Omega 69 (2017) 126–140

Contents lists available at ScienceDirect

Omega
journal homepage: www.elsevier.com/locate/omega

A unicriterion analysis based on the PROMETHEE principles

for multicriteria ordered clustering$
Mohamed Ayman Boujelben
MODEOR, Institut des Hautes Études Commerciales, Université de Sfax, Route Sidi Mansour, BP 967, 3018 Sfax, Tunisia

art ic l e i nf o a b s t r a c t

Article history: We consider multicriteria clustering problems where the groups are defined by a total order. We assume
Received 14 January 2016 that the pairwise comparisons between the alternatives are expressed by a valued preference model built
Accepted 15 August 2016 by PROMETHEE. In order to analyze the characteristics of each cluster on the different criteria, we
Available online 24 August 2016
propose three concepts based on the PROMETHEE principles: the preference profile of a cluster, the
Keywords: similarity profile of a cluster and the inconsistency profile of a cluster. These notions are illustrated based
Multicriteria ordered clustering on a real application on the supplier segmentation.
PROMETHEE & 2016 Elsevier Ltd. All rights reserved.
Supplier segmentation

1. Introduction classical clustering with MCDA. It aims at obtaining homogenous

groups of actions according to several criteria. These groups called
Clustering [1], also referred to unsupervised classification, is clusters are assumed to be unknown a priori, i.e., no information is
the process of regrouping similar objects into undefined groups given on them. An interesting theoretical formulation of this
called clusters. This topic has widely been studied in data mining problem has been proposed in [13,9]. In both works, the authors
and applied in a variety of problems especially in artificial intel- distinguish three main types of problems: nominal or no-
ligence, pattern recognition, image analysis, marketing, etc. Due to relational clustering (no relations between the groups), relational
its numerous applications, many approaches have been proposed clustering (asymmetric relations on the groups) and ordered
such as the k-means and the hierarchical algorithms. As the clustering (total or partial order of the clusters). In this paper, we
objective is to obtain a partition of similar objects, distance or deal with the total ordered type.
similarity measures are generally used to quantify the similarity or Multicriteria ordered clustering can be encountered in different
the dissimilarity between the objects. applications such as the problem of customer satisfaction where
In the field of multicriteria decision aid (MCDA) [2], the the objective is to find a partition representing low, moderate and
assignment of alternatives into predefined classes is a classical high customer satisfaction groups. As stressed in De Smet et al.'s
problem encountered in several fields such as financial manage- work [10], this topic can be considered as a complementary tool
ment and economics, medicine, production systems and technical for the sorting since it allows obtaining ordered clusters and as a
diagnosis, etc. [3]. This issue, known as classification, has exten- consequence identifying for each cluster a subset of reference
alternatives that can be used later for sorting purposes. In addi-
sively been studied in the literature. Several methods have been
tion, according to this work, multicriteria ordered clustering can
developed in this context such as ELECTRE TRI [4], ELECTRE-SORT
be viewed as an additional perspective to understand the different
[5], PROAFTN [6], PAIRCLASS [7], UTADIS [8], etc. Generally, these
developed rankings especially those that give a total ranking of the
approaches address two types of problems: nominal classification
countries according to the economy, the human development
(i.e., no relations between the classes) and sorting (i.e., total or
index (HDI), the environmental performance index (EPI), etc.
partial order of the categories). In both cases, the classes are
Grouping these countries into ordered clusters can give another
assumed to be defined a priori by a set of reference alternatives
manner to interpret the results more accurately.
(limit or central profiles, typical alternatives for each class, etc.).
The analysis of each cluster characteristics on the different
Recently, several authors have addressed the problem of find-
criteria is a crucial question that should be addressed in multi-
ing a data structure from a multicriteria perspective [9–12]. This criteria ordered clustering. As far as we know, no special attention
research topic, known as multicriteria clustering, combines the has been paid to this problem despite the original approaches that
have been developed. Different points can be discussed at this
☆
This manuscript was processed by Associate Editor Doumpos. level including the preferential power or weakness of a cluster on
E-mail address: [email protected] each criterion, the criteria where the alternatives of the same

