Extremely Imbalanced Smell-based Defect Prediction

Abstract: In continuous integration/continuous delivery, one of the main requirements for high-speed delivery of software is to find bugs efficiently. For this reason, multiple solutions were introduced in the literature. For instance, defect prediction approaches based on bad code smells detected in modules from each version of the software. Nevertheless, these approaches do not consider the problem where there may exist an extremely higher percentage of non-defective modules compared to defective modules. Given that, each version of the software may only have a small number of defects. As a result, in this thesis, we introduce a new model with an autoencoder algorithm that uses design and implementation smells to detect defective modules. Therefore, we trained five autoencoders with distinct architectures. Ad- ditionally, for evaluation, we compared each model against autoencoders with the same architecture, trained with traditional object-oriented metrics and the combination of both. Our analysis did not show promising results, as the use of only smells and the combination of features did not provide an improve- ment compared with the use of metrics. However, we introduce a starting point for smell-based defect prediction in the context of dataset imbalance. Furthermore, we introduce a baseline for future work.

Dataset Description:

We provide three datasets. The first results from the extraction of traditional object-oriented metrics (metric.csv). The second results from the extraction of design and implementation smells (smell.csv). The third is the combination of all the features (metricsmell.csv). Moreover, these features were extracted from Designite and Bugsdorjar software archives.