discoCVI 0.2.1
- New vignette, “Evaluating density-based clusterings with discoCVI”,
with a complete workflow: clustering with HDBSCAN, converting noise
labels, choosing a hyperparameter with DISCO, and inspecting pointwise
and noise scores (
vignette("discoCVI")).
discoCVI 0.2.0
disco_samples() and disco_score() no
longer build any n x n matrix. The minimum spanning tree of the
mutual-reachability graph is computed in compiled code (Rcpp) with
distances evaluated on the fly, and all quantities needed by DISCO (sums
and minima of density-connectivity distances per cluster) are obtained
in a single pass over the resulting single-linkage merge tree. Memory
now grows linearly with the number of points (times the number of
clusters) instead of quadratically. On synthetic benchmarks, 0.2.0 is 23
to 138 times faster than 0.1.1 (n = 500 to 4,000; 1,222 MB vs 93 MB peak
memory at n = 4,000), and 100,000 points can be scored in about 50
seconds with about 200 MB peak memory for the whole R process.extract_dc_distances_from_mst() fills the DC-distance
matrix block by block in Kruskal order instead of running a
breadth-first search from every point.compute_dc_distances() no longer builds the
mutual-reachability matrix.p_cluster() is vectorised (one matrix product instead
of nested loops).
Bug fixes
- Points that form a singleton cluster now receive a cluster score of
0, as defined in the DISCO paper (Beer et al., Section 3, “Edge cases”)
and as in scikit-learn’s
silhouette_samples(). Version
0.1.1 gave them a score of 1, which inflated disco_score()
for labelings containing singleton clusters (for example from DBSCAN
with a small minPts). Labelings without singleton clusters
are not affected.
Correctness
- Apart from the singleton fix above, all results are identical to
version 0.1.1 up to floating-point rounding. A test suite comparing
every exported function against the 0.1.1 implementation has been added
(
tests/testthat).
Dependencies
- New dependency: Rcpp (compiled code).
discoCVI 0.1.1