An Analysis of Machine Learning Algorithms for Condensing Reverse Engineered Class Diagrams

Abstract

There is a range of techniques available to reverse engineer software designs from source code. However, these approaches generate highly detailed representations. The condensing of reverse engineered representations into more highlevel design information would enhance the understandability of reverse engineered diagrams. This paper describes an automated approach for condensing reverse engineered diagrams into diagrams that look as if they are constructed as forward designed UML models. To this end, we propose a machine learning approach. The training set of this approach consists of a set of forward designed UML class diagrams and reverse engineered class diagrams (for the same system). Based on this training set, the method 'learns' to select the key classes for inclusion in the class diagrams. In this paper, we study a set of nine classification algorithms from the machine learning community and evaluate which algorithms perform best for predicting the key classes in a class diagram.

Condensing reverse engineered class diagrams through class name based abstraction

“…Prior study [13] shows that the suitable classification algorithm for class inclusion/exclusion based on design metrics are Random Forests and Nearest Neighbor. However, when using just text metrics Logistic Regressions also provides suitable results, and when design and text metrics are combined Decision Stumps also perform surprisingly well (Table III).…”

“…In our previous work [13], we proposed an approach to simplify reverse engineered class diagrams based on static analysis. We used forward designs as ground truth for what classes are most important, and then used machine learning techniques to learn the relationship between class characteristics measured in object-oriented design metrics, and the importance of the class (in forward design or not).…”

“…Related to domain) to be key features to decide on the importance of a particular class. The previous study [13] is used as a baseline and we also follow the experimental setup for this research. We derive several text metrics from the set of class names.…”

TagCombine: Recommending Tags to Contents in Software Information Sites

“…To measure the performance of Sim TagCombine , we use top-K prediction accuracies, which follows some previous studies [21][22][23][24] . Top-K prediction accuracy is the percentage of questions in the test set where their ground truth similar questions are ranked in the top-k positions in the returned ranked lists of similar questions.…”

“…We have double checked our experiments and datasets, and still there could be errors that we have not noticed. We use 10-fold cross validation [15] to evaluate the performance of our approach, which is a standard validation approach used in many previous studies [22][23][24][25] .…”