http://dx.doi.org/10.1016/j.omega.2016.08.007
0305-0483/& 2016 Elsevier Ltd. All rights reserved.
 M.A. Boujelben / Omega 69 (2017) 126–140 127

group are more or less similar, the discriminating power of the based on the comparison of their related central profiles and their
criteria and their conflicting aspects, and the inconsistency level of average net flow scores. In addition, De Smet et al. [10] developed
each cluster on each criterion. A unicriterion analysis that takes an exact algorithm for the multicriteria ordered clustering pro-
into account all these points helps certainly the decision-maker blem. This approach allows finding the best partition that mini-
(DM) to better understand the obtained partition. mizes the inconsistencies between the order of the clusters and
The objective of this work is to propose convenient measures the valued preference degrees describing the pairwise compar-
dealing with this question based on the PROMETHEE principles isons of the alternatives. A detailed presentation of this algorithm
[14]. We assume that the preference model performing the pair- is given in Section 6.2.
wise comparisons between the alternatives is built using PRO- A general modeling on the topic of multicriteria clustering was
METHEE. At first, we define the concept of preference profile of a also proposed by Meyer and Olteanu [9]. Based on the possible
cluster that allows analyzing the quality (preferential power or binary outranking relations that can be built between the alter-
weakness) of each group according to the different criteria. In natives, this work offers an algorithm that can be adapted
addition, it helps the DM to determine the criteria expressing according to the type of clustering: no-relational, relational, total
similar, independent or conflicting preferences and their related or partial ordered. For that purpose, they proposed a fitness
differentiation power. In order to analyze the similarity of each function reflecting the characteristics of each structure. For the
group on each criterion, we introduce the notion of similarity total ordered case, the fitness function was defined such that the
profile of a cluster. Finally, the third concept called the inconsistency preference relation between the clusters is transitive and pre-
profile of a cluster is used to measure the inconsistency level of ferentially consistent.
each group with regard to the total ranking of the clusters on the Finally, let us mention a recent work on this topic developed by
different criteria. An application on the supplier segmentation Boujelben and De Smet [20]. This approach extends the use of the
problem is also proposed to illustrate these different measures. k-means algorithm [1] with the concepts of DISSET method [21] in
This paper is organized as follows: in Section 2 we briefly order to obtain both precise and disjunctive ordered partitions. In
describe the main multicriteria ordered clustering approaches and such a situation, the actions can be assigned even to a pair of
the preference modeling in this field. We also introduce the pro- groups (and not only to precise clusters).
blem of interpreting the results of an ordered partition. The pro-
posed measures are then presented in Sections 3–5. Finally, the 2.2. Preference modeling
application that illustrates these measures is given in Section 6.
The preference modeling is a central component of the multi-
criteria decision analysis allowing to compare the alternatives. In
2. Multicriteria ordered clustering the field of multicriteria ordered clustering, most of the existing
approaches are based on pairwise comparisons between the set of
As its name implies, multicriteria ordered clustering is the actions to be regrouped or between a set of typical alternatives
process of regrouping alternatives into ordered groups. As men- characterizing the clusters. These comparisons can be exploited to
