A0404
Title: Modified silhouette score for evaluating cluster solutions
Authors: Czarinne Antoinette Antonio - University of the Philippines Diliman (Philippines) [presenting]
Joseph Ryan Lansangan - University of the Philippines (Philippines)
Abstract: The assessment of the quality of a clustering solution and the proper identification of the number of clusters to be used is a crucial step in doing cluster analysis. A class of silhouette-based indices, as a modification to the widely used silhouette index, is developed to measure cluster validity. The performance of the proposed indices is demonstrated via a simulation study and through the application to actual data sets. The results revealed that the use of the second and third nearest cluster in the computation instead of just the nearest neighboring cluster relative to observation was advantageous in identifying the number of natural clusters as a viable choice in the cluster analysis. Each of the proposed indices was useful in the presence of noisy data and not well-separated clusters. Further, dimension reduction techniques employed in the calculation of the distance measures provided an added benefit when dealing with high-dimensional data.
