mlr3cluster 0.5.0
New learners
clust.gmeans: G-means clustering from the gmeans
package, which extends k-means by automatically determining the number
of clusters.clust.kmeans_rcpp: K-means clustering via
ClusterR::KMeans_rcpp() from the ClusterR package.clust.kmodes: K-modes clustering for categorical data
from the klaR package.
New measures
clust.sse_ratio: Within-cluster sum of squares divided
by total sum of squares, a normalized measure of cluster
compactness.
Other improvements
clust.cobweb now exposes the
output_debug_info parameter, matching the other RWeka-based
learners.clust.cobweb, clust.em,
clust.ff, and clust.SimpleKMeans now support
marshaling.clust.kcca now exposes simulated annealing through
classify = "simann" and the simann
parameter.clust.meanshift now predicts on new data. It runs mean
shift from each observation with the bandwidth and convergence threshold
used for training, then assigns the observation to the mode it reaches.
Previously, prediction warned and returned the training partition.clust.movMF now exposes the ids parameter,
which initializes the EM algorithm from fixed component
memberships.clust.protoclust now predicts on new data by assigning
each observation to the cluster with the nearest prototype. Previously,
prediction warned and returned the training partition. Its model now
stores the fitted object and training data in a list.clust.skmeans now exposes start for
selecting the initialization and maxchains for the
"pclust" and "LIHC" methods.clust.stdbscan now supports observation weights through
weights, matching clust.dbscan.
Bug fixes
- Clustering measures now handle character features consistently with
factor and ordered features. Distance-based measures use Gower distances
instead of treating character values as missing, while measures that
require numeric features report an informative error.
as_prediction_clust() now converts whole-numbered
partition values to integers instead of rejecting
them.clust.agnes, clust.diana,
clust.genie, and clust.hclust now validate
k against the number of training observations when
predicting. They previously used the number of rows in the new data even
though cutree() cuts the training tree.clust.ap now breaks similarity ties deterministically
when predicting, matching the assignment rule of the apcluster
package.clust.ap no longer fails to train on tasks with a
feature named m.clust.bico and clust.kcca now require
k >= 2, matching their existing training requirements.
Invalid values now fail when set rather than during training.clust.clara and clust.pam now reject
prediction when trained with stand = TRUE, because
clue::cl_predict() would otherwise assign clusters using
unstandardized new data.clust.cmeans now returns a plain probability matrix
without clue’s internal attributes.clust.cobweb, clust.ff, and
clust.xmeans now allow the seed S = 0, which
Weka accepts, matching the seed bound of clust.em and
clust.SimpleKMeans.clust.dbscan_fpc now rejects prediction when trained
with scale = TRUE. This prevents comparisons between
unscaled new data and eps, which applies to the scaled
training data.clust.hclust and clust.protoclust now
require the Minkowski power p to be greater than 0,
matching stats::dist().clust.hdbscan now requires minPts >= 2,
matching dbscan::hdbscan(). Smaller values now fail when
set rather than during training.clust.kcca now validates initcent as a
single string and documents its default as "randomcent".
Invalid values now fail when set rather than producing an obscure S4
validity error during training.clust.kcca no longer exposes the ntry and
min.size parameters, which are
flexclust::qtclust() control options that
flexclust::kcca() ignores.clust.kkmeans now resolves prediction distance ties
with "first" instead of "random", so
predictions are reproducible without setting a seed.clust.mclust now stores $assignments as an
integer vector, consistent with the other learners, instead of the
double vector returned by mclust.clust.optics, clust.som, and
clust.tclust now coerce logical features to numeric, so
tasks with only logical features train and predict instead of erroring
in the upstream packages.clust.SimpleKMeans now declares
min_density as a nonnegative double instead of a positive
integer, since Weka accepts fractional canopy densities.PredictionClust now preserves measure weights from
tasks with a weights_measure column in
$weights and as.data.table() instead of
dropping them.PredictionClust now errors during construction when a
partition label has no matching prob column, instead of
silently accepting inconsistent predictions.
