Distance-Based Learning for Mixed-Type Data

Provides tools for constructing, computing, and using distance measures for numerical, categorical, and mixed-type data. The package implements a flexible framework in which continuous and categorical components can be combined under additive, commensurable, and association-aware specifications. Supported methods include classical distances such as Gower, Euclidean, Manhattan, and Mahalanobis-type distances; categorical dissimilarities such as simple matching, occurrence-frequency, and association-based measures; and mixed-type presets designed to reduce biases due to variable type, scale, distribution, redundancy, and number of categories. The package also provides scaling options, supervised and unsupervised distance constructions, leave-one-variable-out tools for distance-based variable importance, and integration with distance-based learning workflows such as nearest-neighbour prediction, partitioning around medoids, and spectral clustering. Methods are motivated by van de Velden, Iodice D'Enza, Markos, and Cavicchia (2026) and related work on categorical and mixed-type dissimilarities.

manydist

Distance-Based Learning for Mixed-Type Data

manydist provides tools for constructing, computing, and using distance measures for numerical, categorical, and mixed-type data.

Installation

# install.packages("palmerpenguins")
install.packages("manydist")

You can also install the development version from GitHub:

# install.packages("remotes")
remotes::install_github("alfonsoIodiceDE/manydist_package", subdir = "manydist")

Initialization

library(manydist)
library(palmerpenguins)
library(dplyr)
library(tidyr)

Getting started

The main function in manydist is mdist(). It computes dissimilarities for numerical, categorical, and mixed-type data.

penguins_small <- penguins |>
  select(-year) |> drop_na()

A mixed-type dissimilarity can be computed with a preset. For example, preset = "gower" computes a Gower-type dissimilarity.

D <- penguins_small |> select(-species) |> 
  mdist(preset = "gower")

as.matrix(D$distance)[1:5,1:5] |> round(digits=2)

##      1    2    3    4    5
## 1 0.00 0.21 0.25 0.24 0.07
## 2 0.21 0.00 0.07 0.09 0.25
## 3 0.25 0.07 0.00 0.06 0.26
## 4 0.24 0.09 0.06 0.00 0.23
## 5 0.07 0.25 0.26 0.23 0.00

Some presets can use a response variable. When a response is supplied, it is used for response-aware distance construction and is not treated as an ordinary predictor.

D_resp <- penguins_small |> mdist(
  response = species,
  preset = "u_dep"
)

D_resp

## MDist object
##   preset : u_dep 
##   number of observations : 333 
##   number of continuous variables   : 4 
##   number of categorical variables   : 2 
##   parameters:
##     - categorical method: tvd
##     - numerical preprocessing: pc_scores
##     - commensurability adjustment: TRUE

Custom specifications can be defined by combining a categorical method, a numerical preprocessing method, and a commensurability rule.

D_custom <- mdist(penguins_small |> select(-species),
                  preset = "custom",
                  method_cat = "matching",
                  method_num = "std",
                  commensurable = TRUE
)

D_custom

## MDist object
##   preset : custom 
##   number of observations : 333 
##   number of continuous variables   : 4 
##   number of categorical variables   : 2 
##   parameters:
##     - categorical method: matching
##     - numerical preprocessing: std
##     - commensurability adjustment: TRUE
##     - number of principal components: 6
##     - inertia threshold: NULL