Estudo empírico sobre o impacto das configurações do processo de obtenção do acoplamento conceitual

Abstract

Context: Coupling is a fundamental property with a large influence on the software evolution and maintenance. Among coupling measurements there is the Conceptual Coupling. The literature shows that is possible find many models to calculate Conceptual Coupling values that combine preprocessing (using CamelCase, steamming, etc), with differents weighting factor methods, such as TF-IDF (term frequency–inverse document frequency) and LSI (Latent Semantic Indexing). Therefore, it is necessary to establish an evaluative study on the impact of choosing the calculation model on the values obtained and software maintenance processes such as recommendations for co-changes. Objective: The aim of this work consists in comparing how pre-processing and weighting factor methods combined in different models impact the values of Conceptual Coupling, verifying whether different models result in distinct values of Conceptual Coupling. Therefore, it’s necessary to stablish an evaluation about the impact of the choice of the calculation model on the values obtained and software maintenance processes such as co-chance recomandations. Method: We selected 61 free software projects hosted on GitHub, developed in C++, Java, Javascript, Python, and Ruby programming languages . From the most recent versions of these projects (3 to 6 versions), the Conceptual Coupling values were obtained among their source code files through different combinations of textual preprocessing (CamelCase, stemming, use of code comment) and weighting factor methods (LSI, TF-IDF). The results of this process were compared by means of statistical tests in order to determine the impact of the variations of the pre-processing and weighting factor methods on the values of the event. Then, these values were used to evaluate the efficiency of the highest values of Conceptual Coupling for the recommendation of Co-Changes. Results: The results indicate that different combinations of pre-processing parameter impact in distinct values of Conceptual Coupling between souce code archives. Moreover, for the languages C++ and Java the pre-processings that result in bigger values of Conceptual Coupling are common. The same occurs between the languages Javascript, Python and Ruby. Moreover, the calculation models that result in the biggest values are those that use LSI weighting factor method, fact that is independent of the project and the programming language that it is developed. Moreover, one evidence is that the best predictions had been carried through values of Conceptual Coupling by models with TF-IDF weighting factor method, although they do not result in the largest values of Conceptual Coupling between source code files. Conclusions: From this work it was possible to evidence the existence of differences between distributions of values of Conceptual Coupling, calculated by means of distinct models. In virtue of this fact, it was possible to select the attainment models that result in the distributions of bigger values of Conceptual Coupling. These values had been used for the accomplishment of predictions of Co-Changes. Although the models that result in the biggest values of Conceptual Coupling to have as LSI weighting factor method, had been the models with TF-IDF weighting factor method the ones that had provided more efficient predictions. This sample that does not have a direct relation between distributions with bigger values of Conceptual Coupling and better predictions of Co-Changes.