I think you're right (although I personally think 10k is a very large number and with low number of features clustering couple 100 samples also makes sense). Feel free to submit a pull request with the fix.

cc @Charlie-XIAO maybe.

Indeed that's my bad. Fixing that SVG is a bit complicated, maybe I'll do that directly.

Well I double checked and saw that previous "no" also points to tough luck: https://scikit-learn.org/1.4/tutorial/machine_learning_map/index.html so it's actually not my typo. I'm thinking maybe it means that MeanShift and Variational BGM models are not suitable for large number of samples? (I'm not so familiar with those algorithms though.) @adrinjalali

On second thought I actually don't think <10K samples would be tough luck... Looking at the starting point of the graph there is actually an arrow pointing to "get more data" when we have less than 50 samples.

Oh that might be true. We'd need to check the implementation and see if that's true. Also, it might be the case that with the new hardware we have, the number threshold is quite a bit higher.

TL;DR: I believe the cheat sheet is correct.

Suppose we are working with an unlabelled dataset with an unknown number of clusters.

For sample_size >= 10K, most scikit-learn clustering algorithms become computationally expensive, so the "tough luck" conclusion seems reasonable. I believe this is what the original author intended.

In the case of 50 < sample_size < 10K, I don't see any issues with using scikit-learn’s MeanShift or VBGMM. Both algorithms are designed for such scenarios. At worst, they can serve as brute-force solutions if no better options are available.

Thanks everyone for the discussions, closing.