tioned above, this order can be total (i.e., a complete ranking of the regroup the alternatives into homogeneous groups, compare the
clusters) or partial (i.e., ordered clusters with some incomparable obtained clusters and define the complete order between them.
groups [15,16]). In what follows, let A ¼ fa1 ; a2 ,..., an g be the set of As stressed above, these comparisons are generally described
alternatives to be regrouped, G ¼ fg 1 ; g 2 ; …; g s g be the set of criteria either by binary relations representing indifference, preference or
and C ¼ fC 1 ; C 2 ; …; C k g be the set of ordered clusters defined such incomparability situations or a valued preference model that gives
as C 1 g C 2 g ⋯ g C k . The symbol g implies that cluster Cl is outranking degrees between each pair of alternatives. In both
better than cluster C l þ 1 ( 8 l ¼ 1; …; k  1). The number of groups is cases, different outranking methods such as ELECTRE III [22] and
assumed to be fixed a priori by the DM. PROMETHEE [14] can be used to construct the preference model.
The use of PROMETHEE in this context is at the core of this paper.
2.1. Brief overview of the main contributions For this reason, we give a brief description of its basic notions that
are necessary to understand the proposed developments.
To our knowledge, the first approach dealing with the ordered PROMETHEE is a convenient framework for modeling pairwise
clustering in the field of MCDA was proposed by Valls and Torra comparisons of a set of alternatives A ¼ fa1 ; a2 ; …; an g evaluated on
[17]. This method uses a traditional clustering approach to regroup a set of criteria G ¼ fg 1 ; g 2 ; …; g s g. The underlying idea of this
the alternatives based on a similarity matrix between them. The approach is to define a valued preference degree π ðai ; aj Þ for each
obtained clusters are then compared to an ideal alternative using a pair of alternatives ai and aj reflecting the preference intensity of ai
distance measure in order to deduce the order of the groups. to aj according to all the criteria. To compute this degree, this
Despite the originality of this contribution, the multicriteria nature approach starts by determining the difference dh ðai ; aj Þ between
of the problem can be viewed as poorly represented since it does the evaluations of ai and aj on each criterion gh:
not integrate the DM's preferences on the set of alternatives. The dh ðai ; aj Þ ¼ g h ðai Þ  g h ðaj Þ. These differences are then transformed
work of De Smet and Gilbart [18] was also among the first con- into unicriterion preference degrees using a preference function
tributions addressing the multicriteria ordered clustering. This P h ðai ; aj Þ ¼ f h ðdh ðai ; aj ÞÞ defined for each criterion gh. This function
approach extends the use of the PROMETHEE method [14] in order varies between 0 and 1: the greater its value, the higher the pre-
to rank groups of alternatives. However, although it takes into ference of ai over aj on gh. It can be obtained using one of the six
account the multicriteria aspect of the problem through the use of different types defined by Brans and Vincke [14]: usual, U-shape,
the notions of PROMETHEE, this method does not perform well for V-shape, level, linear and Gaussian (see Fig. 1). The use of this
large data sets [10]. preference function P h ðai ; aj Þ is the basis of the proposed measures.
Fernandez et al. [19] also proposed an algorithm to obtain a Finally, the preference degree π ðai ; aj Þ is computed by aggregating
total ordered partition of the alternatives. Based on a valued all the unicriterion preference degrees P h ðai ; aj Þ using the following
indifference relation inspired by outranking methods, this method weighted sum:
starts by regrouping the alternatives which are sufficiently similar X
s
in a preference sense and sufficiently dissimilar to the alternatives π ðai ; aj Þ ¼ wh  P h ðai ; aj Þ ð1Þ
belonging to other clusters. The obtained clusters are then ranked h¼1
128 M.A. Boujelben / Omega 69 (2017) 126–140

