WIP: Initial commit of DBScan

[xd009642]

Looked to the k-means implementation to keep the design consistent. Comments need to be filled in and tests. Also distance is currently just euclidean distance but it's probably a good idea for there to some sort of distance trait or enum so people can pass in what distance function they want for this or other algorithms.

I'll carry on filling in the rest just figured early feedback was better than later 😄

[xd009642]

Also, the dbscan implementation in sklearn isn't one that you can fit then predict, fitting is part of the prediction stage so I've replicated that. Incremental DBScan is an extension over the original algorithm

[codecov]

Codecov Report

Merging #12 into master will decrease coverage by 2.22%.
The diff coverage is 94.44%.

@@            Coverage Diff             @@
##           master      #12      +/-   ##
==========================================
- Coverage   96.68%   94.46%   -2.23%     
==========================================
  Files           7       10       +3     
  Lines         181      271      +90     
==========================================
+ Hits          175      256      +81     
- Misses          6       15       +9

Impacted Files	Coverage Δ
linfa-clustering/examples/kmeans.rs	100% <ø> (ø)
linfa-clustering/examples/dbscan.rs	100% <100%> (ø)
linfa-clustering/src/dbscan/hyperparameters.rs	86.66% <86.66%> (ø)
linfa-clustering/src/dbscan/algorithm.rs	95.38% <95.38%> (ø)
linfa-clustering/src/k_means/algorithm.rs	90.9% <0%> (-4.05%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update ec5d491...d7f8236. Read the comment docs.

[LukeMathWalker]

Nice! I'll have a look straight away 😁

There are other interesting implementation in Python-land that we should keep in mind - in particular, HDBSCAN and pyclustering.
HDBSCAN also supports assigning a cluster to a new point after the initial fitting round - see here.

[LukeMathWalker]

It would be interesting to add a benchmark for DBSCAN as well - I believe we can do some optimisations in a couple of places, but I'd avoid starting with them before we can measure if the gain is real 👍

[xd009642]

So I realised when writing a quick benchmark that predict was a bad function name as a free function because of the public reexports, so I've currently renamed it to dbscan benchmark coming soon (basically copying the k_means one)

[LukeMathWalker]

Yeah, we can work on the naming - I would probably suggest to wrap it in a struct anyway, for ease of saving/loading as well as future extensibility. But we can figure this out once we nailed down the algorithm implementation (I think we are close 😁)

[xd009642]

Right benchmark added which should look very familiar

[xd009642]

Here's a picture of the change for 10,000 points

[LukeMathWalker]

I think we are there from an algorithmic point of view 😀

Next steps to get this ready to be merged:

• an example in the examples category;
• wrapping it in a struct for consistency/serialisation/future extensibility.

Then we are good to go 🚀

[xd009642]

I think we are there from an algorithmic point of view

I realised a mistake in my implementation of the algorithm! Only minor but the number of neighbours needs to be taken into account for each element in the search queue. I've added that and an example. Currently writing the doc comments then I'll push something

[LukeMathWalker]

Merged - thanks for all your work here @xd009642! 🙏