Hierarchical topological clustering

Topological methods have the potential of exploring data clouds without making assumptions on their the structure. Here we propose a hierarchical topological clustering algorithm that can be implemented with any distance choice. The persistence of outliers and clusters of arbitrary shape is inferred from the resulting hierarchy. We demonstrate the potential of the algorithm on selected datasets in which outliers play relevant roles, consisting of images, medical and economic data. These methods can provide meaningful clusters in situations in which other techniques fail to do so.

💡 Research Summary

The paper introduces a novel clustering framework called Hierarchical Topological Clustering (HTC) that leverages concepts from persistent homology, specifically the H₀ (connected components) of a Vietoris‑Rips filtration, to produce a fully hierarchical view of clusters and outliers without the need for extensive hyper‑parameter tuning. The authors begin by reviewing the limitations of traditional clustering methods—partition‑based algorithms such as K‑means that assume convex cluster shapes, hierarchical agglomerative schemes that depend on linkage choices, and density‑based approaches like DBSCAN that require careful selection of ε and minimum‑point thresholds. They argue that topological data analysis (TDA) offers a principled way to capture shape information across multiple scales, and they propose to use the evolution of connected components as the clustering criterion.

The methodological core consists of the following steps. First, a distance metric d is chosen (Euclidean, cosine, or any problem‑specific measure) and the full N × N distance matrix for the point cloud X = {x₁,…,x_N} is computed. The maximum pairwise distance r_max and the minimum non‑zero distance r_min define a filtration grid of M = ⌈r_max / r_min⌉ equally spaced radii r_m = m·h, where h = r_max / M. For each radius r_m, an undirected graph is built by connecting any two points whose distance is ≤ r_m; the connected components of this graph constitute the clusters at that scale. By tracking how components merge as r increases, a dendrogram‑like hierarchy is obtained. The algorithm records, for every filtration step, the cluster membership of each point, enabling the user to inspect the exact composition of clusters at any scale. The authors provide a concrete pseudo‑code (Figure 1) that iteratively updates a cluster‑link matrix L_m, performs depth‑first merging of linked clusters, and stops when a single component remains. The computational cost is bounded by O(M·N²) in the worst case, with a lower practical bound because the number of clusters typically shrinks rapidly as r grows. Memory consumption is dominated by the distance matrix, making the method most suitable for small to moderate data sets; for larger data, the authors suggest hybridizing HTC with density‑based pre‑filtering or using approximate nearest‑neighbor graphs.

A key conceptual advantage is that “persistence”—the range of r values over which a component remains separate—directly quantifies the significance of a cluster or outlier. Components that survive until large r are interpreted as meaningful outliers (singletons or small isolated groups), while early‑merging components correspond to dense regions. This provides a natural, data‑driven way to identify both clusters of arbitrary shape and rare but important observations without manually setting thresholds.

The authors validate HTC on three distinct domains. In a synthetic 2‑D example derived from a histopathology image, points represent the interface between healthy and malignant epithelial cells. HTC cleanly separates the main interface (a large, elongated cluster) from detached islands of malignant cells that have infiltrated healthy tissue. The merging order reveals how deep the malignant cells have penetrated, a nuance that K‑means, average‑link hierarchical clustering, and DBSCAN (even after extensive hyper‑parameter search) fail to capture. In an image‑compression experiment, a series of grayscale images are progressively compressed by retaining a varying number k of singular values. Using the Wasserstein‑1 distance to compare images, HTC produces a hierarchy that reflects perceptual quality loss: clusters of images with similar compression artifacts persist over wide ranges of k, while sudden splits indicate noticeable degradation. Finally, the method is applied to gene‑expression and macro‑economic indicator data sets. In both cases, HTC identifies biologically or economically meaningful subpopulations and flags a small number of samples as outliers, again without any a priori choice of cluster number or density thresholds.

The paper also discusses limitations. Since only H₀ is used, higher‑dimensional topological features such as loops or voids (captured by H₁, H₂) are ignored, potentially missing subtle structural information in data with intrinsic cycles. Moreover, the O(N²) distance matrix becomes prohibitive for very large N, and the authors acknowledge that scalability is an open issue. Future work is suggested in two directions: (i) extending the framework to incorporate higher‑order persistent homology, possibly by integrating barcode distances into the merging criterion; and (ii) employing sparse graph representations (e.g., k‑nearest‑neighbor graphs) or parallel computation to reduce memory and time footprints.

In summary, Hierarchical Topological Clustering offers a mathematically grounded, parameter‑light alternative to conventional clustering. By exploiting the natural hierarchy of connected components in a Vietoris‑Rips filtration, it simultaneously yields a dendrogram of clusters, quantifies cluster significance through persistence, and isolates outliers in a principled way. The experimental results across imaging, biomedical, and economic domains demonstrate that HTC can reveal structure that standard methods either miss or require extensive tuning to uncover. The approach enriches the toolbox of data scientists seeking robust, shape‑aware clustering, especially in contexts where outlier detection is as critical as cluster discovery.