Econometrics and Statistics 30 (2024) 110–123

Contents lists available at ScienceDirect

Econometrics and Statistics

journal homepage: www.elsevier.com/locate/ecosta

Fuzzy k-Means: history and applications

Maria Brigida Ferraro a
Dipartimento di Scienze Statistiche, Sapienza Università di Roma, P.le Aldo Moro, 5 00185, Rome, Italy

a r t i c l e i n f o a b s t r a c t

Article history: The fuzzy approach to clustering arises to cope with situations where objects have not a
Received 18 February 2021 clear assignment. Unlike the hard/standard approach where each object can only belong
Revised 21 November 2021
to exactly one cluster, in a fuzzy setting, the assignment is soft; that is, each object is
Accepted 21 November 2021
assigned to all clusters with certain membership degrees varying in the unit interval. The
Available online 26 November 2021
best known fuzzy clustering algorithm is the fuzzy k-means (FkM), or fuzzy c-means. It is
Keywords: a generalization of the classical k-means method. Starting from the FkM algorithm, and in
Fuzzy clustering more than 40 years, several variants have been proposed. The peculiarity of such different
Fuzzy k-Means proposals depends on the type of data to deal with, and on the cluster shape. The aim is to
Mixed data show fuzzy clustering alternatives to manage different kinds of data, ranging from numeric,
Fuzzy data categorical or mixed data to more complex data structures, such as interval-valued, fuzzy-
Functional data valued or functional data, together with some robust methods. Furthermore, the case of
Double clustering two-mode clustering is illustrated in a fuzzy setting.
© 2021 EcoSta Econometrics and Statistics. Published by Elsevier B.V. All rights reserved.

1. Introduction

The hard/standard approach to clustering consists in assigning each object to one and only one cluster. Thus, some objects
are forced to be assigned to a given cluster despite being far from the prototype. The soft approach to clustering allows us to
assign units to all the clusters with a degree ranging in [0,1]. In this case, the assignment is referred to as a membership and
not as a simply allocation, as in the hard approach. There exist several types of soft clustering: fuzzy, possibilistic, and rough
clustering, among others (see, for a review, Ferraro and Giordani (2020)). In a fuzzy setting, the membership degrees are
not based on probabilistic assumptions but just on the distances between objects and prototypes. The possibilistic approach
differs from the fuzzy one only because some constraints on membership degrees are relaxed, while in the rough clustering,
there are not degrees taking values in the unit interval but objects with no clear assignment (belonging to the boundary
region between two clusters) are associated to more than one cluster. Even if it is not always recognized as a soft approach,
model-based clustering also provides a soft partition. Unlike the previous ones, it is based on probabilistic assumptions. In
such a clustering approach, each cluster is viewed as a component of a mixture model, and the the posterior probability of
a component membership may play the same role as the membership degree in the fuzzy clustering.
In this work, the focus is on the fuzzy approach. The first and most known fuzzy clustering algorithm is the generalization
of the standard k-Means MacQueen (1967); Bock (2007), the fuzzy k-Means (FkM). It was introduced in Dunn (1973) but
deeply analyzed and improved in Bezdek (1974, 1981). Fuzzy clustering has also been shown to be a valuable tool from a
computational point of view, making the clustering algorithm to become more efficient Klawonn et al. (2015). There exist
about 11140 0 0 documents containing “Fuzzy k-means” (or “Fuzzy c-means”) on Google, and 178740 on Google Scholar. The
time series of the number of documents on Google Scholar and Scopus is reported in Figure 1, showing an exponential
growth.

E-mail address: mariabrigida.ferraro@uniroma1.it

https://doi.org/10.1016/j.ecosta.2021.11.008
2452-3062/© 2021 EcoSta Econometrics and Statistics. Published by Elsevier B.V. All rights reserved.
M.B. Ferraro Econometrics and Statistics 30 (2024) 110–123

Fig. 1. Number of documents on Google Scholar and Scopus containing ”Fuzzy k-Means” (or ”Fuzzy c-Means”).

Fig. 2. Documents on Scopus containing ”Fuzzy k-Means” (or ”Fuzzy c-Means”) in the Article title, in the Abstract or in the Keywords, by Subject Area.

Focusing on Scopus, there are near 160 0 0 documents (about 2500 Open Access) containing “Fuzzy k-Means” (or “Fuzzy
c-Means”) in the Article title, in the Abstract or in the Keywords (standard default search in Scopus). Regarding the areas in
which this method is more popular, Figure 2 reports the standard search on Scopus containing “Fuzzy k-Means” (or “Fuzzy
c-Means”) by subjects. About 33% of the documents are in computer science, 23.7% in engineering and 12.3% in mathematics
plus some application areas that underline its multidisciplinarity.
A review of the Fuzzy c-Means algorithm from 20 0 0 to 2014 is provided in Nayak et al. (2015). Starting from the FkM,
several variants have been proposed based on different distance measures, different prototype definitions and different kinds
of data. Most proposals are devoted to numerical data: the Gustafson-Kessel Fuzzy k-Means (GK-FkM) Gustafson and Kessel
(1979), the Entropic Fuzzy k-Means (EFkM) Li and Mukaidono (1995, 1999), the Fuzzy k-Means with Polynomial Fuzzifier
(FkMPF) Klawonn and Höppner (2003); Winkler et al. (2011), the Fuzzy k-Medoids (FkMed) Krishnapuram et al. (2001), the
Fuzzy k-Means with Noise cluster (FkMN) Davé (1991), among others.
While fuzzy cluster analysis techniques for object data (units by variables) matrices have received significant interest,
fuzzy clustering of relational data has received less attention, because in most cases, an object matrix, and not pure rela-
tional data, is available. Relational data are represented by a measure of similarity (or dissimilarity) between the elements
that is sometimes obtained from the objects themselves. An example of this case is represented by the Euclidean distance
between objects. Proposals of fuzzy clustering techniques for relational data are provided, for example, in Roubens (1978);
Windham (1985); Hathaway et al. (1989); Kaufman and Rousseuw (1990); Hathaway and Bezdek (1994); Hathaway et al.
(1996); Davé and Sen (2002).
The above fuzzy clustering methods for relational data can also be used for other kinds of data, for example, categorical
or mixed data, as long as an appropriate dissimilarity measure is defined, but there also exist specific proposals for categor-
ical or mixed data. This is the case of the fuzzy k-modes Huang and Ng (1999) for categorical and the fuzzy k-prototypes Ji
et al. (2012) for mixed data, among others.
In more than 40 years, several variants of FkM have been extended to more complex data structures such as fuzzy-valued,
interval-valued and functional data.
In many real-life situations, some measurements may be imprecise and some observations may be vaguely defined. In
such contexts, it is appropriate to represent the information by means of either interval data or fuzzy data instead of consid-

