Branch detection updates #667
Conversation
|
Check out this pull request on See visual diffs & provide feedback on Jupyter Notebooks. Powered by ReviewNB |
| class BranchDetectionData(object): | ||
| """Input data for branch detection functionality. | ||
|
|
||
| Recreates and caches internal data structures from the clustering stage. | ||
|
|
There was a problem hiding this comment.
Moved BranchDetectionData to a new file so branches.py can import hdbscan_.py without cyclical imports.
| self._finite_index = get_finite_row_indices(X) | ||
| clean_data = X[self._finite_index] | ||
| finite_index = get_finite_row_indices(X) | ||
| clean_data = X[finite_index] | ||
| internal_to_raw = { | ||
| x: y for x, y in zip(range(len(self._finite_index)), self._finite_index) | ||
| x: y for x, y in zip(range(len(finite_index)), finite_index) | ||
| } | ||
| outliers = list(set(range(X.shape[0])) - set(self._finite_index)) | ||
| outliers = list(set(range(X.shape[0])) - set(finite_index)) |
There was a problem hiding this comment.
I now recover the finite index from the condensed tree, so finite_index does not have to be stored explicitly anymore. This reverts changes I made when I introduced the branch detection code.
| assert_array_almost_equal, | ||
| assert_raises, | ||
| assert_array_almost_equal |
There was a problem hiding this comment.
assert_raises gives import error on CI/CD. I replaced it with pytest.raises in all tests.
JelmerBot
force-pushed
the
dev/flasc-updates
branch
from
4dfbe7b to
95a70f3
Compare
JelmerBot
force-pushed
the
dev/flasc-updates
branch
from
95a70f3 to
3ebc8c5
Compare
|
This looks great. Let me know when you are ready to have it merged. |
|
I think this is ready now. It contains breaking name changes for the |
This PR solves #660, adding
labelsandprobabilityparameters toBranchDetector.fit()that override the input HDBSCAN object. Cases where overridden clusters form multiple connected components in the minimum spanning tree are detected. The component labels are returned as branch labels in that case. The condensed and linkage trees for those clusters are set to None, allowing scripts to detect what happened.While working on this PR, I noticed that the branching code could be simplified extensively. This PR revert some of the changes I made when I introduced the branch detection code. I also found and fixed some issues with the hierarchy simplification code that applies a persistence threshold and added a persistence threshold parameter to the clustering code.
Finally, I made small changes in
_hdbscan_boruvka.pyxto expose the computed core distances and neighbours. This allows me to use the implementation in another project.