CSBIterKmeans: A New Clustering Algorithm Based on Quantitative Assessment of the Clustering Quality

摘要

In this paper we introduce a clustering algorithm CSBIterKmeans1 based on the well-known k-means algorithm. Our approach is based on the validation of the clustering result by combining two "antipodal" validation metrics, cluster separation and cluster compactness, to determine autonomously the "best" number of clusters and hence dispense with the number of clusters as input parameter. We report about our first results with a collection of audio features extracted from songs and discuss the performance of the algorithm with different numbers of features and objects.