111
M.B. Ferraro Econometrics and Statistics 30 (2024) 110–123

ering crisp/hard values Zadeh (1965). Several statistical techniques have been introduced for fuzzy/interval data (seeColubi
(2021) for a review in statistics and econometrics). In a clustering framework, extensions of the FkM algorithm for fuzzy data
are provided in Sato and Sato (1995); Hathaway et al. (1996); Yang and Ko (1996); Pedrycz et al. (1998); Takata et al. (1998);
Yang and Liu (1999); Takata et al. (2001); Auephanwiriyakul and Keller (2002); Hung and Yang (2005); D’Urso and Giordani
(2006b); Pelekis et al. (2008); Gonzalez-Rodriguez et al. (2009); Coppi et al. (2012); Fritz et al. (2013); D’Urso and de Gio-
vanni (2014); Giordani and Ramos-Guajardo (2016); Ferraro and Giordani (2017); D’Urso and Leski (2020); Ramos-Guajardo
and Ferraro (2020). If we consider intervals as a particular type of fuzzy data, we can find proposals of FkM-type algorithms
in Gao et al. (20 0 0); D’Urso and Giordani (2006a); D’Urso et al. (2015). On the other hand, if intervals are managed as sym-
bolic data Bock and Diday (20 0 0), some fuzzy clustering methods are illustrated in de Carvalho (2007); de Carvalho and
Tenorio (2010); Pimentel and de Souza (2014); de Carvalho and Simoes (2017).
Among complex data, functional data also deserve attention. Advances in data collection and storage have led to an
increasing amount of this kind of data. Hence, appropriate statistical methods are needed Ramsay and Silverman (2005);
Ferraty and Vieu (2006). Nowadays, it is usual to encounter functional data in many fields such as engineering, economics,
finance biology, medicine, or meteorology. Some proposals of fuzzy clustering of functional data are based on transforming
the data before applying a fuzzy k-means type algorithm (see, for example, Di Maio et al. (2011) and Giordani et al. (2020)).
A different proposal is provided in Tokushige et al. (2007). The authors suggest using a dissimilarity matrix that is itself a
function. Consequently, the cluster prototypes and the membership degrees are also defined as functions.
The above proposals of clustering techniques are also known as one-mode (or one-way) clustering to distinguish them
from two-mode clustering. Two-mode clustering is helpful to synthesize more complex data structures. In particular, we
refer to data matrices whose two modes have an interchangeable role. The approach is symmetrical, and the aim is to look
for blocks (or sub-matrices) characterized by internal cohesion and external separation. In some situations, for example,
units can be similar only on a subgroup of features, and some variables can be associated only within a subgroup of units.
This is the case of products and customers in market basket analysis or genes and samples in DNA microarray analysis. In
this setting, an extension of the FkM is the Fuzzy double k-Means introduced in Ferraro et al. (2015). Further variants and
robust versions are provided in Ferraro et al. (2021).
The paper is structured as follows. Section 2 is devoted to the FkM method and its variants for numerical data. Fur-
thermore, the proposals for relational data are briefly described in Subsection 2.1. Section 3 contains a review of the mod-
ifications of FkM for categorical or mixed data. Fuzzy clustering of fuzzy and interval data are reported in Section 4 and
Subsection 4.1, respectively. The case of functional data is addressed in Section 5, whilst Section 6 contains some extensions
of the FkM algorithm for two-mode clustering of a data matrix. Some applications are reported in Subsections 2.2, 4.2 and
5.1. Finally, some concluding remarks and open problems are addressed in Section 7.

2. Fuzzy k-Means and its variants

The fuzzy approach to clustering is based on a soft assignment of units to clusters. Most of the proposals consider
object data, a typical unit-variable data matrix, with numerical variables. As already stated, the first and best known fuzzy
clustering algorithm is the Fuzzy k-Means (FkM) Bezdek (1974, 1981). Given an (n × p) matrix, X, where n and p are the
number of units and variables, respectively, the aim is to partition the units into k groups, where each group is characterized
by a prototype (centroid). Each row vector xi = [xi1 , xi2 , . . . , xip ] represents the i-th observation. The optimization problem of
the FkM can be formalized as:

n 
k
min JFkM = um
ig
d2 (xi , hg ),
U, H i=1 g=1

s.t. uig ∈ [0, 1], i = 1, . . . , n, g = 1, . . . , k, (1)


k
uig = 1, i = 1, . . . , n,
g=1

where d2 (xi , hg ) = xi − hg 2 is the squared Euclidean distance between unit i and prototype g.
In (1), the (n × k) matrix U denotes the membership degree matrix, where each element uig ∈ [0, 1] represents the mem-
bership degree of unit i to cluster g. The row-wise sums of U are equal to 1. The (k × p) matrix H denotes the prototype
(centroid) matrix, where each row hg = [hg1 , hg2 , . . . , hgp ] (g = 1, . . . , k) is the prototype of cluster g. Finally, the parameter
m > 1 is used to tune the fuzziness of the obtained partition. The greater the value of m, the further away from 1 and 0 the
membership degrees are. For m → 1 FkM reduces to the k-means (kM) algorithm. Empirical results depend on the selection
of this parameter, which is commonly chosen between 1.5 and 2 in practice, as shown in Pal and Bezdek (1995).
The optimal solution can be found by means of an iterative algorithm, where the updates are obtained through the
Lagrangian multiplier method. The centroid and the membership degree updates are

n
um x
ig i
i=1
hg = , g = 1, . . . , k, (2)
n
um
ig
i=1

112
M.B. Ferraro Econometrics and Statistics 30 (2024) 110–123

and
1
uig =   m1−1 , i = 1, . . . , n, g = 1, . . . , k. (3)

k
d2 (xi ,hg )
g =1 (
d2 xi ,hg )

Starting from the FkM, several variants have been proposed in the last decades. Since the initial FkM usually identifies
clusters of spherical shape, some of the variants are based on a non-Euclidean distance measure in order to overcome such
a drawback. This is the case of the Gustafson and Kessel proposal Gustafson and Kessel (1979): the FkM with covariance
matrices (FkM-GK). It consists in replacing the Euclidean distance with the Mahalanobis one:

2
dM (xi , hg ) = (xi − hg ) Fg (xi − hg ),
where Fg is a symmetric and definite positive matrix. The optimal solution of Fg depends on the inverse of the fuzzy co-
variance matrix of the g-th cluster. To avoid numerical problems that may arise when updating Fg , an improved version of
the FkM-GK is suggested in Babuska et al. (2002).The improvement consists in constraining the condition number of Fg to
be higher than a pre-specified threshold.
A further proposal to deal with different cluster shapes is provided in Gath and Geva (1989). In this case, the Euclidean
distance is replaced with a distance norm based on fuzzy maximum likelihood estimates: the Gauss distance.
Since the FkM algorithm involves the parameter m lacking a physical meaning, an alternative, based on an entropic
regularization to replace the fuzziness parameter m, has been proposed in Li and Mukaidono (1995, 1999). The entropic
FkM has been proven to be connected to the EM algorithm used for mixture models Hathaway (1986). An entropic version
of the FkM-GK is proposed in Ferraro and Giordani (2013a). It should be used in the case of non-spherical shape clusters.
A limitation of the above FkM type algorithms is the assignment of the units to all the clusters with non-zero member-
ship degrees. In order to overcome it, a generalization of the FkM algorithm is introduced in Klawonn and Höppner (2003).
In particular, a polynomial fuzzifier is used in the minimization problem. In general, a fuzzifier function is a continuous,
strictly increasing function f : [0, 1] −→ [0, 1] with f (0 ) = 0 and f (1 ) = 1. In the FkM case, f (uig ) = um
ig
. The polynomial
−β 2 2β
fuzzifier function is defined as f (uig ) = ( 11+ β uig + u ),
1+β ig
with β ∈ [0, 1]. For β = 0 we obtain the FkM with parameter m
equal to 2 and for β = 1 the hard/classical kM.
Most of the above algorithms are implemented in the R package fclust Ferraro and Giordani (2015); Ferraro et al. (2019).
As it happens for the k-means type methods, their fuzzy versions are not robust to outliers. This is due to the unit-sum
constraints of the membership degrees that force outliers to be assigned to clusters. Furthermore, since the centroids are
weighted means, they are affected by anomalous points.
There exist several proposals of robust fuzzy clustering methods. A timid kind of robustification consists in replacing the
centroids (means) with the medoids. The Fuzzy k-Medoids (FkMed) algorithm Krishnapuram et al. (2001) is a generalization
of the classical k-Medoids Kaufman and Rousseuw (1990). In Wu and Yang (2002), however, an alternative FkM is proposed,
by replacing the Euclidean distance in (1) with a “robust” distance measure, the exponential distance: dexp (xi , hg ) = 1 −
exp(−γ d (xi , hg )), where γ is a positive constant. Another robust approach consists in adding a noise cluster containing
all the units considered as outliers Davé (1991). It is important to highlight that the noise cluster is not a proper cluster
characterized by homogeneity (compactness). On the other hand, a trimmed approach of fuzzy clustering is suggested in
Fritz et al. (2013). The idea is to trim a fixed proportion of observations. In this case, the trimming proportion is to be fixed,
whilst in the previous one (noise cluster) we have to fix the distance of the observations from the noise prototype.
The possibilistic approach to clustering relaxes the unit-sum constraint, which boosts robustness. The membership de-
grees do not longer depend on the distance to all prototypes but only on the distance to the prototype of the belonging
cluster. This explains why they are also called degrees of typicality. The best known possibilistic clustering algorithm is the
possibilistic extension of the k-Means, the Possibilistic k-Means Krishnapuram and Keller (1993, 1996). Further possibilistic
clustering methods are provided in Yang and Wu (2006); Timm et al. (2004); Wachs et al. (2006). The first one Yang and
Wu (2006) is based on the exponential distance, robust to noise and outliers. The second one Timm et al. (2004) is based
on the introduction of a repulsion term, in order to overcome the risk of obtaining a trivial solution with coincident clusters
Barni et al. (1996). The third proposal Wachs et al. (2006) is an extension of the method introduced in Yang and Wu (2006).
Finally, in order to take into account the benefits of the fuzzy and possibilistic approaches and to overcome their draw-
backs, some hybridizations are proposed: the Fuzzy Possibilistic k-Means Pal et al. (1997), the Possibilistic Fuzzy k-Means
Pal et al. (2005) and the Modified Fuzzy and Possibilistic k-Means Saad and Alimi (2009), among others. The possibilistic
and hybrid proposals are implemented in the R package ppclust Cebeci (2019).