mlr3cluster 0.4.1
Bug fixes
clust.agnes, clust.diana,
clust.genie, clust.hclust, and
clust.protoclust now cut the tree at the current
k when predicting, so changing k between
training and prediction takes effect.clust.ap no longer errors when affinity propagation
finds a single cluster.clust.bico now reclusters the BICO coreset with k-means
via stream::DSC_TwoStage(), so k determines
the number of predicted clusters instead of returning the raw
micro-clusters. Training warns when the coreset is too small to produce
k clusters.clust.clara now errors informatively when predicting
with metric = "jaccard" or medoids.x = FALSE
instead of failing with an obscure upstream error.clust.dbscan_fpc now errors informatively when
predicting after training with seeds = FALSE instead of
failing with an obscure upstream error.clust.flexmix now trains via
flexmix::stepFlexmix(), so nrep actually runs
repeated EM initializations and keeps the best fit instead of being
silently ignored. Setting nrep together with
cluster now errors.clust.genie, clust.kkmeans, and
clust.kproto now error informatively during training when
the task has fewer than 2 features, which the upstream implementations
cannot handle.clust.kkmeans now predicts by assigning observations to
the nearest cluster centroid in the kernel-induced feature space, since
input-space distances produced wrong assignments for nonlinear kernels.
The model is now a list with the fitted object, training data, and
per-cluster kernel statistics.clust.movMF now derives the stored
$assignments from the predict method, so predicting on the
training data yields the training assignments.clust.som now predicts and derives the stored
$assignments via kohonen::map(), so models
trained with keep.data = FALSE no longer fail at predict or
store empty assignments.clust.stdbscan now errors during training when the task
does not have exactly 3 features (two spatial coordinates and one
temporal coordinate) instead of silently using the wrong columns.clust.tclust no longer exposes the
iter.max parameter, which is deprecated in tclust 2.0 in
favor of niter1, niter2, and
nkeep.clust.wss and clust.entropy now return
NaN instead of 0 for empty predictions, which for
clust.wss silently skewed aggregated resampling
scores.PredictionClust: combined and empty prediction data now
retain the PredictionData class, so resample()
no longer errors on resampling iterations with an empty test set.PredictionClust: when the partition is derived from a
probability matrix, it now uses the cluster labels from the column names
instead of the column positions.PredictionClust: empty prob predictions now combine
with non-empty ones without error and no longer serialize a spurious
prob.V1 column via as.data.table().PredictionClust: as.data.table() no longer
drops the partition column for prob-only predictions constructed with
check = FALSE, returning NA partitions
instead.
mlr3cluster 0.4.0
New learners
clust.flexmix: Finite mixture model clustering from the
flexmix package.clust.genie: Genie hierarchical clustering from the
genieclust package.clust.kcca: K-centroids cluster analysis from the
flexclust package, supporting k-means, k-medians, spherical, Jaccard,
and extended Jaccard families.clust.movMF: Von Mises-Fisher mixture clustering from
the movMF package.clust.skmeans: Spherical k-means clustering from the
skmeans package.clust.som: Self-organizing maps from the kohonen
package.clust.stdbscan: ST-DBSCAN spatio-temporal clustering
from the stdbscan package (#83).clust.tclust: Robust trimmed clustering from the tclust
package.
New measures
clust.avg_between: Average between-cluster
distance.clust.avg_within: Average within-cluster distance.clust.davies_bouldin: Davies-Bouldin index.clust.dunn2: Alternative Dunn index using average
distances.clust.entropy: Cluster size distribution entropy.clust.pearsongamma: Pearson Gamma correlation between
distances and cluster membership.clust.wb_ratio: Within/between distance ratio.
Other improvements
- Learners no longer store the training data or dissimilarity matrix
in the model by default:
clust.agnes,
clust.diana, clust.fanny, and
clust.pam now expose keep.diss and
keep.data, clust.clara and
clust.kproto expose keep.data, and
clust.ap exposes includeSim, all initialized
to FALSE. Set the respective parameter to TRUE
to restore the previous behavior.
- Clustering quality measures
clust.ch,
clust.dunn, and clust.wss are now computed
natively instead of relying on fpc::cluster.stats(). The
fpc package is no longer a hard dependency.
clust.cobweb, clust.em,
clust.ff, clust.SimpleKMeans, and
clust.xmeans now declare the missings
property, since Weka handles missing attribute values natively.clust.diana gains the stop.at.k parameter
from cluster::diana().clust.em drops the exclusive property and
clust.MBatchKMeans drops fuzzy. Use the
prob predict type to select learners with soft
memberships.
