Scholarlink

<p>Density-based clustering algorithms can recognize clusters of arbitrary shapes and efficiently deal with noisy points through local density analysis. However, density-based clustering algorithms tend to necessitate substantial tuning of multiple parameters and incur high computational costs due to their time complexity in nearest neighbor searches. To address these limitations, we propose an adaptive representative point-driven granular-ball clustering algorithm (ARP-GB). The algorithm initiates with local density estimation using k-nearest neighbors (KNN) distances, enabling adaptive selection of representative points across diverse density regions. Next, granular balls (GBs) are generated by assigning non-representative points to their nearest representative point based on Euclidean distance in a single step. Subsequent dynamic refinement of GB boundaries via adaptive radius reduction ensures precise cluster delineation. Then, ARP-GB redefines point membership within GBs. Finally, it dynamically performs label propagation governed by overlap rates directly on the GBs, and assigns non-representative points to their nearest representative points. By leveraging approximate nearest neighbor search and direct clustering through GB, ARP-GB achieves substantial computational efficiency improvements. Notably, it requires only a single parameter <i>k</i> for KNN density estimation, which helps reduce parameter dependency compared to conventional density-based methods. Experimental results from parameter sensitivity studies demonstrate the algorithm’s robustness against parameter variations. Experiments on synthetic and real-world datasets confirm that ARP-GB maintains competitive clustering accuracy while achieving substantial efficiency gains.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An adaptive representative point-driven granular-ball clustering algorithm

  • 期刊论文

  • 发表日期：

    2026年5月16日

查看全文
  • Zhonglin Wang,
  • Ping Zhu

摘要

Density-based clustering algorithms can recognize clusters of arbitrary shapes and efficiently deal with noisy points through local density analysis. However, density-based clustering algorithms tend to necessitate substantial tuning of multiple parameters and incur high computational costs due to their time complexity in nearest neighbor searches. To address these limitations, we propose an adaptive representative point-driven granular-ball clustering algorithm (ARP-GB). The algorithm initiates with local density estimation using k-nearest neighbors (KNN) distances, enabling adaptive selection of representative points across diverse density regions. Next, granular balls (GBs) are generated by assigning non-representative points to their nearest representative point based on Euclidean distance in a single step. Subsequent dynamic refinement of GB boundaries via adaptive radius reduction ensures precise cluster delineation. Then, ARP-GB redefines point membership within GBs. Finally, it dynamically performs label propagation governed by overlap rates directly on the GBs, and assigns non-representative points to their nearest representative points. By leveraging approximate nearest neighbor search and direct clustering through GB, ARP-GB achieves substantial computational efficiency improvements. Notably, it requires only a single parameter k for KNN density estimation, which helps reduce parameter dependency compared to conventional density-based methods. Experimental results from parameter sensitivity studies demonstrate the algorithm’s robustness against parameter variations. Experiments on synthetic and real-world datasets confirm that ARP-GB maintains competitive clustering accuracy while achieving substantial efficiency gains.