2.1. Fuzzy clustering algorithms for relational data

In several practical applications, the information is not available in terms of object data but only as relational data.
Relational data are pair-wise relations (similarity or dissimilarity/distance) between units. There exist different proposals of
fuzzy clustering algorithms for such a kind of data. A Relational duals of the FkM algorithm (RFkM) is proposed in Hathaway

113
M.B. Ferraro Econometrics and Statistics 30 (2024) 110–123

Table 1
FAO dataset: prototypes of the four clusters obtained with FkM

Exports Imports Product. PCFat SupplyDay SupplyYear PCProtein

Clus 1 0.22 0.13 0.35 3.69 504.02 63.33 15.58

Clus 2 0.08 0.11 0.15 4.21 707.69 91.13 21.79
Clus 3 0.07 0.18 0.09 7.59 924.79 117.35 27.57
Clus 4 0.37 0.04 0.59 3.90 888.57 115.30 27.46

et al. (1989). It is based on the following minimization problem:


n 
n
um um r (xi ,xi )

k
i=1 i =1
ig i g
min JRFkM = 
n ,
U g=1 2 um
ig
i=1

s.t. uig ∈ [0, 1], i = 1, . . . , n, g = 1, . . . , k, (4)


k
uig = 1, i = 1, . . . , n.
g=1

Unfortunately, the above method requires that r (xi , xi ) is the squared Euclidean distance. This is too restrictive in sev-
eral situations. A generalization of the RFkM algorithm is introduced in Hathaway and Bezdek (1994). It allows the use of
arbitrary dissimilarity data and consists in a modification of the algorithm by means of transformation of the dissimilarity
data. Further proposals are provided in Kaufman and Rousseuw (1990) and in Hathaway et al. (1996). The first one is known
as FANNY. It is characterized by a parameter of fuzziness m = 2 and the relational data usually are computed by using an
L1 norm. The second one illustrates how the FkM method can be applied to relational data. A modification of FANNY, called
FRC, is proposed in Davé and Sen (2002). In particular, a general fuzzifier is used and the relational data may be obtained
from any dissimilarity measure. As for FkM, the optimal solution is obtained by means of a Lagrangian function where
only the unit-sum constraints of the membership degrees are involved. When the relational data are Euclidean, the non-
negativity condition of the membership degrees is automatically satisfied, as in FkM. In the case of non-Euclidean distances,
neither RFkM nor FRC automatically satisfy that constraint. A solution is proposed in Davé and Sen (2002). A robust version
of fuzzy relational clustering is also provided by means of the introduction of a noise cluster Davé (1991). Obviously, the FRC
algorithm and its robust version can be applied to all kinds of data. Depending on the data, different dissimilarity matrices
may be used as input argument of the function. This includes categorical, mixed or more complex data. In addition, even
when data are numerical, several non-Euclidean distances can ben used in order to take into account, for example, different
cluster shapes. In this respect, a proposal of fuzzy clustering for nonlinearly separable data based on the geodesic distance
is introduced in Ferraro and Giordani (2019).

2.2. FkM: a real-case study

This section is devoted to a real-case study. We consider a food balance sheet provided by the Food and Agriculture
Organization (FAO). In particular, we analyze the “wheat and products” of European countries in 2018. The food balance
sheet shows for each food item the sources of supply and its utilization. The aim is to find homogeneous groups of the
39 European countries characterized by similar behaviour related to production, imports and exports of wheat, fat sup-
ply quantity (g/capita/day), food supply (kcal/capita/day), food supply quantity (kg/capita/yr) and protein supply quantity
(g/capita/day). The first three variables are normalized (divided) by population. We use the function FKM of the fclust pack-
age Ferraro et al. (2019). By inspecting the values of Fuzzy Silhouette Campello and Hruschka (2006), a cluster validity index
used to evaluate the partition quality, the optimal number of clusters, corresponding to the maximum value of the index, is
k = 4. The first cluster is composed by 7 countries: Belarus, Czechia, Denmark, Estonia, Latvia, Netherlands and Republic of
Moldova. Austria, Bosnia and Herzegovina, Croatia, Finland, Germany, North Macedonia, Norway, Poland, Portugal, Slovakia,
Spain, Sweden, Switzerland, United Kingdom of Great Britain and Northen Ireland are all contained in the second cluster.
The third cluster contains Albania, Belgium, Greece, Iceland, Ireland, Italy, Luxembourg and Malta. The remaining 8 countries
belong to the forth cluster: Bulgaria, France, Hungary, Lithuania, Romania, Russian Federation, Serbia and Ukraine.
The prototypes/centroids of the clusters are reported in Table 1. As we can notice, countries in Cluster 1 are characterized
by average values of imports, exports and production of wheat and related products, and the lowest values of supplies.
Cluster 4 contains countries with the highest production and exports and the lowest import quantity. Countries in Cluster
2 and Cluster 3 present a similar export quantity, but the first ones have a lower value of imports and higher production.
Furthermore, Cluster 3 is also characterized by the highest values of fat, protein and food supply quantities.
By inspecting the membership degree matrix (not reported here for the sake of brevity), there are three countries not
clear assigned (highest membership degree lower than 0.5): Belgium, Czechia and the Netherlands. In particular, Belgium is
assigned to Cluster 3 with a membership degree equal to 0.37 and its membership degree to Cluster 2 is 0.36. Czechia has
intermediate characteristics between Cluster 1 (uig = 0.47) and Cluster 2 (uig = 0.46). Finally, the Netherlands is assigned to

114
M.B. Ferraro Econometrics and Statistics 30 (2024) 110–123

Table 2
Feature values of Belgium, Czechia and Netherlands.

Exports Imports Product. PCFat SupplyDay SupplyYear PCProtein

Belgium 0.23 0.47 0.14 2.56 781 110.51 23.37

Czechia 0.21 0.05 0.41 4.80 647 84.36 18.44
Netherlands 0.09 0.43 0.06 2.46 584 68.42 18.30

Cluster 1 (uig = 0.46), but the membership degree to Cluster 2 is 0.39. This can also be seen by looking at the feature values
assumed by these three countries, reported in Table 2.

3. Extensions of Fuzzy k-Means for categorical/mixed data

This section is devoted to fuzzy clustering proposals for categorical or mixed data. As already observed in Subsection 2.1,
the fuzzy relational clustering methods can be applied by using specific dissimilarity/similarity measures for categorical or
mixed data. However, specific algorithms exist. An extension of the k-Means algorithm to clustering large data sets with
categorical variables is proposed in Huang (1998). It is known as k-Modes. A fuzzy version is introduced in Huang and Ng
(1999), the Fuzzy k-Modes (FkMo). It consists in using a simple matching measure for categorical data and replacing the
means with the modes. The optimization problem is


n 
k
ig c ( i
min JFkMo = um d x , hg ),
U,H i=1 g=1

s.t. uig ∈ [0, 1], i = 1, . . . , n, g = 1, . . . , k, (5)


k
uig = 1, i = 1, . . . , n,
g=1
 
where dc xi , hg is the simple matching dissimilarity measure. It is defined as


p
dc (xi , hg ) = δ (xi j , hg j ), (6)
j=1

where

0, if xi j = hg j ,
δ (xi j , hg j ) = (7)
1, if xi j = hg j .

