#HDBSCAN
Agora isso virou uma função Python para scikit-learn e hdbscan. Em 10 mil linhas roda em menos de 1 segundo.

Código, explicação ilustrada e artigo no primeiro comentário.

🇺🇸 English below

You use HDBSCAN and never touch min_samples? That's not your fault.
November 9, 2026 at 9:04 PM
🇧🇷 Você usa HDBSCAN para clustering e nunca mexeu no min_samples? A culpa não é sua.

Os próprios autores do método o tratam como um fator de suavização ao qual métodos parecidos costumam ser robustos. As documentações do hdbscan e do BERTopic dão destaque a outro parâmetro.
November 9, 2026 at 9:04 PM
Código / Code: github.com/provezano/m...

No terminal / In the terminal:
pip install git+github.com/provezano/m...

No Python / In Python:
from minpts_check import check_min_samples
GitHub - provezano/minpts-check: Does your HDBSCAN* clustering survive a small change of min_samples? One-run predictor and certificate (Coutinho 2026).
Does your HDBSCAN* clustering survive a small change of min_samples? One-run predictor and certificate (Coutinho 2026). - provezano/minpts-check
github.com
November 9, 2026 at 9:04 PM
→ a mathematical condition, checked without refitting, guarantees when the next value leaves the clusters unchanged.

It is now one Python function for scikit-learn and hdbscan. On 10,000 rows it runs in under a second.

Code, illustrated guide and paper below:
November 9, 2026 at 9:04 PM
Its own authors treat it as a smoothing factor that methods are typically robust to, the hdbscan and BERTopic docs put another parameter first, and if you leave it unset the libraries quietly copy min_cluster_size into it.
November 9, 2026 at 9:04 PM
For #30DayMapChallenge day 4, “Clusters”, @fil.rezo.net ported HDBSCAN to a notebook and wrote some examples. observablehq.com/@observableh...
November 6, 2026 at 5:16 AM
check_min_samples(X, min_samples=30, min_cluster_size=30).summary()

Explicação ilustrada, com demo no navegador / Illustrated guide with an in-browser demo:
🇺🇸 dataunicorn.com.br/en/papers/t...
The shrinking cluster · Felipe Coutinho
Interactive explainer: why HDBSCAN*’s flat clustering changes when minPts goes up one step, how to predict which cluster is at risk, and how to prove that a step is safe.
dataunicorn.com.br
November 9, 2026 at 9:04 PM
Tracking Vietnamese news trends with embeddings, HDBSCAN and an LLM

How I group articles from ~95 Vietnamese news feeds into topics, score which ones are rising or fading, and have an LLM explain why.

#Embeddings #HDBSCAN #LLM
Tracking Vietnamese news trends with embeddings, HDBSCAN and an LLM
How I group articles from ~95 Vietnamese news feeds into topics, score which ones are rising or fading, and have an LLM explain why.
vuminhhoa.com
November 1, 2026 at 6:31 PM
ポルトガルの上部ジュラ系(約 1 億 5000 万年前)ギマロータ炭鉱から産出した 74 点の獣脚類恐竜の歯写真を使用。AI に合成幾何学図形を学習させる転移学習技術を用い、歯の 5 方向(表、裏、前、後ろ、断面)の画像から綺麗な二値シルエット(影絵)を自動抽出しました。
シルエットから歯の上・中・下の幅や高さ、輪郭角度など 120 個の特徴量を自動計測し、数学的処理(UMAP・HDBSCAN)を行うことで、標本群から 5 つの明確な形態グループ(細長型、低背型、極厚断面型など)を客観的に検出。
October 28, 2026 at 11:13 PM
It has been a really fun ride to work on this, and I've logged everything I learned along the way in the dsp-kitchen docs.

