An Approach for Clustering Class Coupling Metrics to Mine Object Oriented Software Components

摘要

Unsupervised learning methods such as clustering techniques are a natural choice for analyzing software quality by mining its related metrics. It is well known that clustering plays an important role in data mining tasks like in data analysis and information retrieval. In this paper, we have proposed an approach to cluster the pool of java classes based on the proximity between them. To know the proximity, coupling between each pair of classes is calculated in terms of weights using the weighted coupling measures. We modified document representations scheme as per our requirement to represent collected class coupling measures before applying k-mean clustering algorithm. In order to, reduce the dependency of k-mean clustering results efficiency on the choice of initial centroids, neighbor and link based criteria's for selecting initial k centroids have been proposed in the context of Object Oriented (OO) design artifacts i.e., classes. We demonstrate our work in detail and compare results of K-mean algorithm based on random and neighbor and link based initial centroids selection criteria's. Further the results of clustering are analyzed through purity and F-measure. It has been observed that definition of neighbor and link can be interpreted well in terms of the coupling between OO classes and produces best K-mean clustering results. Our approach of software component clustering may become an integral part of a framework to analyze and predict software quality attributes.