Empirical Testing of a Clustering Algorithm for Large UML Class Diagrams; an Experiment

Abstract

One important part of developing information systems is to get as much insight as possible about the problem, and possible solutions, in an early phase. To get this insight, the actors involved need good and understandable models. One popular modeling approach is UML class diagrams. A problem with UML class diagrams is that they tend to get very large when used to model large-scale commercial applications. In the absence of suitable mechanisms for complexity management, such models tend to be represented as single, interconnected diagrams. Diagrams of this kind are difficult for stakeholders to understand and maintain. There have been developed algorithms for filtering large ER diagrams, and the aim of this project has been to try if one of these algorithms can be used for filtering UML class diagrams as well. This paper describes a laboratory experiment which compares the efficiency of two different representation methods for documentation and maintenance of large data models. The representation methods compared are the ordinary UML class diagram, and the Leveled Data Model. The methods are compared using a range of performance based and perception based variables. The results show that the Leveled Data Model is not suited for modeling large generalization hierarchies. For other kinds of relations, the efficiency of the two modeling methods is the same. The participants preferred to use the ordinary UML diagram to solve the experimental tasks.