tidyclust 0.3.0

Deprecation

• finalize_model_tidyclust() and finalize_workflow_tidyclust() are deprecated. Use tune::finalize_model() and tune::finalize_workflow() instead, which now support cluster_spec objects natively. (#223)

New Models and Engines

• New db_clust() clustering specification for fitting DBSCAN models, with engines "dbscan" and "hdbscan". (#209, #238)

• New gm_clust() clustering specification for fitting Gaussian mixture models, with engine "mclust". (#209)

• New mean_shift() clustering specification for fitting mean shift models, which iteratively shift observations toward regions of high density and determine the number of clusters automatically. Engines "LPCM" and "meanShiftR" are supported. (#240, #244)

Improvements

• Added dials parameter constructors radius(), min_points(), circular(), zero_covariance(), shared_orientation(), shared_shape(), and shared_size() so that tuning parameters for db_clust() and gm_clust() resolve to real parameter objects rather than erroring on unexported dials:: names.

• Added a "Getting started with tidyclust" vignette (vignette("tidyclust")). (#232)

• Added butcher support for cluster_fit objects. axe_data() removes the training data stored in the fit, and axe_env() clears the environment reference from the preprocessing terms. (#126)

• contr_one_hot() is now exported, fixing the indicators = "one_hot" code path in .convert_form_to_x_fit() and .convert_form_to_x_new(). (#218)

• extract_cluster_assignment(), extract_centroids(), and predict() gain a labels argument, a character vector of cluster labels that overrides the auto-generated prefix-based labels. (#148)

• hier_clust() gains a dist_fun argument for specifying a custom distance function. (#70)

• hier_clust() documentation now clarifies that predict() may not match extract_cluster_assignment() on training data: predict() uses a distance-based heuristic while extract_cluster_assignment() uses cutree() based on the dendrogram structure. (#208)

• The dist_fun argument accepted by cluster metrics is now documented, including how to use {philentropy} to supply custom distance methods. See vignette("tuning_and_metrics", package = "tidyclust") for examples. (#185)

• tune_cluster() now supports parallel processing via the mirai package in addition to future. (#220)

• tune_cluster() now warns when passed an apparent() resample. Metrics from apparent resamples are excluded by collect_metrics(summarize = TRUE) (the default) since tune 1.2.0, which caused unexpected NA values. Use collect_metrics(summarize = FALSE) to see per-resample metrics. (#193)

• The .notes column returned by tune_cluster() now includes a trace column containing backtraces for errors and warnings, making it easier to debug failures. (#220)

Bug Fixes

• Fixed bug when trying to tune the linkage_method argument. (#206, @lgaborini)

• silhouette_avg() now has direction = "maximize" instead of direction = "zero", so that show_best() and select_best() correctly return models with the highest silhouette values. (#212, @dnldelarosa)

• sse_within_total() now correctly applies a custom dist_fun when new_data is NULL by using training data stored in the model. (#184)

Breaking Changes

• The foreach package is no longer supported for parallel processing in tune_cluster(). Use the future or mirai packages instead. See ?tune::parallelism for details. (#220)

• The .config column produced by tune_cluster() has changed from the Preprocessor{num}_Model{num} pattern to pre{num}_mod{num}_post{num} to align with updates in the tune package. (#220)