dist_fun from public knn_dist(). Supplying new_data selects
tabular prediction using mdist() internally; dist_args passes its distance
settings. Without new_data, x contains precomputed test-to-training
distances. The lower-level tidymodels engine interfaces are unchanged.mode from all_dist_method_specs(). The default preset = NULL
returns custom combinations, a vector of names returns only those named
presets, and preset = "all" returns all named presets. Include "custom"
in a preset selection to also generate component-filtered custom rows.
Component filters without custom rows are rejected rather than ignored."std" and "pc_scores",
omitting "robust" as redundant. This changes the candidate grid only;
mdist() still supports all its numerical preprocessing choices.response_aware_only mode.HLeucl and hl
with direct coordinate differences. This prevents zero test-to-training
distances observed with larger training sets in the previous backend,
for both commensurable and non-commensurable HLeucl.mdist() and step_mdist(),
including indicator-based categorical methods. Existing variable-wise
training averages (including the zero diagonal) are preserved.knn_dist() for direct classification, class probabilities, and
regression from test-to-training distances, or from tabular predictors using
mdist() internally. The response argument
selects and excludes the outcome column without manual predictor selection.
Existing fit/predict engine
functions remain available for tidymodels and compatibility.Removed the redundant spec_type field from distance-specification grids
and benchmark results. preset = "custom" selects explicit method settings;
named presets select their predefined settings. Legacy spec_type columns
are ignored on input. Update grid filters to use preset instead.
Changed LOVO and LOVO-comparison plots so reorder = TRUE ranks variables
within their categorical and numerical groups, preserving the meaning of
the background bands. Comparison rankings use the mean metric across
methods; top_n selection remains global.
Changed the default fill gradient for pairwise benchmark heatmaps to
coral (#E76F51) for lower values and blue (#008CFF) for higher values,
with regular-weight white cell labels. This applies to all metrics plotted by
autoplot.MDistBenchmark(), including clustering agreement.
Left self-comparison tiles on the main diagonal blank. ARI heatmaps now use a fixed 0 to 1 colour scale instead of the observed range. Negative ARIs remain labelled and use the low-end colour.
Changed MDS-based diagnostics in lovo_mdist() and
compare_lovo_mdist() to be opt-in. Set mds = TRUE to compute
congruence and alienation diagnostics; dims = 2 remains the default MDS
dimensionality when MDS is requested. Distance-based LOVO diagnostics remain
available by default, and clustering diagnostics remain controlled by
cluster_k.
Added compact print() and summary() methods for MDistBenchmark
objects. The full benchmark remains available as a tibble, while interactive
output emphasizes run status and failures. summary() displays one labelled
upper-triangular table per diagnostic, with blank diagonal and lower triangle.
Use metric to select measures (default: all; "ari" selects all available
clustering ARIs), and digits to control rounding. It still invisibly returns
the complete, unrounded pairwise-results tibble for further analysis.
Removed benchmark_comparisons(). Use pairs <- summary(result) instead.
This is a breaking API change; autoplot() continues to work directly on
the benchmark object.
Added the "u_dep_bw" preset for association-aware distances with block-wise
numerical commensurability. The preset computes Manhattan distances on
whitened principal-component scores and scales the complete numerical block
so that its mean over distinct training pairs equals the number of original
numerical variables. It retains the response-aware categorical construction
used by "u_dep".
"u_dep_bw" supports principal-component selection through ncomp or
threshold. For new observations, the PCA transformation and numerical block
scaling estimated from the training data are reused.
benchmark_mdist() to compare every pair of successful distance
specifications using mean absolute distance differences, symmetric relative
distance, multidimensional-scaling congruence, and alienation.benchmark_mdist(). Supplying
cluster_k computes pairwise adjusted Rand indices for PAM, hierarchical,
and/or spectral clustering; clustering is skipped when cluster_k = NULL.benchmark_comparisons() to extract the pairwise diagnostics stored in
an MDistBenchmark result without recomputing the distances.autoplot() method for MDistBenchmark objects, with annotated
heatmaps for distance, geometry, and clustering-agreement diagnostics.step_mdist() so response-aware specifications can obtain a single
outcome directly from the recipe formula during preparation. The fitted
response-aware profiles are reused when new data are baked, so assessment
and test outcomes are neither required nor used.response_used argument to step_mdist(), allowing response use to
be disabled explicitly.method_num to override the default standardization of the
"euclidean" preset for numerical-only data. In particular,
method_num = "none" computes ordinary Euclidean distances on the original
variables.wdi_2022, a documented snapshot of selected 2022 World Development
Indicators for reproducible mixed-type distance examples.step_mdist() workflows and the
pairwise benchmarking interface.manydist from a package focused on mixed-type distance construction to a broader framework for distance-based learning with mixed-type data.mdist() interface and documentation for mixed-type distance construction.step_mdist() for integrating manydist distances into recipes and tidymodels workflows.nearest_neighbor_dist() and related prediction functions for nearest-neighbour models based on precomputed or manydist-generated distances.pam_dist() for partitioning around medoids using manydist dissimilarities.spectral_dist() and spectral_from_dist() for spectral clustering from distance matrices.lovo_mdist() for leave-one-variable-out diagnostics of distance matrices.compare_lovo_mdist() and lovo_method_spec() for comparing LOVO diagnostics across multiple distance specifications.gen_mixed() and generate_dataset() for generating mixed-type example and simulation data.benchmark_mdist() for benchmarking distance specifications across datasets and method grids.all_dist_method_specs() and distance-method metadata helpers.0.5.0 release.