A demo of HDBSCAN oversensitivity #
Interactively plots HDBSCAN clustering solutions given a ground truth of two real clusters of n_points points each (just simple Gaussians) and a variable number of noise points. You can also tweak the HDBSCAN settings live! Notice how smaller values of min_cluster_size and min_samples, while useful to detect smaller clusters, also drive up the false positive rate.
Matplotlib stuff is based on this tutorial.
Other things to note:
- The data is 5D, but only two dimensions are displayed (making it more similar to Gaia data)
- Unclustered noise points (i.e. not assigned to a cluster by HDBSCAN) are displayed as faint black points; clustered points are displayed with colours. There are only 10 colours, so it will start reusing them eventually!
- There are way fewer points involved (a few thousand by default) vs. what you see in Gaia. You can turn up n_points massively to e.g. 2000 to simulate this better and see more false positives, but... good luck to your computer =) this example is contrived to run efficiently so don't expect values of min_cluster_size and min_samples to behave the same way they do for Gaia works.
Quick example #
This is what the clustering solution should look like...
... but when n_noise_points >> n_points in each cluster (250 in this case), all hell breaks loose!
Requirements #
- Python 3.6+
- Numpy
- Matplotlib
- HDBSCAN
Running #
Tweak the settings at the top of the script if you'd like, and run it with
python3 hdbscan_issue_demo.py
If you can't get plotting to work (make sure matplotlib is using a backend like Tkinter, and displaying in its own window), you can also try running the simple demo of sliders with
python3 slider_example.py