of Ceará, Fortaleza-Ceará, Brazil
Keyword(s):
Change-prone Class, Machine Learning, Deep Learning, Recurrent Algorithm, Time-series.
Abstract:
During the development and maintenance of a large software project, changes can occur due to bug fix, code refactoring, or new features. In this scenario, the prediction of change-prone classes can be very useful in guiding the development team since it can focus its efforts on these pieces of software to improve their quality and make them more flexible for future changes. A considerable number of related works uses machine learning techniques to predict change-prone classes based on different kinds of metrics. However, the related works use a standard data structure, in which each instance contains the metric values for a particular class in a specific release as independent variables. Thus, these works are ignoring the temporal dependencies between the instances. In this context, we propose two novel approaches, called Concatenated and Recurrent, using time-series in order to keep the temporal dependence between the instances to improve the performance of the predictive models. Th
e Recurrent Approach works for imbalanced datasets without the need for resampling. Our results show that the Area Under the Curve (AUC) of both proposed approaches has improved in all evaluated datasets, and they can be up to 23.6% more effective than the standard approach in state-of-art.
(More)