Abstract

Support vector clustering (SVC) is a powerful algorithm for density-based clustering, offering advantages such as handling arbitrary cluster shapes and determining the number of classes without prior knowledge. However, its practical application is hindered by the high computational cost of the labeling phase and its limited ability to address overlapping clusters and outliers. To overcome these challenges, this article introduces a novel labeling algorithm called fuzzy support vector clustering (FuSVC). FuSVC utilizes fuzzy sets to generate atomic fuzzy clusters based on critical points of the SVC decision function, subsequently shaping final crisp clusters using the similarity of these fuzzy partitions. Through more than 300 computational experiments, FuSVC demonstrates robustness in outlier detection, effectively handles overlapping clusters, and provides membership degrees for clusters with diverse silhouettes. By addressing limitations inherent in state-of-the-art labeling rules, FuSVC significantly enhances the applicability and scalability of SVC, showcasing its potential for efficient and high-quality clustering in diverse real-world applications.