References · inspirations

Talks that shaped this one

The people whose work built the stack at the heart of the talk — plus a rolling list of recent clustering developments worth following.

← back to the main page

What's new in clustering

A rolling list of recent developments — papers, algorithms, libraries. Newest first.

ACM DL · DOI 10.1145/3772363.3798816

CHI EA 2026 series · paywalled · 2026

Flagged for the rolling list. Page is behind the ACM paywall; title and one-line summary to be filled in once Seth has access.

Inspirational talks

Thumbnail for Leland McInnes UMAP talk

UMAP: Uniform Manifold Approximation and Projection

Leland McInnes

The canonical UMAP talk by its creator. Explains why the knobs behave the way they do — the foundation under the n_neighbors and min_dist sweeps in The Good. Same lab also brought us HDBSCAN and EVoC.

Thumbnail for Leland McInnes HDBSCAN SciPy 2016 talk

High Quality, High Performance Clustering with HDBSCAN

Leland McInnes · SciPy 2016

The original SciPy introduction to HDBSCAN, by its other co-author. Why density-based wins on irregular real-world data, and what makes the hierarchical extension a step-change over DBSCAN. Pair with Healy's 2018 walkthrough below for the full intuition.

Thumbnail for John Healy HDBSCAN talk

HDBSCAN, Fast Density Based Clustering — the How and the Why

John Healy · PyData NYC 2018

The canonical HDBSCAN walkthrough by its co-author. Why density-based, why hierarchical, and why the noise label matters — the intuitions sitting under The Good's clustering stage.

Thumbnail for Leland McInnes Topological Techniques talk

Topological Techniques for Unsupervised Learning

Leland McInnes · PyData LA 2019

The math under the whole Tutte stack — the same topological lens that produces UMAP also produces HDBSCAN and later EVoC. Watch this once and the family resemblance between the libraries stops being a coincidence.

Thumbnail for Leland McInnes Data Mapping talk

Data Mapping for Data Exploration

Leland McInnes · PyData Seattle 2023

The pitch for treating the 2-D map itself as the EDA artifact — not the pipeline output. The framing behind the interactive datamapplot views in The Beautiful, and the reason the 100K map is the centerpiece, not a postscript.

Thumbnail for Maarten Grootendorst BERTopic talk

BERTopic: Topic Modeling with Transformers

Maarten Grootendorst

BERTopic's author walking through the design: embeddings → UMAP → HDBSCAN → c-TF-IDF. The "modern pipeline" argument in The Good traces back to this framing.

Guides & references

Foundational learning resources — the docs and courses to start with if clustering is new, or to bookmark if you're going deeper.

Clustering · scikit-learn user guide

scikit-learn documentation

The canonical Python reference. Walks every clustering algorithm scikit-learn ships — k-means, DBSCAN, OPTICS, hierarchical, spectral, mean-shift, affinity propagation, Birch, HDBSCAN — with the famous comparison-plot grid showing where each one wins and fails. The single best one-page survey of clustering in Python.

Clustering Algorithms · Google ML Crash Course

Google for Developers

Google's introductory unit on clustering. Lays out the four families — centroid-based, density-based, distribution-based, hierarchical — with concrete examples and the trade-offs between them. The clearest conceptual ladder for anyone new to the topic.

A Comprehensive Survey of Clustering Algorithms

Dongkuan Xu & Yingjie Tian · Annals of Data Science · 2015

The academic deep dive. Covers the field end-to-end — definitions, distance and similarity measures, evaluation indicators, traditional algorithms (partition, hierarchical, density-based, grid-based, model-based) and modern ones (high-dimensional, large-scale, stream). The reference to reach for when "which algorithm should I use?" needs a rigorous answer; pair with the sklearn guide for the Python equivalents.

Articles & benchmarks

clustering_arena

Dima Shulga · GitHub

Head-to-head comparison of clustering algorithms on selected real-world datasets, with ground-truth labels. The kind of benchmark that makes the case in The Bad concrete — k-means doesn't just lose by a hair, it lands on the wrong structure.

You Are Probably Using the Wrong Clustering Algorithm

Dima Shulga · Medium (Data and Beyond)

The companion essay to clustering_arena. The argument: k-means is fast and explainable, but on most real datasets it confidently assigns points to clusters that contradict the actual structure — and with no noise label, you don't see the failure.

Podcast · Learning from Machine Learning

Learning from Machine Learning

My podcast on the insights behind a career in machine learning and data science — industry practitioners on what they've learned, what the field is becoming, and how it's reshaping the world around us.