The matrix H in (5) contains the modes. The main difference with the FkM algorithm is the update of H. In detail, denoting
(n j )
with DOM j = {a(j1 ) , a(j2 ) , · · · , a j } the set of n j categories of the j-th categorical variable, j = 1, · · · , p, the loss function JFkMo
(r )
is minimized iff hg j = a j ∈ DOM j where
 
um
ig ≥ um
ig 1 ≤ t ≤ nj (8)
(r ) (t )
i,xi j =a j i,xi j =a j

for 1 ≤ j ≤ p.
An extension of the FkMo is proposed in Kim et al. (2004). In particular, the clusters of categorical data are represented
by means of fuzzy centroids instead of the hard ones.
In the case of mixed data (units described by both numerical and categorical variables) neither the FkM nor the FkMo
can be used. For this reason, a combination of both methods is introduced Huang (1997): the (hard) k-Prototype algo-
rithm. A fuzzy extension, the Fuzzy k-Prototype (FkP), is proposed in Ji et al. (2012). In this case, each unit vector is
xi = [xri1 , xri2 , · · · , xrip , xci,p +1 , · · · xcip ], where the first p1 are numerical variables, denoted by superscript r, and the remaining
1 1
p − p1 are the categorical ones, denoted by superscript c.
The minimization problem is


n 
k
ig rc ( i
min JFkP = um d x , hg ),
U,H i=1 g=1

s.t. uig ∈ [0, 1], i = 1, . . . , n, g = 1, . . . , k, (9)


k
uig = 1, i = 1, . . . , n,
g=1

115
M.B. Ferraro Econometrics and Statistics 30 (2024) 110–123

Fig. 3. Three examples of LR fuzzy data: a trapezoidal fuzzy number (left), a triangular fuzzy number (center), an interval (right).

 
where drc xi , hg is defined as


p1

p
drc (xi , hg ) = (w j (xri j − hrg j ))2 + δ (xci j , hcg j ). (10)
j=1 j= p1 +1

The membership degree update is the same as the FkM algorithm with dissimilarity drc (xi , hg ). The prototype consists of
two parts. The first p1 elements are computed as weighted means of the numerical variables (as in FkM) and the remaining
p − p1 ones are the fuzzy modes introduced in Kim et al. (2004).

4. Fuzzy k-Means for fuzzy data and its variants

In various practical applications in economics, social science, biology, ecology or medical science, many useful variables
are vague or imprecise, and it is easier to capture the vagueness/imprecision by means of more complex data than to discard
it and obtain precise data. Imprecise data may be formalized by means of fuzzy numbers Zadeh (1965). The space of fuzzy
numbers, denoted by Fc (R ), is composed by the mappings U : R → [0, 1] such that for each α ∈ (0, 1] the so-called α -level
set (or α -cut) Uα = {x ∈ R|U (x ) ≥ α} belongs to the class of nonempty compact intervals in R (denoted by Kc (R )). The 0-
level, U0 , is the closure of the support of U .
The most used family of fuzzy numbers is the so-called class of LR-fuzzy numbers. An LR-fuzzy number U is determined
by four values, U ≡ (c1 , c2 , r, l )LR . In detail, c1 and c2 are the left and the right centers of the 1-level of U and represent the
location of the fuzzy number, while the right and left spreads of U , r and l are associated with the imprecision of U .
The membership degree of x to U is defined as
⎧ c − x
⎪
⎪ L
1
, if x < c1 ,
⎪
⎨ l
μU (x ) = 1, if c1 ≤ x ≤ c2 , (11)
⎪
⎪  
⎪
⎩ R x − c2 , if x > c2 ,
r

where L : R → [0, 1] (and R) is a convex upper semi-continuous function so that L(0 ) = 1 and L(x ) = 0, for all x ∈ R \ [0, 1]
(see Zimmermann (1996)). If L(z ) = 1 − z and R(z ) = 1 − z for 0 ≤ z ≤ 1, then U is a trapezoidal fuzzy number when c1 = c2
and a triangular fuzzy number when c1 = c2 = c. Furthermore, LR fuzzy numbers are a generalization of intervals. An interval
is got when c1 = c2 and l = r = 0 (see Figure 3).
The usual arithmetic between fuzzy numbers is a level-wise extension of the standard non-linear arithmetic for intervals
(Nguyen (1978); Zadeh (1965)). Given U , V ∈ Fc (R ) and λ ∈ R, the sum and the product by a scalar can be defined so that
for each α ∈ [0, 1] it is fulfilled that
   
U + λV = Uα + λVα = u + λv : u ∈ Uα , v ∈ Vα . (12)
α

 Given
 n units  described by p LR  fuzzy variables, a fuzzy data matrix can be defined as X =
xi j ≡ c 1 i j , c 2 i j , l i j , r i j , i = 1, . . . , n, j = 1, . . . , p , where xi j is the value of the LR fuzzy variable j observed on the i-
LR
th unit with left center c1i j , right center c2i j , and left and right spreads li j and ri j , respectively. The matrix X can be
characterized by four matrices: X ≡ (C1 , C2 , L, R )LR . Each observation i is expressed as xi ≡ (c1i , c2i , li , ri )LR .

116
M.B. Ferraro Econometrics and Statistics 30 (2024) 110–123

Let dF (·, · ) be a distance measure between fuzzy numbers, the Fuzzy k-Means for Fuzzy data (FkM-F) is formalized by
means of the following optimization problem:

n 
k  
min JF kM−F = um d2 xi , hg ,
ig F
U,H i=1 g=1

s.t. uig ≥ 0, i = 1, . . . , n, g = 1, . . . , k, (13)


k
uig = 1, i = 1, . . . , n,
g=1

where uig is the membership degree of observation i to cluster g, stored in the (n × k ) matrix U, and H =
     
C C C C
hg j ≡ hg1j , hg2j , hLg j , hRg j , g = 1, ..., k, j = 1, ..., p is the prototype matrix. In particular, hg j ≡ hg1j , hg2j , hLg j , hRg j repre-
LR LR
C C
sents the value of the j-th LR fuzzy variable of the g-th centroid with left center hg1j , right center hg2j , left spread hLg j and
C C
right spread hRg j . Hence, hg ≡ (hg1 , hg2 , hLg , hRg )LR is the fuzzy vector of length p for centroid g.
The proposal of FkM-F in Coppi et al. (2012) is based on a weighted dissimilarity measure for fuzzy data: dw2 ( x , x ).
i i
Given two LR fuzzy observations, xi and xi , it is defined as a weighted sum of the squared Euclidean distances between
centers and spreads:
2
rl dw (xi , xi ) = wC2 [d2 (c1i , c1i ) + d2 (c2i , c2i )] + w2S [d2 (li , li ) + d2 (ri , ri )], (14)
where d (·, · ) is the standard Euclidean distance (for non-fuzzy data), and wC and wS are weights for the center component
and the spread component.
As for FkM, the iterative solution of the constrained quadratic minimization problem (13) is obtained through the La-
grangian multiplier method Coppi et al. (2012).
A particular case of the FkM-F method is introduced in D’Urso and Giordani (2006b) for LR1 symmetric fuzzy data. A
symmetric LR1 fuzzy number is an LR fuzzy number with c1 = c2 and l = r. A weighted dissimilarity measure taking into
account two components, the center and the spread, is also adopted in this case.
As for the numerical data case, since the dissimilarity (14) is based on the Euclidean distance, it does not allow us
to recognize non-spherical shape clusters. To overcome this limitation, a generalization of dw2 (·, · ) is introduced in Ramos-

Guajardo and Ferraro (2020). It entails the cluster covariance matrices and is defined as
2
dM,w (xi , xi ) = wC2 [dM
2
( c 1 i , c 1 i ) + dM
2
(c2i , c2i )] + w2S [dM
2
(li , li ) + dM
2
(ri , ri )], (15)
where dM (·, · ) is the usual Mahalanobis distance.
In the case of outliers, as stated in Section 2, the fuzzy k-means type methods fail due to the unit-sum constraints of the
membership degrees. When dealing with fuzzy data, we have to face with three kinds of outliers: outliers with respect to
centers (location), outliers with respect to spreads (imprecision/size) and outliers with respect to both centers and spreads
(location and imprecision/size).
Robust clustering methods are provided in D’Urso and de Giovanni (2014). The authors introduce a generalization of the
FkMed Krishnapuram et al. (2001) to the case of fuzzy data (FkMed-F), by using the dissimilarity dw 2 (·, · ) Coppi et al. (2012).