None of this would exist without open source: Zarr & zarrs, CubeCL, Burn, GPUI, Kilosort, EMUsort, MountainSort, SpyKING CIRCUS, Tridesclous, SpikeInterface, HDBSCAN and more!
October 26, 2026 at 5:08 PM
Method: a deliberately broad search (phrase, hashtag, ten languages, black + October + AI terms), then BERTopic: multilingual embeddings, UMAP, HDBSCAN clusters. Clusters where posts naming the event were over-represented became candidates; I confirmed them by reading.

3/3
October 23, 2026 at 7:52 AM
Photometric Identification and Dynamical Purification of Blue Horizontal Branch Stars in DESI MWS DR1. Shuai Zhang et. al. https://arxiv.org/abs/2610.15031
October 15, 2026 at 1:50 AM
I am going to be in Boston for the IEEE VIS conference Nov 9 -13 this year! I am keen to talk dimension reduction, clustering, and data maps (UMAP, HDBSCAN, DataMapPlot). If you are going to be at VIS, or are just in Boston, and want to grab a coffee, please reach out. DMs open.
September 25, 2026 at 2:14 AM
The python package is well documented and feels modern, and most of the obvious python LDAs use stochastic variational. Both BERTopic and SV LDA are good enough that they seem to work, but not really that good. The weakness in BERTopic is that embedding data doesn't really work with HDBSCAN.
September 22, 2026 at 4:57 PM
Turns out I'd archived two years of Claude Code sessions and read almost none of them, so I built a tool to see the shape of a text pile instead of just querying it: clusters, a 3D map, and a tiny model that needs bullying into not inventing its labels.
How to Map a Pile of Text Nobody Has Read
A step-by-step guide to turning any corpus into a named, explorable 3D map — embeddings, UMAP, HDBSCAN, and a 3B model that will lie to you unless you make it prove itself. With a CC0 tool, a downloadable public-domain dataset, and the story of the thesaurus I published as a Vict
nakomis.substack.com
September 22, 2026 at 9:38 AM
ML folks: I'm trying to run HDBSCAN on a dataset of 10s or 100s of millions of points. I'm considering NVIDIA's cuML implementation for GPU acceleration, vs. fast_hdbscan's parallelized CPU implementation. Do folks have any insights into runtime vs. GPU memory vs. RAM usage on specific hardware?
September 18, 2026 at 8:22 PM
Unit 42 Maps 40,000 Cloud Identities: From ML Clustering to SQL Heuristics
Unit 42 Maps 40,000 Cloud Identities: From ML Clustering to SQL Heuristics
Unit 42 researchers analyzed over 40,000 AWS identities using UMAP and HDBSCAN, extracting patterns convertible into SQL heuristics to detect threats without heavy ML pipelines.
deafnews.it
September 14, 2026 at 11:19 AM
📢 Meet the VISxAI Keynote: Leland McInnes

Leland McInnes is a researcher at the Tutte Institute for Mathematics and Computing and the author of UMAP and co-author of HDBSCAN. His work spans dimension reduction, clustering, and visual exploration of high-dimensional data.

🔗 Details: visxai.io
9th Workshop on Visualization for AI Explainability
The role of visualization in artificial intelligence (AI) gained significant attention in recent years. With the growing complexity of AI models, the critical need for understanding their inner-workin...
visxai.io
September 11, 2026 at 12:19 AM
Not sure how well it maps to neuroscience, but we use UMAP+HDBSCAN when we want to cluster data with an unknown number of subgroups. UMAP as a nonlinear approach will give you better separation of clusters, but as nonlinear is less repeatable vs PCA.
May 11, 2025 at 12:22 PM
We then found that applying iterative unsupervised clustering (using UMAP and HDBSCAN) to the acoustic features was useful for separating false-positive and true detections, and in the end you are also left with individual note annotations that have essentially been automatically tagged!
January 21, 2025 at 9:46 PM
working on an interactive fiction thing. got a 500+ notes/ideas for it. during lunch claude built a semantic search system to surface relevant stuff while i write a scene. i asked for a visualisation of the embeddings. the 2d map is much more interesting to explore than i expected❤️
January 30, 2026 at 3:54 PM