A framework for experimental evaluation of clustering techniques

Experimental evaluation of clustering techniques far component recovery is necessary in order to analyze their strengths and weaknesses in comparison to other techniques. For comparable evaluations of automatic clustering techniques, a common reference corpus of freely available systems is needed for which the actual components are known. The reference corpus is used to measure recall and precision of automatic techniques. For this measurement, a standard scheme for comparing the components recovered by a clustering technique to components in the reference corpus is required. This paper describes both the process of setting up reference corpora and ways of measuring recall and precision of automatic clustering techniques. For methods with human intervention, controlled experiments should be conducted. This paper additionally proposes a controlled experiment as a standard for evaluating manual and semi-automatic component recovery methods that can be conducted cost-effectively.