Fig. 1. PROMETHEE preference functions – case of criteria that have to be maximized.

where wh is the criterion weight of gh given by the DM and defined In this section, we propose the concept of preference profile of
Ps
such that wh 40 and h ¼ 1 wh ¼ 1. Obviously, the following prop- a cluster in order to analyze the preferential information contained
erties are verified: 0 r π ðai ; aj Þ r 1 and 0 r π ðai ; aj Þ þ π ðaj ; ai Þ r 1. in each group. This notion allows characterizing each cluster by
Moreover, it is not difficult to deduce that the indifference degree unicriterion flows reflecting its preferential quality. The idea
between ai and aj is equal to 1  π ðai ; aj Þ  π ðaj ; ai Þ. behind it was inspired by the GAIA visual modeling method which
is a powerful tool representing the rankings of PROMETHEE I and II
2.3. Interpreting the results of an ordered partition [24]. This approach is based essentially on the notion of uni-
criterion flows defined for each alternative. This work extends the
Interpreting clustering results is one of the most important use of these flows to characterize each cluster.
issues in the classical clustering allowing the evaluation of a given
partition. Several concepts have been developed in this context in
3.1. Unicriterion flows of a cluster
order to assess the quality of the clustering results and to make
interpretations [23]. These measures can be viewed as convenient
tools at the disposal of the DM to validate the quality of clustering Definition 3.1. The unicriterion preference index of cluster Cl to
algorithms. cluster Cf, denoted P h ðC l ; C f Þ, is the degree that measures how the
In the field of multicriteria ordered clustering, the analysis of a nl actions of cluster Cl are preferred to the nf actions of cluster Cf on
given partition is an essential issue to be explored. Since the criterion gh. It is given by:
obtained clusters are not known a priori, their evaluation is nee- P P
ded because it gives a global overview on their characteristics. This ai A C l aj A C f P h ðai ; aj Þ
P h ðC l ; C f Þ ¼ ð2Þ
paper aims at presenting some concepts allowing to interpret the nl  nf
results of an ordered partition obtained based on the PROMETHEE
principles. These notions take into account the information
underlying the order of the clusters and the DM's preferences on Definition 3.2. The unicriterion positive flow of cluster Cl, denoted
the decision alternatives. They give a unicriterion analysis on each ϕhþ ðC l Þ, is the degree that measures how Cl outranks all the other
group according to three components: the preferential quality, the clusters on criterion gh, i.e., how the nl actions of Cl are preferred to
similarity and the inconsistency. all the n  nl actions belonging to C⧹fC l g on gh: the higher its value,
The proposed measures can be considered as quality measures the better Cl on gh. It is defined as:
allowing to evaluate, interpret and further understand the results P P
a AC a A C⧹fC l g P h ðai ; aj Þ
of a given ordered partition. In addition, they can be used to ϕhþ ðC l Þ ¼ i l j ð3Þ
nl  ðn  nl Þ
compare ordered partitions obtained by different clustering
approaches. They can also be useful in the comparison of two
clustering results obtained by the same approach with different Definition 3.3. The unicriterion negative flow of cluster Cl, deno-

numbers of clusters. ted ϕh ðC l Þ, is the degree that measures how Cl is outranked by all
the other clusters on criterion gh, i.e., how the nl actions of Cl are
preferred by all the n  nl actions belonging to C⧹fC l g on gh: the
3. Preference profile of a cluster lower its value, the better Cl on gh. Formally:
P P
a AC a A C⧹fC l g P h ðaj ; ai Þ
The preferential order between the clusters is an important ϕh ðC l Þ ¼ i l j ð4Þ
feature of an ordered partition that should be analyzed. Defining nl  ðn  nl Þ
the subsets of criteria that constitute the power and the weakness
of each group is additional information that helps the DM to better
þ
understand the preferential structure of the partition, i.e., why We note that ϕh ðC l Þ allows taking into account the outranking
each group is preferred to and/or preferred by others. More spe- character of cluster Cl on criterion gh even on the clusters which

cifically, a unicriterion analysis of the preferential information are preferred to Cl. Similarly, ϕh ðC l Þ considers the outranked
underlying the ordered partition allows the DM to identify the character of Cl on gh even on the clusters which are preferred by Cl.
þ 
reasons leading to this order. Obviously, ϕh ðC l Þ and ϕh ðC l Þ belong to ½0; 1. In addition, it is easy
 M.A. Boujelben / Omega 69 (2017) 126–140 129

to verify the following formulas:  If covðϕh ðÞ; ϕr ðÞÞ is highly negative, the criteria gh and gr are
P strongly negatively correlated and thus they are conflicting
C A C⧹fC l g nf  P h ðC l ; C f Þ
ϕhþ ðC l Þ ¼ f ð5Þ because they express opposite preferences (ϕh ðÞ tends to
n  nl increase as ϕr ðÞ decreases, and vice versa). This is the case of a
P cluster which is good on criterion gh (in comparison with the
C f A C⧹fC l g nf  P h ðC f ; C l Þ
ϕh ðC l Þ ¼ ð6Þ other groups) and not good on criterion gr.
n  nl

