Changelog¶

1.x.x (2025-xx-xx)¶

[Python and R] A specialised (slightly faster due to better locality of reference) version of the Jarnik-Prim algorithm for computing Euclidean MSTs has been added and now it is applied automatically where applicable.
[Python and R] The mlpack package is now required.
[R] The mst family of methods has been refactored. mst.default is now equipped with the algorithm argument (default: "auto"), which allows for choosing between the Jarnik and the Dual-Tree Boruvka algorithm from mlpack.

[BACKWARD INCOMPATIBILITY] [Python and R] Inequality measures are no longer referred to as inequity measures.
[BACKWARD INCOMPATIBILITY] [Python and R] Some external cluster validity measures were renamed: adjusted_asymmetric_accuracy -> normalized_clustering_accuracy, normalized_accuracy -> normalized_pivoted_accuracy.
[BACKWARD INCOMPATIBILITY] [Python] compare_partitions2 has been removed, as compare_partitions and other partition similarity scores now support both pairs of label vectors (x, y) and confusion matrices (x=C, y=None).
[Python and R] New parameter to pair_sets_index: clipped.
In normalizing_permutation and external cluster validity measures, the input matrices can now be of the type double.
[BUGFIX] [Python] #80: Fixed adjustment for nmslib_n_neighbors in small samples.
[BUGFIX] [Python] #82: cluster_validity submodule not imported.
[BUGFIX] Some external cluster validity measures now handle NaNs better and are slightly less prone to round-off errors.

[Python] The GIc algorithm is no longer marked as experimental; its description is provided in https://doi.org/10.1007/s00357-024-09483-1.

[R] mst.default now throws an error if any element in the input matrix is missing/infinite.
[Python] The call to mlpack.emst that stopped working with the new version of mlpack has been fixed.

[Python and R] adjusted_asymmetric_accuracy now accepts confusion matrices with fewer columns than rows. Such “missing” columns are now treated as if they were filled with 0s.
[Python and R] pair_sets_index, and normalized_accuracy return the same results for non-symmetric confusion matrices and transposes thereof.

[Python and R] New function: adjusted_asymmetric_accuracy.
[Python and R] Implementations of the so-called internal cluster validity measures discussed in DOI: 10.1016/j.ins.2021.10.004; see our (GitHub-only) CVI package for R. In particular, the generalised Dunn indices are based on the code originally authored by Maciej Bartoszuk. Thanks.

Functions added (cluster_validity module): calinski_harabasz_index, dunnowa_index, generalised_dunn_index, negated_ball_hall_index, negated_davies_bouldin_index, negated_wcss_index, silhouette_index, silhouette_w_index, wcnn_index.

These cluster validity measures are discussed in more detail at https://clustering-benchmarks.gagolewski.com/.
[BACKWARD INCOMPATIBILITY] normalized_confusion_matrix now solves the maximal assignment problem instead of applying the somewhat primitive partial pivoting.
[Python and R] New function: normalizing_permutation
[R] New function: normalized_confusion_matrix.
[Python and R] New parameter to pair_sets_index: simplified.
[Python] New parameters to plots.plot_scatter: axis, title, xlabel, ylabel, xlim, ylim.

[GENERAL] A paper on the genieclust package is now available: M. Gagolewski, genieclust: Fast and robust hierarchical clustering, SoftwareX 15, 100722, 2021, DOI: 10.1016/j.softx.2021.100722.
[Python] plots.plot_scatter now uses a more accessible default palette (from R 4.0.0).
[Python and R] New function: devergottini_index.

[R] Use mlpack instead of RcppMLPACK (#72). This package is merely suggested, not dependent upon.

[BUGFIX] [Python] Add code coverage CI. Fix some minor inconsistencies. Automate the bdist build chain.
[R] Update DESCRIPTION to meet the CRAN policies.

[GENERAL] The package has been completely rewritten. The core functionality is now implemented in C++ (with OpenMP).
[GENERAL] Clustering with respect to HDBSCAN*-like mutual reachability distances is supported.
[GENERAL] The parallelised Jarnik-Prim algorithm now supports on-the-fly distance computations. Euclidean minimum spanning tree can be determined with mlpack, which is much faster in low-dimensional spaces.
[R] R version is now available.
[Python] [Experimental] The GIc algorithm proposed by Anna Cena in her 2018 PhD thesis is added.
[Python] Approximate version based on nearest neighbour graphs produced by nmslib is added.