Modular network models for class dependencies in software

Software clustering algorithms can automatically decompose a software system into modules by analyzing the network of dependencies between its components (e.g., classes in object-oriented systems). Empirical evaluation of these algorithms is difficult because few software systems have reference decompositions to be compared with the decompositions found by the algorithms. Alternatively, the algorithms can be evaluated by applying them on computer-generated networks with built-in decompositions, but the validity of this approach depends on the similarity between real and computer-generated networks. In this paper we present three network models and show that, with a proper choice of parameters, they can generate networks that are indistinguishable from class dependency networks.