Definition 3.4. The unicriterion net flow of cluster Cl, denoted

4. Similarity profile of a cluster
ϕh ðC l Þ, is the degree that measures the preferential quality of Cl in
þ
the partition on criterion gh. It is the balance between ϕh ðC l Þ and
 In multicriteria ordered clustering, the similarity is an impor-
ϕh ðC l Þ: tant feature characterizing the clusters. A high similarity level on
each group reflects a good quality of a given partition in terms of
ϕh ðC l Þ ¼ ϕhþ ðC l Þ  ϕh ðC l Þ ð7Þ homogeneity. Analyzing it according to each group on each cri-
terion certainly helps the DM to identify the criteria where the
alternatives of the same group are more or less similar, and
This degree takes into account the outranking and outranked therefore to make interpretations based on this information.
characters of Cl with regard to the other clusters on gh. It varies The second development in this work is devoted to the notion
between  1 and 1: the higher its value, the better the preferential of similarity of a cluster Cl on the different criteria, i.e., the simi-
quality of Cl on criterion gh. We speak about preferential power if larity that can exist between its nl alternatives. Obviously, we
ϕh ðC l Þ 4 0 and preferential weakness if ϕh ðC l Þ o 0. In addition, ϕh speak about this notion only when the cluster is composed of at
ðC l Þ allows the definition of the preference profile related to each least one pair of alternatives. The similarity of a cluster with only
cluster. one action does not make a sense. That is why we assume in this
section that that nl a 1.
Definition 3.5. The preference profile of cluster Cl, denoted In what follows, we define unicriterion similarity degrees for
PRPðC l Þ, is defined by all the values of the unicriterion net flows each pair of alternatives ai and aj belonging to the same cluster Cl.
ϕh ðC l Þ, i.e.: These measures are used to deduce unicriterion similarity degrees
for each cluster Cl and then to define the notion of similarity
PRPðC l Þ ¼ ðϕ1 ðC l Þ; ϕ2 ðC l Þ; …; ϕs ðC l ÞÞ ð8Þ profile of a cluster. Two types of similarity are distinguished: the
intra-similarity (within the cluster) and the inter-similarity
(between the cluster and the other groups).
PRPðC l Þ permits to analyze the preferential power or weakness
of Cl with regard to the other clusters on the different criteria. 4.1. Unicriterion intra-similarity of a cluster

