kProtoClust: Towards Adaptive k-Prototype Clustering without Known k

Towards optimal k-prototype discovery, k-means-like algorithms give us inspirations of central samples collection, yet the unstable seed samples selection, the hypothesis of a circle-like pattern, and the unknown K are still challenges, particularly for non-predetermined data patterns. We propose an adaptive k-prototype clustering method (kProtoClust) which launches cluster exploration with a sketchy division of K clusters and finds evidence for splitting and merging. On behalf of a group of data samples, support vectors and outliers from the perspective of support vector data description are not the appropriate candidates for prototypes, while inner samples become the first candidates for instability reduction of seeds. Different from the representation of samples in traditional, we extend sample selection by encouraging fictitious samples to emphasize the representativeness of patterns. To get out of the circle-like pattern limitation, we introduce a convex decomposition-based strategy of one-cluster-multiple-prototypes in which convex hulls of varying sizes are prototypes, and accurate connection analysis makes the support of arbitrary cluster shapes possible. Inspired by geometry, the three presented strategies make kProtoClust bypassing the K dependence well with the global and local position relationship analysis for data samples. Experimental results on twelve datasets of irregular cluster shape or high dimension suggest that kProtoClust handles arbitrary cluster shapes with prominent accuracy even without the prior knowledge K.