Um estudo exploratório sobre a criação de modelos de predição cruzada de defeitos apoiada por uma medida de correlação

Abstract

Defect prediction in software is a difficult task, especially when the target project is in early stage of development, because in this stage the project does not have historical data that can be used to train a prediction model. In this case, it is necessary to find projects that matches to the target project in order to use their historical data. This aproach is named cross-project prediction. This work aims to analyze an alternative method to clustering the projects. Our proposal is to group prediction models so that they can share their sets of training data. The groups were created by clustering algorithm called Basic Sequential Algorithmic Scheme (BSAS), using a measure of correlation known as Matthews Correlation Coefficient (MCC) to compute similarity models. When grouped, the training data models results to a single training set, giving rise to models cross-project prediction. The results obtained by analyzing the MCC values computed by prediction models indicate that the created models have low predictive performance, while the indicators used in similar studies indicate models with good performance. Therefore, we can concluded that the MCC correlation measure is a more robust measure of performance than other prediction models, and it contributes to the reduction of the class imbalance problem.