LoginSign Up
← Back to all sparks
M

mlr3cluster

ANALYTICS
Velocity0.0

Cluster analysis learners and measures for the mlr3 machine learning framework

github.com/mlr-org/mlr3cluster ↗

mlr3cluster went from a handful of clusterers to covering the field

clusteringmlr3machine-learningr-statslearnersunsupervised
◆Current state
mlr3cluster supplies clustering learners to the mlr3 framework. Over three releases it added roughly a dozen learners — CLARA, k-prototypes, spectral, then a batch of nine covering finite mixtures, spherical and directional families, self-organising maps, spatio-temporal DBSCAN and robust trimmed clustering. The newest release fixes predict-time behaviour across the hierarchical learners.
◆Where it's heading
The package is at the tail end of a coverage push, and the emphasis has shifted from adding algorithms to making the ones it has behave correctly at prediction time — cutting trees at the current k, reclustering coresets, failing informatively on unsupported metric combinations. That is the normal sequence after a rapid expansion.
◆Prediction
Expect further predict-path corrections and parameter-set alignment across the newly added learners before any more algorithms arrive.

◆Recent moves

  1. 1mo ago

    Hierarchical learners now honour k at prediction time

    Fixes a class of prediction bugs across the newly added learners: trees are cut at the current k, BICO reclusters its coreset so k means what it says, and unsupported CLARA combinations error clearly. Cleanup directly attributable to the fast expansion in 0.3.0 and 0.4.0.

    View source ↗
  2. 2mo ago

    Nine new clustering learners in one release

    ⚡ SPARK

    The release that took mlr3cluster from partial to broad coverage, pulling in families the framework previously could not express at all.

    View source ↗
  3. 5mo ago

    CLARA, k-prototypes and spectral clustering learners added

    Adds three learners covering large-sample medoids, mixed-type data and spectral methods, and corrects the silhouette measure's range. The release that opened the expansion 0.4.0 completed.

    View source ↗
  4. 6mo ago

    Typed error classes and probabilistic EM assignments

    Introduces Mlr3Error and Mlr3Warning condition classes, adds the protoclust learner and gives EM probabilistic assignments. Framework plumbing alongside the learner additions.

    View source ↗
  5. 8mo ago

    HDBSCAN gains cluster_selection_epsilon

    Exposes an HDBSCAN tuning parameter and sets a sensible minPts default. Small, but it makes an existing learner tunable in the way its upstream package intends.

    View source ↗
  6. 1y ago

    Mclust learner brought in line with paradox conventions

    A convention fix in how the Mclust learner sets its control default. No user-visible effect.

    View source ↗