3.2. Specific interpretations for the criteria The unicriterion intra-similarity degree of ai ; aj A C l is measured
by the indifference degree of ai and aj on criterion gh. It is given by:
The preference profiles of all the clusters can be grouped in the
SAh ðai ; aj Þ ¼ 1  P h ðai ; aj Þ  P h ðaj ; ai Þ ð9Þ
following matrix:
g1 g2 … gs This notion allows deducing the unicriterion intra-similarity
0 1 degree of Cl that measures the similarity of its nl actions within
C 1 ϕ1 ðC 1 Þ ϕ2 ðC 1 Þ … ϕs ðC 1 Þ
B C it on criterion gh.
C 2 B ϕ1 ðC 2 Þ ϕ2 ðC 2 Þ … ϕs ðC 2 Þ C
B C
B
⋮ @ ⋮ ⋮ … ⋮ C A Definition 4.1. The unicriterion intra-similarity degree of Cl,
C k ϕ1 ðC k Þ ϕ2 ðC k Þ … ϕs ðC k Þ denoted SAh ðC l Þ, is the average of all the similarity degrees SAh ðai ; aj Þ
of the nl  ðnl 1Þ=2 pairs of alternatives ai ; aj A C l :
This matrix is particularly useful to measure the discriminating P P
SAh ðai ; aj Þ
power of each criterion gh based on its related vector ϕh ðÞ. It is the
aj A C l
ai A C l aj a ai
SAh ðC l Þ ¼ ð10Þ
standard deviation of the values of ϕh ðÞ: the higher its value, the nl  ðnl 1Þ
more criterion gh differentiates between the clusters.
In addition, it is possible to deduce if two criteria gh and gr are
similar, independent or conflicting in terms of preferences based 4.2. Unicriterion inter-similarity of a cluster
on the covariance of their associated vectors ϕh ðÞ and ϕr ðÞ. This
measure allows making the following interpretations: The unicriterion inter-similarity of ai ; aj A C l can be measured
through the unicriterion outranking and outranked characters of
 If covðϕh ðÞ; ϕr ðÞÞ is highly positive, the criteria gh and gr are this pair of alternatives to all the n  nl actions belonging to the
strongly positively correlated and therefore they are similar other clusters, i.e., to all ak such that ak 2
= Cl:
because they express similar preferences (both unicriterion net
flows ϕh ðÞ and ϕr ðÞ tend to increase or decrease together). This  The unicriterion outranking inter-similarity measures the simi-
means that if a cluster is better (or worse) than the other groups larity of ai and aj on criterion gh when comparing their
on criterion gh, it is also better (or worse) than them on criterion preference power with regard to all ak 2 = C l , i.e., when comparing
gr. the unicriterion preference values P h ðai ; ak Þ and P h ðaj ; ak Þ for all
 If covðϕh ðÞ; ϕr ðÞÞ tends to zero, the criteria gh and gr are inde- ak 2
= C l . In this case, the dissimilarity
 of ai and aj with  respect to
pendent. In this case, there is no relation between the pre- ak is defined as the distance P h ðai ; ak Þ  P h ðaj ; ak Þ and therefore
ferences of criterion gh and those of criterion gr expressed on the similarity is equal to 1  P h ðai ; ak Þ  P h ðaj ; ak Þ. As a result,
the clusters. the similarity with regard to all ak 2 = C l is the average of all the
130 M.A. Boujelben / Omega 69 (2017) 126–140

similarities of ai ; aj A C l to all ak 2 = C l . More formally: degrees Sh ðC l Þ, i.e.:

P  
  SPðC l Þ ¼ ðS1 ðC l Þ; S2 ðC l Þ; …; Ss ðC l ÞÞ ð16Þ
þ = C l 1  P h ðai ; ak Þ  P h ðaj ; ak Þ
ak 2
SEh ðai ; aj Þ ¼ ð11Þ
n  nl

SPðC l Þ allows investigating the global similarity of Cl on the

 The unicriterion outranked inter-similarity measures the simi- different criteria. Moreover, it helps the DM to detect the criteria
larity of ai and aj on criterion gh when comparing their where the nl alternatives of Cl are more or less similar.
preference weakness related to all ak 2 = C l , i.e., when comparing
the unicriterion preference values P h ðak ; ai Þ and P h ðak ; aj Þ for all
ak 2
= C l . In this case, the dissimilarity
 of ai and aj with  regard to ak 5. Inconsistency profile of a cluster
is defined as the distance P h ðak ; ai Þ  P h ðak ; aj Þ and thus the
 