This is the starting point of the other three robust proposals. The first one is based on the following ’robust’ (squared)
distance measure:
 
−exp (xi , xi ) = 1 − exp −γ dw (xi , xi ) ,
2 2
dw (16)
where γ is a positive constant determined according to the variability of the data. It is an extension of the distance intro-
duced in Wu and Yang (2002). The method is called the Smoothed Fuzzy k-Medoids for Fuzzy Data. The second proposal
deals with outliers by means of the noise cluster Davé (1991) whilst the third one is a Trimmed Fuzzy k-Medoids for Fuzzy
Data. In addition, a Fuzzy k-Medoids for fuzzy data based on a combination of Huber’s M-estimator and Yager’s ordered
weighted averaging operators is provided in D’Urso and Leski (2020).
Further robust proposals for fuzzy data are addressed in Hung et al. (2010) and in Zarandi and Razaee (2011).
As for the numerical data, also in the case of fuzzy data, by relaxing the unit-sum constraints of the membership degrees
we obtain the Possibilistic k-Means clustering method for Fuzzy data (PkM-F). In Ferraro and Giordani (2017) the PkM-F has
been defined by using the distance dw (·, · ). There exists another possibilistic clustering method for LR fuzzy data proposed
in Coppi et al. (2012) taking inspiration from Yang and Wu (2006). The two proposals can be formulated in the same way
except for the second term of the cost function.
In order to avoid the coincident cluster problem Barni et al. (1996), a possibilistic clustering method with repulsion con-
straints for symmetric triangular fuzzy data is developed in Ferraro and Giordani (2013b). A different strategy for preventing
coincident clusters is the hybridization of the fuzzy and possibilistic approaches Ferraro and Giordani (2017). The last pro-
posal exploits the benefits of both approaches, fuzzy and possibilistic. On the one hand, the fuzzy approach is helpful to
find the best fuzzy partition. On the other hand, the possibilistic one helps us to identify outliers.
Finally, a mention should be made of the proposals of fuzzy clustering of fuzzy data based on hypothesis tests Gonzalez-
Rodriguez et al. (2009); Giordani and Ramos-Guajardo (2016).

117
M.B. Ferraro Econometrics and Statistics 30 (2024) 110–123

4.1. Fuzzy k-Means for interval data and its variants

In several contexts, we face with interval data: fluctuations, ranges, grouped data, among others. An interval I is for-
malized by [infI , supI ]. Intervals can be seen as a particular case of LR fuzzy sets. In particular, when l = r = 0, c1 and c2
represent the infimum and the supremum of the interval, respectively. An interval can also be formalized by means of the
mid, (c1 + c2 )/2, and the spread (c2 − c1 )/2. The clustering methods reviewed in Section 4 can be obviously used for inter-
val data. Some of them are previously introduced for this kind of data and then generalized to the case of fuzzy data. In a
multidimensional setting, in the case of interval data, each observation is represented by a hyperrectangle in R p .
In Gao et al. (20 0 0) an FkM clustering algorithm for interval-valued (and fuzzy-valued) data is proposed. In particular,
the authors propose to preprocess the data by a feature mapping technique.
A robust FkM clustering method for interval data is provided in D’Urso and Giordani (2006a). The optimization problem
of the FkM of interval data is analogous to that in (13) except for the dissimilarity measure. Furthermore, the authors adopt
the noise cluster approach Davé (1991) to make robust the clustering algorithm and to take into account the presence
of outliers. The same dissimilarity measure is also used in D’Urso et al. (2015). The authors propose two fuzzy clustering
methods. The first one is the FkMed for interval-valued data, a timid kind of robustification. The second one is a more robust
proposal: a trimmed FkMed algorithm.
For the sake of completeness, intervals can also be seen as a special case of symbolic data Bock and Diday (20 0 0). There
exist several proposals of fuzzy k-means for symbolic interval data (see, for example, de Carvalho (2007); de Carvalho and
Tenorio (2010); Pimentel and de Souza (2014); de Carvalho and Simoes (2017)).

4.2. FkM-F: a real-case study

This section is devoted to an application of the FkM algorithm for fuzzy data. In particular, the data refers to a survey
about the personal evaluation of a set of n = 27 students regarding different aspects of 9 specialized courses on Soft Com-
puting which has been carried out in the European Centre for Soft Computing (Mieres, Spain) in 2008. The courses are:
“Soft-Computing: A History of an Interdisciplinary Field”, “Fuzzy Set Theory-Fuzzy Systems”, “Neural Networks and Neuro-
Fuzzy Systems”, “Evolutionary Computation and Genetic Fuzzy Systems”, “Probability and Statistics for Soft-Computing”,
“Fuzzy Classification and Ensembles”, “Regression and System Modeling”, “Time Series”, “Frequent Item-Set Mining”. The
students are asked to answer the questions of the survey by using trapezoidal fuzzy sets. Here we focus our attention on
the overall rating of each course.
We partition the data into two clusters. By inspecting the membership degree matrix (not reported here for the sake
of brevity), we can note that a cluster is composed by 9 students and the remaining 18 students belong to the other one.
There are 3 students, 1 in Cluster 1 and 2 in Cluster 2, with a membership degree around 0.6, thus they have intermediate
characteristics between the two clusters. Analyzing the following values of the prototypes, we can characterize the obtained
partition:
 
C1
53.73 63.96 49.24 72.68 57.43 63.37 67.50 74.41 55.03
H =
67.70 72.12 72.43 81.81 76.31 70.39 75.15 84.24 74.74
 
C2 57.63 69.72 55.82 82.16 62.26 70.50 72.23 81.24 61.19
H =
75.50 78.45 77.53 88.35 80.67 77.91 82.75 90.09 81.66
 
10.82 8.28 7.92 8.35 8.67 10.82 8.57 9.08 9.64
HL =
7.37 7.51 8.44 7.17 7.84 7.54 7.26 7.26 6.66
 
R
13.01 9.79 8.15 8.31 9.01 11.66 9.56 8.52 11.12
H =
6.99 6.49 7.44 6.28 7.56 7.64 7.03 6.05 5.89
In particular, the overall evaluation of the courses expressed by students in Cluster 1 are, in general, lower than those
of students in Cluster 2. Looking at the values of the spreads (stored in HL and HR ) of the prototypes, we can note that the
evaluations of students in Cluster 1 are more imprecise.

5. Fuzzy k-Means for functional data

In the last decades, functional data analysis has received a great deal of attention. Functional data encounter a complexity
that is not easy to manage. The observations are supposed to be functions (on a continuous domain such as time or space),
but in practice, the sampled curves are observed on a finite set of points. The usual methods for multivariate data are
not suitable for functional data but can be applied to discrete measurements. In D’Urso et al. (2018), for example, some
fuzzy clustering algorithms are proposed for multivariate time-varying data, corresponding to discrete time data instead of
functional data. On the other hand, there also exist proposals of fuzzy clustering methods for times series (see, for example,
D’Urso et al. (2017)). Although conceptually different, a time series can be seen as a realization of a functional random

118
M.B. Ferraro Econometrics and Statistics 30 (2024) 110–123

element: namely, when the time series generation process, involving concepts such as autocorrelation, is not considered,
they are just (sampled) functions defined over a time domain, and clustering methods developed for functional data can
be applied and conversely. In contrast, when the clustering method regards the time series generation process, it cannot be
applied for general functional data.
Most of the clustering proposals for functional data are based on dissimilarities that do not depend on time t. So, the
time dependency is neglected and an L2 distance measure is used. Some dissimilarities for functional data are introduced
in Tokushige et al. (2002) and also in Ferraty and Vieu (2006). In this case, the FkM method and its variants can be easily
adapted. On the other hand, functional data can be transformed and then the usual fuzzy clustering algorithms can be
employed. This is, for example, the case of the proposals in Di Maio et al. (2011) and Giordani et al. (2020). In the first
one, the FkM algorithm is used for clustering functional principal components of selected relevant process variables. In
Giordani et al. (2020) the authors introduce an FkMed method for functional data based on two steps. In the first one,
the functions are fitted to the observed data by means of B-splines. Secondly, the usual FkMed algorithm is applied to the
B-spline coefficients obtained in the previous step.
A different approach is provided in Tokushige et al. (2007). In detail, the dissimilarity between functional data is consid-
ered as a function. The authors address crisp and fuzzy k-means clustering algorithms for multivariate functional data. The
aim is to partition n objects, which are represented as a p-vector of continuous functions of t ∈ [a, b],

xi (t ) = (xi1 (t ), xi2 (t ), · · · , xip (t )), ( i = 1, · · · , n ), (17)

into k clusters. The optimization problem is
 b
n 
k

ig d (xi (t ), hg (t ))dt,
um 2
min JFkM−Func =
U,H a
i=1 g=1

s.t. uig (t ) ∈ [0, 1], i = 1, . . . , n, g = 1, . . . , k,

 b
k
1
uig (t )dt = 1, i = 1, . . . , n,
b−a a g=1

where hg (t ) is the centroid of the g-th cluster. By minimize the loss function, the obtained cluster prototypes

hg (t ) = (hg1 (t ), hg2 (t ), · · · , hgp (t )) (g = 1, · · · , k, t ∈ [a, b] )

and the membership degrees

uig (t ) (i = 1, · · · , n, g = 1, · · · , k, t ∈ [a, b] )
are defined as functions of t ∈ [a, b].

