This work describes clustering methods for constructing a reliable classification of data. The underlying scheme is based on qualitative and quantitative measures of similarity in order to create classes. The methods take into consideration the objectives of the classification study that have not been dealt with by previous work. They are computationally efficient and have high classification accuracy. The basic ideas of the presented methods are simple yet effective. They use a control-generate-test strategy, which involves the generation of a set of dynamic filters (concepts) and then testing these concepts on given data vectors (components) for detecting clusters of these components. The SOM (self-organizing map) method and some well-known clustering techniques are used to show the clustering capability of these methods. In particular, simulation results from the application of these methods to computer workload characterization indicate that the conceptual clustering has a higher potential for producing meaningful categories of the workload components than the traditional clustering methods.
clustering, workload modeling, computer performance, neural networks.
to main index of abstracts.