Scholarlink

<p>Topological methods have the potential of exploring the shape of datasets across different scales by analyzing topological features (connected components, holes) that persist at different spatial resolutions. Given a distance defined on a dataset, we propose a hierarchical topological clustering algorithm that creates clusters from the components of an associated one parameter family of simplicial complexes. Unlike DBSCAN, this strategy is suitable to understand multiclustering phenomena in hierarchical data and does not require blind parameter choices. Cluster configurations that persist for large parameter ranges are considered dominant. While popular methods like K-means or K-medoids fail to identify non convex clusters, the proposed algorithm creates clusters of any shape. Moreover, its stability under noise is quantifiable through distances defined on persistence diagrams, a property not enjoyed by standard hierarchical clustering. Unlike linkage strategies such as centroid or median, this algorithm can be implemented with any distance choice, not just Euclidean distances. The complexity of the algorithm is bounded from above by <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(MN^2\)</EquationSource> </InlineEquation>, <i>N</i> being the amount of data and <i>M</i> the resolution levels. Persistent clusters formed by one or a few points are candidates to be meaningful outliers, that is, outliers with a particular significance for the dataset under study (such as key genes or trade partners), rather than noise. We demonstrate through MATLAB tests the potential of the algorithm on datasets of moderate size in which multiclustering and outliers are relevant, such as swarm motion data, gene and microscope data for cancer research, compressed images, and trade volume data, using different distances (Euclidean distance, Wasserstein distance, Fermat distance).</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

Hierarchical topological clustering and meaningful outliers

  • 期刊论文

  • 发表日期：

    2026年2月7日

查看全文
  • Ana Carpio,
  • Gema Duro

摘要

Topological methods have the potential of exploring the shape of datasets across different scales by analyzing topological features (connected components, holes) that persist at different spatial resolutions. Given a distance defined on a dataset, we propose a hierarchical topological clustering algorithm that creates clusters from the components of an associated one parameter family of simplicial complexes. Unlike DBSCAN, this strategy is suitable to understand multiclustering phenomena in hierarchical data and does not require blind parameter choices. Cluster configurations that persist for large parameter ranges are considered dominant. While popular methods like K-means or K-medoids fail to identify non convex clusters, the proposed algorithm creates clusters of any shape. Moreover, its stability under noise is quantifiable through distances defined on persistence diagrams, a property not enjoyed by standard hierarchical clustering. Unlike linkage strategies such as centroid or median, this algorithm can be implemented with any distance choice, not just Euclidean distances. The complexity of the algorithm is bounded from above by \(MN^2\) , N being the amount of data and M the resolution levels. Persistent clusters formed by one or a few points are candidates to be meaningful outliers, that is, outliers with a particular significance for the dataset under study (such as key genes or trade partners), rather than noise. We demonstrate through MATLAB tests the potential of the algorithm on datasets of moderate size in which multiclustering and outliers are relevant, such as swarm motion data, gene and microscope data for cancer research, compressed images, and trade volume data, using different distances (Euclidean distance, Wasserstein distance, Fermat distance).