HDBSCAN, Fast Density Based Clustering, the How and the Why
summary
HDBSCAN becomes a practical tool by converting the naive O(n^2) approach into an O(n log n) implementation using dual‑tree Borůvka and spatial indexing (kd/cover trees) to build an MST under mutual reachability distance. The library lands in scikit-learn-contrib, exposes the condensed dendrogram and persistence‑based cluster selection, and shows n log n scaling in experiments with runtimes similar to a well‑tuned DBSCAN. Reproducible notebooks and slides are provided.
AI-written summary · about
chapters (24)
- 0:00Introduction
- 1:06Clustering algorithms overview
- 3:12Centroid-based limits
- 4:37Density-based clustering with DBSCAN
- 6:40Hierarchical density-based clustering introduction
- 7:22Why hierarchical density?
- 9:50Defining clusters via density
- 12:00Local density in HDBSCAN
- 13:46Tree formation in HDBSCAN
- 15:08Mutual reachability distance
- 16:31Pruning the dendrogram
- 17:34Cluster persistence
- 18:35Accelerating HDBSCAN with spanning trees
- 21:26Dual tree Boruvka
- 22:52Performance validation
- 24:16HDBSCAN vs others
- 26:46Why HDBSCAN beats DBSCAN
- 27:48Installation and reproducibility
- 28:12Q&A: Reachability distance
- 29:18Q&A: Triangle inequality
- 30:23Q&A: Large-scale MSTs
- 31:12Q&A: Spectral clustering
- 32:15Q&A: UMAP vs t-SNE
- 33:44Closing
related talks
35mClustering: A Guide for the Perplexed
26mUnleash Big Data Clustering
01h 20mClustering Data
25mThe Art of Clustering: The Good, The Bad and The Beautiful
47mAssessing the quality of a clustering
26mUsing Random Projections to Make Sense of High-Dimensional ...
26mMathematical Approaches to Clustering
34mParallel Efficient Clustering with Graph-Based Approximate Nearest Neighbor Search