Scholarlink

<p>This paper introduces a <i>convex hull-based modified agglomerative clustering (hereafter HCL/HAL)</i>. Our proposed <i>HCL/HAL</i> consists two novel algorithms together, viz., Hull Centroid Linkage (hereafter <i>HCL</i>) and Hull Average Linkage (hereafter <i>HAL</i>). This twin algorithm integrates convex hull-based centroid computation to enhance clustering quality. <i>HAL</i> calculates centroids using boundary vertices of convex hulls, while <i>HCL</i> focuses on calculating the average distance between the hull vertices of two clusters, making it more efficient for clustering formation. The proposed methods were compared against standard agglomerative clustering (single, complete, average, and Ward linkage) using various internal and external metrics related to clustering quality on various diverse datasets. The results show <i>HCL/HAL</i> not only achieves superior or competitive performance with high scores for various metrices on the datasets, but it also improved clustering quality especially for external metrics. These findings suggest <i>HCL/HAL’s</i> potential for applications requiring precise cluster boundary definitions focusing on scalability in high-dimensional spaces. The code of <i>HCL/HAL</i> <i>developed by us is available at</i> <a href="https://github.com/rsrohansaha/HCL-HAL">https://github.com/rsrohansaha/HCL-HAL</a>.</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

HCL/HAL: a computational geometry-based twin approach for improving agglomerative clustering

  • 期刊论文

  • 发表日期：

    2026年7月21日

查看全文
  • Rohan Saha,
  • Barun Kr. Paul,
  • Arnab Maitra,
  • Chandan Chakraborty,
  • Kinsuk Giri

摘要

This paper introduces a convex hull-based modified agglomerative clustering (hereafter HCL/HAL). Our proposed HCL/HAL consists two novel algorithms together, viz., Hull Centroid Linkage (hereafter HCL) and Hull Average Linkage (hereafter HAL). This twin algorithm integrates convex hull-based centroid computation to enhance clustering quality. HAL calculates centroids using boundary vertices of convex hulls, while HCL focuses on calculating the average distance between the hull vertices of two clusters, making it more efficient for clustering formation. The proposed methods were compared against standard agglomerative clustering (single, complete, average, and Ward linkage) using various internal and external metrics related to clustering quality on various diverse datasets. The results show HCL/HAL not only achieves superior or competitive performance with high scores for various metrices on the datasets, but it also improved clustering quality especially for external metrics. These findings suggest HCL/HAL’s potential for applications requiring precise cluster boundary definitions focusing on scalability in high-dimensional spaces. The code of HCL/HAL developed by us is available at https://github.com/rsrohansaha/HCL-HAL.