Condensing Class Diagrams with Minimal Manual Labeling Cost.
Xinli Yang,David Lo,Xin Xia,Jianling Sun
DOI: https://doi.org/10.1109/compsac.2016.83
2016-01-01
Abstract:Traditionally, to better understand the design of a project, developers can reconstruct a class diagram from source code using a reverse engineering technique. However, the raw diagram is often perplexing because there are too many classes in it. Condensing the reverse engineered class diagram into a compact class diagram which contains only the important classes would enhance the understandability of the corresponding project.A number of recent works have proposed several supervised machine learning solutions that can be used for condensing reverse engineered class diagrams given a set of classes that are manually labeled as important or not. However, a challenge impacts the practicality of the proposed solutions, which is the expensive cost for manual labeling of training samples. More training samples will lead to better performance, but means higher manual labeling cost. Too much manual labeling will make the problem pointless since the aim is to automatically identify important classes.In this paper, to bridge this research gap, we propose a novel approach MCCondenser which only requires a small amount of training data but can still achieve a reasonably good performance. MCCondenser firstly selects a small proportion of all data, which are the most representative, as training data in an unsupervised way using k-means clustering. Next, it uses ensemble learning to handle the class imbalance problem so that a suitable classifier can be constructed based on the limited training data. To evaluate the performance of MCCondenser, we use datasets from nine open source projects, i.e., ArgoUML, JavaClient, JGAP, JPMC, Mars, Maze, Neuroph, Wro4J and xUML, containing a total of 2640 classes. We compare MCCondenser with two baseline approaches proposed by Thung et al., both of which are state-of-the-art approaches aimed to reduce the manual labeling cost. The experimental results show that MCCondenser can achieve an average AUC score of 0.73, which improves those of the two baselines by nearly 20% and 10% respectively.