[BUG] DBSCAN results incorrect

[cjnolet]

@daxiongshu ran our DBSCAN & k-means implementations against [1] and found that our results do not match, even for datasets as small as size 2^10.

[1] https://scikit-learn.org/stable/auto_examples/cluster/plot_cluster_comparison.html

[cjnolet]

branch-0.5 should be compared against 0.4 release.

[dantegd]

what is [1]?

[cjnolet]

Updated original comment

[cjnolet]

I ran @daxionshu's notebook against branches 0.5, 0.4, 0.3. This means this has been broken since before the refactor.

I believe the sklearn toy datasets should be tested even on the C++ side. That way when results don't match it's very clear to see which layer bugs were introduced.

[cjnolet]

@teju85, have you gotten a chance to look at this or #63 yet? It looks like a fix for this is slated for 0.5. Referencing #83 to reproduce the problem.

[teju85]

@cjnolet which of these issues against dbscan needs to be prioritized? 54, 63 or 80?

[teju85]

Also, is there a standalone python script that could repro this mismatch? (Sorry, if you have had it somewhere already!)

[cjnolet]

I’m going to go ahead and close this for now since we have discussed how the subtle differences in eps affect the results.