Bug fixes
mlr3cluster is now added to
mlr_reflections$loaded_packages to fix errors when using
the package in parallel.as_prediction_clust.data.frame() no longer errors with
unused argument (with = FALSE) when given a plain
data.frame.clust.cmeans now reports a proper error message when an
invalid weights value is given instead of failing with a
type error.clust.cmeans, clust.kkmeans, and
clust.kmeans now accept a matrix of initial cluster centers
for the centers parameter, matching the upstream
functions.clust.cobweb now declares the hierarchical
property instead of partitional, and
clust.meanshift declares density instead of
partitional.clust.dbscan, clust.dbscan_fpc,
clust.hdbscan, and clust.optics now declare
the partial property instead of complete,
since these algorithms can leave observations unassigned (noise points
labeled 0).clust.featureless now returns prob
predictions whose most probable cluster matches the predicted
partition, with cluster column names consistent with the
other learners supporting the prob predict type.clust.silhouette now returns NaN instead
of 0 when all observations belong to a single cluster,
since the silhouette width is undefined for k < 2.
mlr3cluster 0.3.0
- feat: Add CLARA clustering learner
clust.clara from the
cluster package.
- feat: Add k-prototypes clustering learner
clust.kproto
from the clustMixType package.
- feat: Add spectral clustering learner
clust.specc from
the kernlab package.
- fix:
LearnerClustDBSCANfpc now correctly passes the
newdata argument in the predict method.
- fix:
LearnerClustKKMeans now correctly passes kernel
parameters via the kpar list to
kernlab::kkmeans().
- fix:
clust.silhouette measure now has the correct range
of [-1, 1].
- docs: Fix typos in measure documentation.
mlr3cluster 0.2.0
- feat:
Mlr3Error and Mlr3Warning classes
for errors and warnings.
- feat: Add protoclust learner from the protoclust package.
- feat: EM learner now supports probabilistic assignments.
- fix: Update learner parameter sets to match upstream package
changes.
- docs: Documentation improvements.
- chore: mlr3cluster now requires R 3.4.0. Following data.table’s
minimum R version.
- chore: mlr3cluster now requires mlr3 (>= 1.3.0) and mlr3misc
(>= 0.19.0).
mlr3cluster 0.1.12
- feat: Add
cluster_selection_epsilon parameter to
HDBSCAN learner and initialize minPts to 5.
- docs: Better learner example section.
mlr3cluster 0.1.11
- fix: Mclust learner no longer sets the control default with a
function not in import to stay compliant with paradox package
conventions.
mlr3cluster 0.1.10
- feat: Add BIRCH learner from the stream package.
- feat: Add BICO learner from the stream package.
mlr3cluster 0.1.9
- feat: Add DBSCAN learner from the fpc package.
- feat: Add HDBSCAN learner from the dbscan package.
- feat: Add OPTICS learner from the dbscan package.
- chore: Compatibility with upcoming paradox release.
- chore: Move to testthat3.
- refactor: General code refactoring.
mlr3cluster 0.1.8
- feat: Add new task based on
ruspini dataset.
mlr3cluster 0.1.7
- chore: Replace ‘clusterCrit’ measures with alternatives from cluster
and fpc packages.
- fix: Remove broken unloading test.
mlr3cluster 0.1.6
- feat: Add states as row names to
usarrests task.
- fix: Remove dictionary items after unloading package.
mlr3cluster 0.1.5
- feat: Add Mclust learner.
- fix: Fix error associated with new dbscan release.
mlr3cluster 0.1.4
- refactor: General code refactoring.
mlr3cluster 0.1.3
- feat: Add filter to
PredictionClust.
- fix: Small bug fixes.
- refactor: General code refactoring.
mlr3cluster 0.1.2
- feat: Add Hclust learner.
- feat: Add within sum of squares measure.
- docs: Add tests and documentation for Hclust.
- docs: Add documentation for WSS measure.
- refactor: Code factor adaptations.
mlr3cluster 0.1.1
- feat: Add eight new learners.
- feat: Add
assignments and save_assignments
fields to LearnerClust class.
mlr3cluster 0.1.0