similarity is equal to 1  P h ðak ; ai Þ  P h ðak ; aj Þ. Consequently, the Multicriteria ordered clustering can be viewed as a process that
similarity with respect to all ak 2 = C l is the average of all the produces an ordered partition (output) based on a preference
similarities of ai ; aj A C l to all ak 2 = C l , i.e.: model (input) describing the comparisons between alternatives. Of
course, the lower the inconsistency (incompatibility) between this
P  
P h ðak ; ai Þ  P h ðak ; aj Þ input and this output, the better the partition. Therefore, analyzing
= Cl 1 
ak 2
SEh ðai ; aj Þ ¼ ð12Þ this inconsistency with regard to the order of the groups provides
n  nl
an idea on the quality of the partition. In addition, a unicriterion
analysis allows the DM to better understand the reasons of this
inconsistency according to each group.
Based on SEhþ ðai ; aj Þ and SEh ðai ; aj Þ, it is possible to compute the In this section, we propose the concept of inconsistency profile
unicriterion outranking and outranked inter-similarity degrees for of a cluster. This notion allows performing a unicriterion analysis
each cluster Cl. These two measures allow quantifying the simi- of the inconsistency of each cluster Cl with respect to the total
larity of the nl actions of Cl when comparing their outranking and order of the partition. As the similarity, the unicriterion incon-
outranked characters to all the n  nl actions of the other clusters sistency of Cl is investigated through two notions: the intra-
on criterion gh. inconsistency (the inconsistency within Cl) and the inter-
inconsistency. The latter covers two situations:
Definition 4.2. The unicriterion outranking inter-similarity degree
of Cl, denoted SEhþ ðC l Þ, is the average of all the similarity degrees
 The inconsistency between Cl and C l þ 1-k the subset of clusters
SEhþ ðai ; aj Þ of the nl  ðnl  1Þ=2 pairs of alternatives ai ; aj A C l :
that are preferred by Cl (with C l aC k ).
P P þ  The inconsistency between Cl and C 1-l  1 the subset of clusters
ai A C l aj A C l SE
h ðai ; aj Þ
þ
SEh ðC l Þ ¼
aj a ai
ð13Þ that are preferred to Cl (with C l a C 1 ).
nl  ðnl  1Þ
Let us note that this notion was inspired by De Smet et al.'s
Definition 4.3. The unicriterion outranked inter-similarity degree method [10] where the inconsistency is defined globally according
of Cl, denoted SEh ðC l Þ, is the average of all the similarity degrees to all the criteria and for each pair of alternatives. In what follows,
SEh ðai ; aj Þ of the nl  ðnl  1Þ=2 pairs of alternatives ai ; aj A C l : we define it according to each criterion and study it with regard to
P P 
each group.
ai A C l aj A C l SE
h ðai ; aj Þ
 aj a ai
SEh ðC l Þ ¼ ð14Þ 5.1. Unicriterion intra-inconsistency of a cluster
nl  ðnl  1Þ

When considering two alternatives ai and aj such that ai ; aj A C l ,

4.3. Unicriterion global similarity of a cluster ai and aj are assumed to be similar or indifferent according to each
criterion gh. Therefore, the unicriterion preference degrees P h ðai ;
This notion is a global measure that takes into account the aj Þ and P h ðaj ; ai Þ should be as low as possible. In this case, the intra-
unicriterion intra-similarity and inter-similarity degrees of cluster inconsistency degree of ai ; aj A C l on gh can be defined as:
Cl. Formally: IAh ðai ; aj Þ ¼ P h ðai ; aj Þ þ P h ðaj ; ai Þ ð17Þ
Definition 4.4. The unicriterion global similarity of cluster Cl, This notion allows deducing the unicriterion intra-inconsistency
denoted Sh ðC l Þ, is defined as the average of SAh ðC l Þ, SEhþ ðC l Þ and degree of Cl that measures the inconsistency of its nl actions within
SEh ðC l Þ: it on criterion gh. Obviously, this intra-inconsistency is equal to
0 when nl ¼1.
SAh ðC l Þ þ SEhþ ðC l Þ þ SEh ðC l Þ
Sh ðC l Þ ¼ ð15Þ
3 Definition 5.1. The unicriterion intra-inconsistency degree of Cl,
denoted IAh ðC l Þ, is the average of all the inconsistency degrees IAh
ðai ; aj Þ of the nl  ðnl  1Þ=2 pairs of alternatives ai ; aj A C l :
The operator “average” is considered in this formula (and not 8
the ”weighted sum”) because SAh ðC l Þ, SEhþ ðC l Þ and SEh ðC l Þ have the >
<P0 if nl ¼ 1
P
same weights (since each value is the average of nl  ðnl  1Þ=2 IAh ðC l Þ ¼ ai A C l
IAh ðai ;aj Þ
aj A C l ð18Þ
>
:
aj a ai
otherwise
degrees). Of course, SAh ðC l Þ, SEhþ ðC l Þ, SEh ðC l Þ and Sh ðC l Þ are num- nl ðnl  1Þ

