Bootstrap-CURE Clustering: An Investigation of Impact of Shrinking on Clustering Performance

Abstract

Hierarchical clustering is one of the most popular techniques in unsupervised segmentation. However, since it has quadratic complexity as it is based on pairwise distance matrix construction, it tends to be less used with really large data cases. CURE clustering tackles this challenge by accelerating the process through a first hierarchical clustering over a smaller sample from which a set of representative points of resulting clusters is obtained and used to estimate the cluster shape. A KNN process with those representative points allows completing the cluster assignment to the remaining points. This clustering technique scales the hierarchical clustering to large datasets. This work is in continuation of the earlier research, Bootstrap-CURE which uses repeated samples in the first part of the process and gains both robustness and representativeness. Also, the proposed approach uses a criterion for automatic identification of the number of clusters from a dendrogram, so that the bootstrap samples can be automatically processed. In this paper, the concept of shrinkage is proposed as a hyperparameter to the Bootstrap-CURE clustering approach. The inclusion of shrinkage brings the proposed clustering technique closer to the original CURE clustering. The impact of shrinkage on the overall performance of Bootstrap-CURE is further explored. A real-life use case from 3D printers is presented to illustrate the performance of the proposed clustering.