kccaExtendedFamily: Extending K-Centroids Clustering to (Mixed-with-)Ordinal Data
In flexord: Flexible Clustering of Ordinal and Mixed-with-Ordinal Data

kccaExtendedFamily

R Documentation

Extending K-Centroids Clustering to (Mixed-with-)Ordinal Data

Description

This wrapper creates objects of class "kccaFamily", which can be used with flexclust::kcca() to conduct K-centroids clustering using the following methods:

kModes (after Weihs et al., 2005)
kGower (Gower's distance after Kaufman & Rousseeuw, 1990, and a user specified centroid)
kGDM2 (GDM2 distance after Walesiak et al., 1993, and a user specified centroid)

Usage

kccaExtendedFamily(which = c('kModes', 'kGDM2', 'kGower'),
                   cent = NULL,
                   preproc = NULL,
                   xrange = NULL,
                   xmethods = NULL,
                   trim = 0, groupFun = 'minSumClusters')

Arguments

`which`	One of either `'kModes'`, `'kGDM2'` or `'kGower'`, the three predefined methods for K-centroids clustering. For more information on each of them, see the Details section.
`cent`	Function for determining cluster centroids. This argument is ignored for `which='kModes'`, and `centMode` is used. For `'kGDM2'` and `'kGower'`, `cent=NULL` defaults to a general purpose optimizer.
`preproc`	Preprocessing function applied to the data before clustering. This argument is ignored for `which='kGower'`. In this case, the default preprocessing proposed by Gower (1971) and Kaufman & Rousseeuw (1990) is conducted. For `'kGDM2'` and `'kModes'`, users can specify preprocessing steps here, though this is not recommended.
`xrange`	The range of the data in `x`. Options are: `"all"`: uses the same minimum and maximum value for each column of `x` by determining the whole range of values in the data object `x`. `"columnwise"`: uses different minimum and maximum values for each column of `x` by determining the columnwise ranges of values in the data object `x`. A vector of `c(min, max)`: specifies the same minimum and maximum value for each column of `x`. A list of vectors `list(c(min1, max1), c(min2, max2),...)` with length `ncol(x)`: specifies different minimum and maximum values for each column of `x`. This argument is ignored for `which='kModes'`. `xrange=NULL` defaults to `"all"` for `'kGDM2'`, and to `"columnwise"` for `'kGower'`.
`xmethods`	An optional character vector of length `ncol(x)` that specifies the distance measure for each column of `x`. Currently only used for `'kGower'`. For `'kGower'`, `xmethods=NULL` results in the use of default methods for each column of `x`. For more information on allowed input values, and default measures, see the Details section.
`trim`	Proportion of points trimmed in robust clustering, wee `flexclust::kccaFamily()`.
`groupFun`	A character string specifying the function for clustering. Default is `'minSumClusters'`, see [flexclust::kccaFamily()].

Details

Wrappers for defining families are obtained by specifying which using:

which='kModes' creates an object for kModes clustering, i.e., K-centroids clustering using Simple Matching Distance (counts of disagreements) and modes as centroids. Argument cent is ignored for this method.
which='kGower' creates an object for performing clustering using Gower's method as described in Kaufman & Rousseeuw (1990):
- Numeric and/or ordinal variables are scaled by \frac{\mathbf{x}-\min{\mathbf{x}}}{\max{\mathbf{x}-\min{\mathbf{x}}}}. Note that for ordinal variables the internal coding with values from 1 up to their maximum level is used.
- Distances are calculated for each column (Euclidean distance, distEuclidean, is recommended for numeric, Manhattan distance, distManhattan for ordinal, Simple Matching Distance, distSimMatch for categorical, and Jaccard distance, distJaccard for asymmetric binary variables), and they are summed up as:
  
  d(x_i, x_k) = \frac{\sum_{j=1}^p \delta_{ikj} d(x_{ij}, x_{kj})}{\sum_{j=1}^p \delta_{ikj}}
  
  where p is the number of variables and with the weight \delta_{ikj} being 1 if both values x_{ij} and x_{kj} are not missing, and in the case of asymmetric binary variables, at least one of them is not 0.
  
  The columnwise distances used can be influenced in two ways: By passing a character vector of length p to xmethods that specifies the distance for each column. Options are: distEuclidean, distManhattan, distJaccard, and distSimMatch. Another option is to not specify any methods within kccaExtendedFamily, but rather pass a "data.frame" as argument x in kcca, where the class of the column is used to infer the distance measure. distEuclidean is used on numeric and integer columns, distManhattan on columns that are coded as ordered factors, distSimMatch is the default for categorically coded columns, and distJaccard is the default for binary coded columns.
  
