Skip to contents

Changelog

manydist 0.5.2

Simplified direct kNN prediction

  • Removed 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.

Simpler specification grids

  • Removed 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.
  • Commensurable custom grids now retain only "std" and "pc_scores", omitting "robust" as redundant. This changes the candidate grid only; mdist() still supports all its numerical preprocessing choices.
  • Select response-aware specifications using method metadata or explicit preset/method names instead of the former response_aware_only mode.

Consistent new-data distances

  • Replaced the Euclidean cross-distance backend used by 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.
  • Corrected Gower’s test-to-training average to divide by the number of original predictors rather than the number of expanded dummy columns. Training-to-training distances and the unaveraged Gower sum are unchanged.

Training-only commensurability

  • Corrected numerical and categorical commensurability so each contribution uses its training-to-training mean before aggregation, never the mean of a test-to-training matrix. This applies to mdist() and step_mdist(), including indicator-based categorical methods. Existing variable-wise training averages (including the zero diagonal) are preserved.
  • Numerical absolute-distance means and categorical means are obtained without allocating observation-level training squares solely for normalization.
  • Robust numerical preprocessing now also uses training medians and IQRs rather than recomputing them on each test batch.

Direct prediction

  • Added 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.
  • Fixed neighbour indexing for a single test observation with multiple neighbours and probability output for a single outcome level. Empty precomputed test matrices are also supported in the kNN helpers.

Diagnostics

  • 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.

Distance construction

  • 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.

manydist 0.5.1

CRAN release: 2026-07-23

Benchmarking and diagnostics

  • Extended benchmark_mdist() to compare every pair of successful distance specifications using mean absolute distance differences, symmetric relative distance, multidimensional-scaling congruence, and alienation.
  • Added optional clustering comparisons to benchmark_mdist(). Supplying cluster_k computes pairwise adjusted Rand indices for PAM, hierarchical, and/or spectral clustering; clustering is skipped when cluster_k = NULL.
  • Added benchmark_comparisons() to extract the pairwise diagnostics stored in an MDistBenchmark result without recomputing the distances.
  • Added an autoplot() method for MDistBenchmark objects, with annotated heatmaps for distance, geometry, and clustering-agreement diagnostics.

Distance construction and recipe workflows

  • Updated 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.
  • Added the response_used argument to step_mdist(), allowing response use to be disabled explicitly.
  • Allowed 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.

Data

  • Added wdi_2022, a documented snapshot of selected 2022 World Development Indicators for reproducible mixed-type distance examples.

Documentation and testing

  • Added focused tests for response-aware step_mdist() workflows and the pairwise benchmarking interface.
  • Expanded the package website with task-oriented articles on distance construction, diagnostics and benchmarking, clustering, and nearest-neighbour workflows.

manydist 0.5.0

CRAN release: 2026-06-09

Major changes

  • Expanded manydist from a package focused on mixed-type distance construction to a broader framework for distance-based learning with mixed-type data.
  • Updated the package title and description to reflect support for distance construction, distance-based modelling workflows, variable-importance diagnostics, and clustering.
  • Changed the package maintainer from Angelos Markos to Alfonso Iodice D’Enza.

Distance construction

  • Added a revised mdist() interface and documentation for mixed-type distance construction.
  • Added support for additional mixed-type distance specifications and presets.
  • Added response-aware distance construction tools for supervised mixed-type workflows.
  • Added interaction-aware distance components for continuous-categorical relationships.
  • Added helper infrastructure for preprocessing and applying mixed-type distance specifications consistently across training and new data.
  • Added utilities for generating and benchmarking distance-method specifications.

Distance-based learning workflows

  • Added step_mdist() for integrating manydist distances into recipes and tidymodels workflows.
  • Added nearest_neighbor_dist() and related prediction functions for nearest-neighbour models based on precomputed or manydist-generated distances.
  • Added pam_dist() for partitioning around medoids using manydist dissimilarities.
  • Added spectral_dist() and spectral_from_dist() for spectral clustering from distance matrices.
  • Added support functions for converting distances to affinities and fitting distance-based clustering models.

Variable importance and diagnostics

  • Added lovo_mdist() for leave-one-variable-out diagnostics of distance matrices.
  • Added compare_lovo_mdist() and lovo_method_spec() for comparing LOVO diagnostics across multiple distance specifications.
  • Added congruence- and alienation-based diagnostics for comparing multidimensional scaling configurations.
  • Added optional clustering-based LOVO diagnostics using PAM, hierarchical clustering, and spectral clustering.

Data generation and benchmarking

Documentation

  • Added documentation for the new modelling, clustering, LOVO, benchmarking, and recipe functions.
  • Updated the package-level description and metadata for the 0.5.0 release.
  • Removed CRAN-inappropriate development files, caches, and vignette outputs from the source build.