Abstract
Numerical classification methods provide essential tools for data analysis in various fields of science. Reallocation algorithms are used for improving an a priori clustering (either a random clustering, or one obtained by another method) by changing iteratively the assignment of objects with the purpose of optimizing a cluster validity criterion. We present two new reallocation algorithms, REMOS1 and REMOS2, which optimize the silhouette width of the objects. REMOS algorithms reallocate ‘misclassified’ objects to their closest cluster; therefore, they increase the overall mean silhouette width of the classification. We compare REMOS algorithms with the recently introduced OPTSIL (Roberts 2015) method from two aspects of performance: optimization efficiency and runtime. We conclude that REMOS algorithms reach similar or higher mean silhouette widths for the final classification, while requiring much shorter computation time.
Abbreviations
- MSW
- mean silhouette width
- PAM
- partitioning around medoids