  For this method, if cent=NULL, a general purpose optimizer with NA omission is applied for centroid calculation.
which='kGDM2' creates an obejct for clustering using the GDM2 distance for ordinal variables. The GMD2 distance was first introduced by Walesiak et al. (1993), and adapted in Ernst et al. (2025), as the distance measure within flexclust::kcca().

This distance respects the ordinal nature of a variable by conducting only relational operations to compare values, such as \leq, \geq and =. By obtaining the relative frequencies and empirical cumulative distributions of x, we allow for comparison of two arbitrary values, and thus are able to conduct K-centroids clustering. For more details, see Ernst et al. (2025).

Also for this method, if cent=NULL, a general purpose optimizer with NA omission will be applied for centroid calculation.

Scale handling. In 'kModes', all variables are treated as unordered factors. In 'kGDM2', all variables are treated as ordered factors, with strict assumptions regarding their ordinality. 'kGower' is currently the only method designed to handle mixed-type data. For ordinal variables, the assumptions are more lax than with GDM2 distance.

NA handling. NA handling via omission and upweighting non-missing variables is currently only implemented for 'kGower'. Within 'kModes', the omission of NA responses can be avoided by coding missings as valid factor levels. For 'kGDM2', currently the only option is to omit missing values completely.

Value

An object of class "kccaFamily".

References

Ernst, D, Ortega Menjivar, L, Scharl, T, Grün, B (2025). Ordinal Clustering with the flex-Scheme. Austrian Journal of Statistics. Submitted manuscript.
Gower, JC (1971). A General Coefficient for Similarity and Some of Its Properties. Biometrics, 27(4), 857-871. \Sexpr[results=rd]{tools:::Rd_expr_doi("10.2307/2528823")}
Kaufman, L, Rousseeuw, P (1990). Finding Groups in Data: An Introduction to Cluster Analysis. Wiley Series in Probability and Statistics. \Sexpr[results=rd]{tools:::Rd_expr_doi("10.1002/9780470316801")}
Leisch, F (2006). A Toolbox for K-Centroids Cluster Analysis. Computational Statistics and Data Analysis, 17(3), 526-544. \Sexpr[results=rd]{tools:::Rd_expr_doi("10.1016/j.csda.2005.10.006")}
Walesiak, M (1993). Statystyczna Analiza Wielowymiarowa w Badaniach Marketingowych. Wydawnictwo Akademii Ekonomicznej, 44-46.
Weihs, C, Ligges, U, Luebke, K, Raabe, N (2005). klaR Analyzing German Business Cycles. In: Data Analysis and Decision Support, Springer: Berlin. 335-343. \Sexpr[results=rd]{tools:::Rd_expr_doi("10.1007/3-540-28397-8_36")}

Examples

# Example 1: kModes
set.seed(123)
dat <- data.frame(cont = sample(1:100, 10, replace=TRUE)/10,
                  bin_sym = as.logical(sample(0:1, 10, replace=TRUE)),
                  bin_asym = as.logical(sample(0:1, 10, replace=TRUE)),                     
                  ord_levmis = factor(sample(1:5, 10, replace=TRUE),
                                      levels=1:6, ordered=TRUE),
                  ord_levfull = factor(sample(1:4, 10, replace=TRUE),
                                       levels=1:4, ordered=TRUE),
                  nom = factor(sample(letters[1:4], 10, replace=TRUE),
                               levels=letters[1:4]))
flexclust::kcca(dat, k=3, family=kccaExtendedFamily('kModes'))

# Example 2: kGDM2
flexclust::kcca(dat, k=3, family=kccaExtendedFamily('kGDM2',
                                                    xrange='columnwise'))
# Example 3: kGower
flexclust::kcca(dat, 3, kccaExtendedFamily('kGower'))
nas <- sample(c(TRUE,FALSE), prod(dim(dat)), replace=TRUE, prob=c(0.1,0.9)) |> 
   matrix(nrow=nrow(dat))
dat[nas] <- NA
flexclust::kcca(dat, 3, kccaExtendedFamily('kGower',
                                           xrange='all'))
flexclust::kcca(dat, 3, kccaExtendedFamily('kGower',
                                           xmethods=c('distEuclidean',
                                                      'distEuclidean',
                                                      'distJaccard',
                                                      'distManhattan',
                                                      'distManhattan',
                                                      'distSimMatch')))
#the case where column 2 is a binary variable, but is symmetric

flexord documentation built on April 3, 2025, 8:53 p.m.