HDBSCAN Clustering

HDBSCAN is a hierarchical density-based clustering method that extracts stable clusters of varying density from a cluster hierarchy and labels low-density cases as noise.

HDBSCAN은 군집 계층에서 밀도가 서로 다른 안정적 군집을 뽑아내고 밀도가 낮은 사례는 잡음으로 표시하는 계층적 밀도 기반 군집 방법입니다.

My data have clusters of different densities plus scattered noise; which clustering method can handle that without choosing the number of clusters?

밀도가 다른 군집과 흩어진 잡음이 섞여 있는데, 군집 수를 정하지 않고 쓸 수 있는 방법이 있나요?

HDBSCAN builds a hierarchy of density-based clusters using mutual reachability distances and selects the clusters that persist longest across density levels, subject to a minimum cluster size. It does not need the number of clusters or a single distance threshold, and cases that belong to no stable cluster are labeled as noise.

When to use HDBSCAN Clustering

• Clusters may have different densities and irregular shapes.
• The data contain noise points that should not be forced into clusters.
• You do not want to fix the number of clusters in advance.

When to use something else

• Every case must be assigned to a cluster; use k-means or a Gaussian mixture model instead.
• Clusters are roughly spherical with similar sizes and you need simple centroids; use k-means instead.

Assumptions and requirements

• A meaningful distance metric on standardized variables.
• A sensible minimum cluster size.
• Clusters are regions of higher density separated by lower density.

What the analysis reports

• Number of clusters
• Number and percentage of noise cases
• Cluster sizes and mean membership probability
• Silhouette of non-noise cases

How to report it

HDBSCAN (minimum cluster size = [m]; N = [n]) identified [k] clusters and classified [value] cases ([value]%) as noise.

Related methods

• DBSCAN Density Based Clustering
• Mean Shift Clustering
• Gaussian Mixture Model (GMM)

Run HDBSCAN Clustering on Quorant