5.1. FkM-Func: a real-case study

This subsection includes the application of the FkM-Func algorithm to the Velib dataset. It contains data from the bike
sharing system of Paris. The data are loading profiles of the bike stations over one week. The data were collected every hour
during the period from Sunday, September 1, 2014, to Sunday, September 7, 2014. The data consists of the loading profiles
(number of available bikes / number of bike docks) of the 1189 stations at 181 time points. We use the dataset velib
contained in the R package funFEM Bouveyron et al. (2014)
The functions are transformed by means of B-splines and then the FkM algorithm is applied to B-spline coefficients. The
optimal number of clusters according to the fuzzy silhouette index is k = 4. The obtained partition is composed of clusters
of size 446, 266, 221 and 256, respectively. The 4 clusters and the corresponding centroids (bold line) are represented in
Figure 4.
The percentage of unclear assignments (object assigned with a membership degree lower than 0.5) is 2.19%. The mem-
bership degree plays an important role in recognizing objects (curves) with intermediate characteristics between two clus-
ters. For example, in Figure 5 are reported two objects (black solid line and dashed red line) whose membership degrees to
Cluster 1 are 0.497 and 0.497, and to Cluster 3 0.413 and 0.428, respectively.

6. Fuzzy Double k-Means

Double clustering, also known as biclustering, two-mode clustering or co-clustering, consists in simultaneously clustering
modes (e.g., units, variables) of a two-mode data matrix. Let X be a data matrix of order (n × p), the objective of a double
clustering algorithm is to simultaneously partition n rows (e.g., units) into k clusters and p columns (e.g., variables) into c
clusters.
Starting from the Double k-Means Vichi et al. (2001), the fuzzy version has been briefly introduced in Ferraro et al.
(2015) and deeply analyzed in Ferraro et al. (2021). The Fuzzy Double k-Means (FDkM) algorithm consists in solving the

119
M.B. Ferraro Econometrics and Statistics 30 (2024) 110–123

Fig. 4. Velib dataset: Cluster 1 (top left), Cluster 2 (top right), Cluster 3 (bottom left) and Cluster 4 (bottom right), obtained with FkM-Func. The bold
curves refer to the corresponding centroids.

following constrained minimization problem:


n 
p 
k 
c  2
JF DkM = xi j − hg f um vl ,
ig j f
i=1 j=1 g=1 f =1
(18)

k 
c
s.t. uig , v j f ∈ [0, 1], uig = 1, v j f = 1,
g=1 f =1

where U = [uig ] is the (n × k ) membership degree matrix for the rows, V = [v j f ] is the ( p × c ) membership degree matrix
for the columns, and H = [hg f ] is the prototype matrix of order (k × c ). In a two-mode setting, the prototypes have size
depending on the numbers of clusters for rows and columns. In this way, the roles of rows and columns are interchangeable.
The parameters m > 1 and l > 1 tune the fuzziness of the two partitions. The solution of (18) leads to a decomposition of
the data matrix into kc blocks.
The FDkM algorithm includes some special cases. When m and l tend to 1, the FDkM solution approaches the DkM one.
When c = p, the columns are not partitioned and FDkM reduces to the FkM algorithm. Furthermore, if c = p and m tends to
1, it corresponds to the standard k-Means.
In Ferraro et al. (2021), a more general Fuzzy Double k-means algorithm with polynomial fuzzifiers is also addressed,
and robust versions are illustrated. The robustness is covered through the noise cluster approach, but we have to stress that,
in this case, the structure is more complex, so three noise clusters are considered.
There also exist in literature proposals of fuzzy double clustering for categorical datasets. In this case, the data are stored
in tables, whose rows and columns are the units and the categories of all the variables, respectively. These tables are known
as cross-classification tables, contingency tables or in general co-occurrence matrices. In Oh et al. (2001) the fuzziness is
represented by an entropy regularization. The constraints on the unit membership degrees and on the category ones are
different. For each unit, the constraint is the usual one: the sum of its membership degrees to all the clusters is equal to
1. On the other hand, for each cluster, the sum of the membership degrees of all the categories to this cluster have to be
equal to 1. This leads to optimize the loss function when only one variable in each cluster is completely relevant and the
remaining ones are irrelevant. Hence, the proposal can be seen as a variable selection procedure rather than a clustering
of variables. In order to overcome numerical instabilities of the above method in the presence of large numbers of units
and categories, a further proposal is introduced in Kummamuru et al. (2003). Finally, in Tjhi and Chen (2005), a single term
fuzzifier is used in the optimization problem.

120
M.B. Ferraro Econometrics and Statistics 30 (2024) 110–123

Fig. 5. Velib dataset: two curves with unclear assignment (black solid line and dashed red line) and centroids of Cluster 1 and Cluster 3 (bold solid lines).

7. Concluding remarks and future directions

Starting from the FkM method, first several variants together with robust proposals for standard object numerical data
have been reviewed. Extensions for different kinds of data are described. In particular, the case of categorical and mixed
data are considered. Then, more complex data structures, such as fuzzy, interval or functional data, are encountered. The
last part is devoted to double clustering of heterogeneous datasets characterized by blocks of rows and columns. It consists
in simultaneously clustering rows and columns. Even if there are huge amount of papers on this topic, there are still several
open problems to deal with.
Data grow so large and complex that it becomes crucial to adapt the existing method to them. On the one hand, the
algorithms have to take into account memory limitations. In a classical clustering setting, a memory-efficient k-means al-
gorithm is implemented in the R package biganalytics Emerson et al. (2020) (related to R package bigmemory Kane et al.
(2013)). It would be interesting to address a fuzzy version of it. On the other hand, the clustering methodology have to
be adapted to the intrinsic characteristics of the data. In terms of data complexity, there is increasing interest, for exam-
ple, in network data. Networks represent a powerful model to describe problems and applications in various fields, such
as economics, biology and technology, among others. Networks can be formalized in several ways and, depending on the
formalization, extensions of FkM type algorithms can be introduced.
Furthermore, the complexity of a structure also refers to heterogeneous two-mode or multi-mode datasets, whose fea-
tures are both categorical and numerical ones. In this connection, some proposals could be found in Vichi et al. (2007);
Papalexakis et al. (2013). It would be very useful to provide two-mode or three-mode FkM type algorithms.

References

Auephanwiriyakul, S., Keller, J.M., 2002. Analysis and efficient implementation of a linguistic fuzzy c-means. IEEE Transactions on Fuzzy Systems 10,
563–582.
Babuska, R., Van der Veen, P.J., Kaymak, U., 2002. Improved covariance estimation for gustafson-kessel clustering. In: Proceedings of the 2002 IEEE interna-
tional conference on fuzzy systems, pp. 1081–1085.
Barni, M., Cappellini, V., Mecocci, A., 1996. Comments on ‘a possibilistic approach to clustering’. IEEE Transactions on Fuzzy Systems 4, 393–396.
Bezdek, J.C., 1974. Cluster validity with fuzzy sets. Journal of Cybernetics 3, 58–73.
Bezdek, J.C., 1981. Pattern Recognition with Fuzzy Objective Function Algorithm. Plenum Press, New York.
Bock, H.H., 2007. Clustering methods: a history of k-means algorithms. In: Selected contributions in data analysis and classification, pp. 161–172.
Bock, H.H., Diday, E., 20 0 0. Analysis of symbolic data. Exploratory methods for extracting statistical information from complex data. Springer-Verlag, Hei-
delberg.
Bouveyron, C., Come, E., Jacques, J., 2014. The discriminative functional mixture model for the analysis of bike sharing systems. Preprint HAL n.01024186,
University Paris Descartes.
Campello, R.J.G.B., Hruschka, E.R., 2006. A fuzzy extension of the silhouette width criterion for cluster analysis. Fuzzy Sets and Systems 157, 2858–2875.
De Carvalho, F.A.T., 2007. Fuzzy c-means clustering methods for symbolic interval data. Pattern Recognition Letters 28, 423–437.
De Carvalho, F.A.T., Simoes, E.C., 2017. Fuzzy clustering of interval-valued data with city-block and hausdorff distances. Neurocomputing 266, 659–673.

