A comparative investigation of hierarchical clustering techniques and dissimilarity measures applied to the cell formation problem

One methodology for identifying alternative cell designs is the use of clustering algorithms coupled with dissimilarity measures. This paper investigates the performance of seven hierarchical clustering techniques (six previously developed and one developed specifically for cell formation) and eight dissimilarity measures (three well-known measures and five versions of a recently developed parametric measure) in the context of cell formation. Twenty-four data sets, at close to 200 partition levels, and ten measures of performance, are used for this purpose. The authors first identify clustering techniques and dissimilarity measures which should not be used for cell formation when binary data are involved. From the remaining clustering techniques and dissimilarity measures, they then identify clustering technique/dissimilarity measure combinations which are consistently good or poor performers when cell characteristics are observed. High internal cell cohesiveness and lwo levels of machine duplication are shown to be conflicting goals. Clustering techniques’ performance dependency on dissimilarity measures, data sets, stopping rules, and metrics are also clearly illustrated. Another result is that choice of clustering technique is more critical than choice of dissimilarity measure. However, differences among clustering techniques (due to chaining tendencies) can be sharply reduced by restricting the solution space for acceptable cell configurations.