Evaluating the effectiveness of multi-level greedy modularity clustering for software architecture recovery

Abstract

Software architecture recovery approaches mainly analyze various types of dependencies among software modules to group them and reason about the high-level structural decomposition of a system. These approaches employ a variety of clustering techniques. In this paper, we present an empirical evaluation of a modularity clustering technique used for software architecture recovery. We use five open source projects as subject systems for which the ground-truth architectures were known. This dataset was previously prepared and used in an empirical study for evaluating four state-of-the-art architecture recovery approaches and their variants as well as two baseline clustering algorithms. We used the same dataset for an evaluation of multi-level greedy modularity clustering. Results showed that MGMC outperforms all the other SAR approaches in terms of accuracy and modularization quality for most of the studied systems. In addition, it scales better to very large systems for which it runs orders-of-magnitude faster than all the other algorithms.