121
M.B. Ferraro Econometrics and Statistics 30 (2024) 110–123

De Carvalho, F.A.T., Tenorio, C.P., 2010. Fuzzy k-means clustering algorithms for interval-valued data based on adaptive quadratic distances. Fuzzy Sets and
Systems 161, 2978–2999.
Cebeci, Z., 2019. Comparison of internal validity indices for fuzzy clustering. Journal of Agricultural Informatics 10, 1–14.
Colubi, A., 2021. Fuzzy sets and (fuzzy) random sets in Econometrics and Statistics. Econometrics and Statistics (in press.).
Coppi, R., D’Urso, P., Giordani, P., 2012. Fuzzy and possibilistic clustering for fuzzy data. Computational Statistics & Data Analysis 56, 915–927.
Davé, R.N., 1991. Characterization and detection of noise in clustering. Pattern Recognition Letters 12, 657–664.
Davé, R.N., Sen, S., 2002. Robust fuzzy clustering of relational data. IEEE Transactions on Fuzzy Systems 10, 713–727.
Di Maio, F., Secchi, P., Vantini, S., Zio, E., 2011. Fuzzy c-means clustering of signal functional principal components for post-processing dynamic scenarios of
a nuclear power plant digital instrumentation and control system. IEEE Transactions on Reliability 60 (2).
Dunn, J., 1973. A fuzzy relative of the isodata process and its use in detecting compact well-separated clusters. Cybernetics and Systems 3, 32–57.
D’Urso, P., Giordani, P., 2006a. A robust fuzzy k-means clustering model for interval valued data. Computational Statistics 21, 255–269.
D’Urso, P., Giordani, P., 2006b. A weighted fuzzy c-means clustering model for fuzzy data. Computational Statistics & Data Analysis 50, 1496–1523.
D’Urso, P., de Giovanni, L., 2014. Robust clustering of imprecise data. Chemometrics and Intelligent Laboratory Systems 136, 58–80.
D’Urso, P., de Giovanni, L., Massari, R., 2015. Trimmed fuzzy clustering for interval-valued data. Advances in Data Analysis and Classification 9, 21–40.
D’Urso, P., de Giovanni, L., Massari, R., 2018. Robust fuzzy clustering of multivariate time trajectories. International Journal of Approximate Reasoning 99,
12–38.
D’Urso, P., Leski, J.M., 2020. Fuzzy clustering of fuzzy data based on robust loss functions and ordered weighted averaging. Fuzzy Sets and Systems 389,
1–28.
D’Urso, P., Maharaj, E.A., Alonso, A.M., 2017. Fuzzy clustering of time series using extremes. Fuzzy Sets and Systems 318, 56–79.
Emerson, J. W., Kane, M. J., & Chandra, S. (2020). biganalytics: Utilities for ’big.matrix’ objects from package ’bigmemory’. Version: 1.1.21 https://cran.
r-project.org/web/packages/biganalytics/index.html.
Ferraro, M.B., Giordani, P., 2013a. A new fuzzy clustering algorithm with entropy regularization. In: Proceedings of the 9th meeting of the classification and
data analysis group, pp. 195–198.
Ferraro, M.B., Giordani, P., 2013b. On possibilistic clustering with repulsion constraints for imprecise data. Information Sciences 245, 63–75.
Ferraro, M.B., Giordani, P., 2015. A toolbox for fuzzy clustering using the r programming language. Fuzzy Sets and Systems 279, 1–16.
Ferraro, M.B., Giordani, P., 2017. Possibilistic and fuzzy clustering methods for robust analysis of non-precise data. International Journal of Approximate
Reasoning 88, 23–38.
Ferraro, M.B., Giordani, P., 2019. A review and proposal of (fuzzy) clustering for nonlinearly separable data. International Journal of Approximate Reasoning
115, 13–31.
Ferraro, M.B., Giordani, P., 2020. Soft clustering. Wiley Interdisciplinary Reviews: Computational Statistics 12, e1480.
Ferraro, M.B., Giordani, P., Serafini, A., 2019. fclust: an r package for fuzzy clustering. The R Journal 11, 205–233.
Ferraro, M.B., Giordani, P., Vichi, V., 2021. A class of two-mode clustering algorithms in a fuzzy setting. Econometrics and Statistics 18, 63–78.
Ferraro, M.B., Vichi, M., Grzegorzewski, P., Gagolewski, M., Hryniewicz, O., Gil, M., 2015. Fuzzy double clustering: a robust proposal. In: Advances in intelli-
gent systems and computing 315. Springer, Cham., pp 225–232.
Ferraty, F., Vieu, P., 2006. Non parametric functional data analysis. Theory and Practice. Springer, Berlin.
Fritz, H., Garcia-Escudero, L.A., Mayo-Iscar, A., 2013. Robust constrained fuzzy clustering. Information Sciences 245, 38–52.
Gao, X., Ji, H., Xie, W., 20 0 0. A novel FCM clustering algorithm for interval-valued data and fuzzy-valued data. in: 5th international conference on signal
processing proceedings. In: 16th world computer congress 20 0 0, beijing, china, pp. 1551-1555.
Gath, I., Geva, A.B., 1989. Fuzzy clustering for the estimation of the parameters of the components of mixtures of normal distributions. Pattern Recognition
Letters 9, 77–86.
Giordani, P., Perna, S., Bianchi, A., Pizzulli, A., Tripodi, S., Matricardi, P.M., 2020. A study of longitudinal mobile health data through fuzzy clustering methods
for functional data: The case of allergic rhinoconjunctivitis in childhood. In: PLos one 15, e0242197.
Giordani, P., Ramos-Guajardo, A.B., 2016. A fuzzy clustering procedure for random fuzzy sets. Fuzzy Sets and Systems 305, 54–69.
Gonzalez-Rodriguez, G., Colubi, A., D’Urso, P., Montenegro, M., 2009. Multi-sample test-based clustering for fuzzy random variables. International Journal of
Approximate Reasoning 50, 721–731.
Gustafson, D.E., Kessel, W.C., 1979. Fuzzy clustering with a fuzzy covariance matrix. In: Proceedings of the 1978 IEEE conference on decision and control
including the 17th symposium on adaptive processes, pp. 761-766.
Hathaway, R.J., 1986. Another interpretation of the EM algorithm for mixture distributions. Statistics & Probability Letters 4, 53–56.
Hathaway, R.J., Bezdek, J.C., 1994. NERF c-means: Non-euclidean relational fuzzy clustering. Pattern Recognition 27, 429–437.
Hathaway, R.J., Bezdek, J.C., Devenport, J.W., 1996. On relational data versions of c-means algorithms. Pattern Recognition Letters 17, 607–612.
Hathaway, R.J., Devenport, J.W., Bezdek, J.C., 1989. Relation duals of the c-means clustering algorithms. Pattern Recognition 22, 205–212.
Huang, Z., 1997. Clustering large data sets with mixed numeric and categorical values. In: Proceedings of the first pacific asia knowledge discovery and data
mining conference. World Scientific, Singapore, pp. 21–34.
Huang, Z., 1998. Extensions to the k-means algorithm for clustering large data sets with categorical values. Data Mining and Knowledge Discovery 2,
283–304.
Huang, Z., Ng, M.K., 1999. A fuzzy k-modes algorithm for clustering categorical data. IEEE Transactions on Fuzzy Systems 7 (4).
Hung, W., Yang, M., Lee, E., 2010. A robust clustering procedure for fuzzy data. Computers & Mathematics with Applications 60, 151–165.
Hung, W.-L., Yang, M.S., 2005. Fuzzy clustering on LR-type fuzzy numbers with an application in taiwanese tea evaluation. Fuzzy Sets and Systems 150,
561–577.
Ji, J., Pang, W., Zhou, C., Han, X., Wang, Z., 2012. A fuzzy k-prototype clustering algorithm for mixed numeric and categorical data. Knowledge-Based Systems
30, 129–135.
Kane, M.J., Emerson, J., Weston, S., 2013. Scalable strategies for computing with massive data. Journal of Statistical Software 55, 1–19.
Kaufman, L., Rousseuw, P.J., 1990. Finding Groups in Data: An Introduction to Cluster Analysis. Wiley, New York.
Kim, D.-W., Lee, K.H., Lee, F., 2004. Fuzzy clustering of categorical data using fuzzy centroids. Pattern Recognition Letters 25, 1263–1271.
Klawonn, F., Höppner, F., 2003. An alternative approach to the fuzzifier in fuzzy clustering to obtain better clustering. In: Proceedings of eusflat conference,
pp. 730–734.
Klawonn, F., Kruse, R., Winkler, R., 2015. Fuzzy clustering: more than just fuzzification. Fuzzy Sets and Systems 281, 272–279.
Krishnapuram, R., Joshi, A., Nasraoui, O., Yi, L., 2001. Low-complexity fuzzy relational clustering algorithms for web mining. IEEE Transactions on Fuzzy
Systems 9, 595–607.
Krishnapuram, R., Keller, J.M., 1993. A possibilistic approach to clustering. IIEEE Transactions on Fuzzy Systems 1, 98–110.
Krishnapuram, R., Keller, J.M., 1996. The possibilistic c-means algorithm: insights and recommendations. IEEE Transactions on Fuzzy Systems 4, 385–393.
Kummamuru, K., Dhawale, A., Krishnapuram, R., 2003. Fuzzy co-clustering of documents and keywords. In: Proceedings of the 12th IEEE international
conference on fuzzy systems, 2003, pp. 772–777.
Li, R.-P., Mukaidono, M., 1995. A maximum-entropy approach to fuzzy clustering. In: Proceedings of 1995 IEEE international conference on fuzzy systems,
pp. 2227-2232.
Li, R.-P., Mukaidono, M., 1999. Gaussian clustering method based on maximum-fuzzy-entropy interpretation. Fuzzy Sets and Systems 102, 253–258.
MacQueen, J., 1967. Some methods for classification and analysis of multivariate observations. Proceedings of the Fifth Berkeley Symposium on Mathematics,
Statistics and Probability 1, 281–298.

