Unsupervised and semi-supervised clustering via density and distance-based label propagation and assignment

Abstract

Density-based clustering is capable of identifying clusters of arbitrary shapes without the need to predefine the number of clusters or their distributions. However, it suffers from varying density and parameter sensitivity. To tackle these challenges, we present the Density and Distance-Based Clustering (DDBC) algorithm, which performs clustering from the backbone to the foliage. Based on the “K_cutoff” neighborhoods of core points, DDBC constructs the cluster backbone through label propagation and subcluster aggregation. Subsequently, we construct cluster prototypes and leverage point-prototype distances to help assign points located outside the backbone. The proposed method effectively mitigates issues related to varying density. Furthermore, we propose a semi-supervised version of DDBC, termed SS-DDBC, which utilizes a few labeled data to guide label propagation and subcluster aggregation. It provides a safe and adaptive approach to leverage class information for semi-supervised clustering. Moreover, we propose automated parameter optimization approaches for DDBC and SS-DDBC, thus addressing the issue of parameter sensitivity. In both unsupervised and semi-supervised settings, we conducted experimental comparisons of DDBC and SS-DDBC with ten state-of-the-art algorithms across a range of benchmark datasets. Both algorithms consistently outperform their competitors in terms of average performance and achieve superior results on the majority of datasets. These experimental results demonstrate the effectiveness of our proposed methods. The source codes for our algorithms are accessible at https://github.com/nblnbl/DDBC.