122
M.B. Ferraro Econometrics and Statistics 30 (2024) 110–123

Nayak, J., Naik, B., Behera, H., 2015. Fuzzy c-means (FCM) clustering algorithm: a decade review from 20 0 0 to 2014. Computational intelligence in data
mining 2, 133–149.
Nguyen, H.T., 1978. A note on the extension principle for fuzzy sets. Journal of Mathematical Analysis and Applications 64, 369–380.
Oh, C.H., Honda, K., Ichihashi, H., 2001. Fuzzy clustering for categorical multivariate data. In: Proceedings of joint 9th IFSA world congress and 20th NAFIPS
international conference, pp. 2154–2159.
Pal, N.R., Bezdek, J.C., 1995. Cluster validity for the fuzzy c-means model. IEEE Transactions on Fuzzy Systems 3, 370–379.
Pal, N.R., Pal, K., Bezdek, J.C., 1997. A mixed c-means clustering model. In: Proceedings of FUZZ-IEEE’97, pp. 11-21.
Pal, N.R., Pal, K., Keller, J.M., Bezdek, J.C., 2005. A possibilistic fuzzy c-means clustering algorithm. IEEE Transactions on Fuzzy Systems 13, 517–530.
Papalexakis, E.E., Sidiropoulos, N.D., Bro, R., 2013. From k-means to higher-way co-clustering: Multilinear decomposition with sparse latent factors. IEEE
Transactions on Signal Processing 61, 493–506.
Pedrycz, W., Bezdek, J.C., Hathaway, R.J., Rogers, G.W., 1998. Two nonparametric models for fusing heterogeneous fuzzy data. IEEE ransactions on Fuzzy
Systems 6, 411–425.
Pelekis, N., Iakovidis, D.K., Kotsifakos, E.E., Kopanakis, I., 2008. Fuzzy clustering of intuitionistic fuzzy data. International Journal of Business Intelligence and
Data Mining 3 (1).
Pimentel, B.A., de Souza, R.M.C.R., 2014. A weighted multivariate fuzzy c-means method in interval-valued scientific production data. Expert Systems with
Applications 41, 3223–3236.
Ramos-Guajardo, A.B., Ferraro, M.B., 2020. A fuzzy clustering approach for fuzzy data based on a generalized distance. Fuzzy Sets and Systems 389, 29–50.
Ramsay, J., Silverman, B.W., 2005. Functional Data Analysis (2nd edn.). Springer, New York.
Roubens, M., 1978. Pattern classification problems and fuzzy sets. Fuzzy Sets and Systems 1, 239–253.
Saad, M.F., Alimi, A.M., 2009. Modified fuzzy possibilistic c-means. In: Proceedings of the international multiconference of engineers and computer scientists,
pp. 18–20.
Sato, M., Sato, Y., 1995. Fuzzy clustering model for fuzzy data. In: Proceedings of 1995 IEEE international conference on fuzzy systems, 2123–2128.
Takata, O., Miyamoto, S., Umayahara, K., 1998. Clustering of data with uncertainties using hausdorff distance. In: Proceedings of the 2nd IEEE international
conference on intelligence processing systems, pp. 67–71.
Takata, O., Miyamoto, S., Umayahara, K., 2001. Fuzzy clustering of data with uncertainties using minimum and maximum distances based on l1 metric. In:
Proceedings of joint 9th IFSA world congress and 20th NAFIPS international conference, pp. 2511–2516.
Timm, H., Borgelt, C., Döring, C., Kruse, R., 2004. An extension to possibilistic fuzzy cluster analysis. Fuzzy Sets and Systems 147, 3–16.
Tjhi, W.-C., Chen, L., 2005. Fuzzy co-clustering of web documents. In: Proceedings of the 2005 international conference on cyberworlds.
Tokushige, S., Inada, K., Yadohisa, H., 2002. Dissimilarity and related methods for functional data. Journal of Japanese Society of Computational Statistics 15,
319–326.
Tokushige, S., Yadohisa, H., Inada, K., 2007. Crisp and fuzzy k-means clustering algorithms for multivariate functional data. Computational Statistics 22,
1–16.
Vichi, M., Borra, S., Rocci, R., Vichi, M., Schader, M., 2001. Double k-means clustering for simultaneous classification of objects and variables. In: Advances
in classification and data analysis. studies in classification, data analysis, and knowledge organization. Springer, Berlin, Heidelberg.
Vichi, M., Rocci, R., Kiers, H.A.L., 2007. Simultaneous component and clustering models for three-way data: Within and between approaches. Journal of
Classification 24, 71–98.
Wachs, J., Shapira, O., Stern, H., Abraham, A., de Baets, B., Köppen, M., Nickolay, B., 2006. A method to enhance the ‘possibilistic c-means with repulsion’ al-
gorithm based on cluster validity index. In: Applied soft computing technologies: The challenge of complexity. advances in soft computing,34. Springer,
Berlin, Heidelberg.
Windham, M.P., 1985. Numerical classification of proximity data with assignment measures. Journal of Classification 2, 157–172.
Winkler, R., Klawonn, F., Kruse, R., 2011. Fuzzy clustering with polynomial fuzzifier function in connection with m-estimators. Applied and Computational
Mathematics 10, 146–163.
Wu, K.-L., Yang, M.S., 2002. Alternative c-means clustering algorithms. Pattern Recognition 35, 2267–2278.
Yang, M.S., Ko, C.H., 1996. On a class of fuzzy c-numbers clustering procedures for fuzzy data. Fuzzy Sets and Systems 84, 49–60.
Yang, M.S., Liu, H.H., 1999. Fuzzy clustering procedures for conical fuzzy vector data. Fuzzy Sets and Systems 106, 189–200.
Yang, M.-S., Wu, K.L., 2006. Unsupervised possibilistic clustering. Pattern Recognition 39, 5–21.
Zadeh, L.A., 1965. Fuzzy sets. Information and Control 8, 338–353.
Zarandi, M., Razaee, Z.S., 2011. A fuzzy clustering model for fuzzy data with outliers. International Journal of Fuzzy Systems 1, 29–42.
Zimmermann, H.J., 1996. Fuzzy Set Theory and its Applications (3rd edition). Kluwer Academic Publishers.

123