bers between 0 and 1: the higher their values, the higher the
similarity of Cl on criterion gh. Moreover, it is possible to define the
similarity profile associated to each cluster based on Sh ðC l Þ. 5.2. Unicriterion inter-inconsistency of a cluster

Definition 4.5. The similarity profile of cluster Cl, denoted SPðC l Þ, When considering two alternatives ai and aj such that ai A C l
is defined by all the values of the unicriterion global similarity and aj A C f , one can distinguish the two following cases:
 M.A. Boujelben / Omega 69 (2017) 126–140 131

 If C l g C f , ai is considered to be better than aj on each criterion Finally, let us note that an illustrative example is given in
gh. Thus, P h ðaj ; ai Þ should be as low as possible and the inter- Appendix A in order to understand the computational details of all
inconsistency of ai to aj on criterion gh occurs when P h ðaj ; ai Þ 4 0. the preference, similarity and inconsistency measures developed
 If C l ! C f , ai is assumed to be worse than aj on each criterion gh. in this paper.
Therefore, P h ðai ; aj Þ should be as low as possible and the inter-
inconsistency of ai to aj on criterion gh occurs when P h ðai ; aj Þ 4 0.
6. Case study: the problem of supplier segmentation
These two types of inconsistencies of ai to aj on gh can be
generalized to measure the unicriterion inter-inconsistencies of Cl 6.1. Description of the problem
with regard to C l þ 1-k and C 1-l  1 . This leads to the following
definitions. The supplier segmentation is one of the main strategic activ-
ities of a firm that consists in creating groups of similar suppliers.
Definition 5.2. The unicriterion inter-inconsistency of cluster Cl
This classification, which succeeds the supplier selection, plays an
with regard to the subset of clusters C l þ 1-k , denoted
important role in the supply chain management since it allows
IEh ðC l =C l þ 1-k Þ, is the degree that measures how Cl is outranked by
choosing the most suitable strategies for handling each group
C l þ 1-k on criterion gh, i.e., how the nl actions of Cl are preferred by
differently [25]. Several approaches have been developed in this
all the nl þ 1-k actions of C l þ 1-k on gh:
8 context to deal with this problem. Kraljic's work [26] was among
>
<P 0 if C l a C k the first contributions addressing this problem based on two cri-
P
IEh ðC l =C l þ 1-k Þ ¼ ai A C l aj A C l þ 1-k P h ðaj ; ai Þ ð19Þ teria: the profit impact and the supply risk. Following this idea, a
>
: otherwise
nl  nl þ 1-k variety of bi-criteria procedures have then been proposed for the
supplier segmentation. A literature review of these approaches can
Definition 5.3. The unicriterion inter-inconsistency of cluster Cl
be found in [27]. Most of them consider four clusters of suppliers
with regard to the subset of clusters C 1-l  1 , denoted
based on two levels (low and high) for each criterion.
IEh ðC l =C 1-l  1 Þ, is the degree that measures how Cl outranks
Recently, Rezaei and Ortt [28] proposed an approach that
C 1-l  1 on criterion gh, i.e., how the nl actions of Cl are preferred to
considers more than two criteria in this problem. These criteria
all the n1-l  1 actions of C 1-l  1 on gh:
8 can be classified into two types: the supplier capabilities (price,
>
<0 if C l a C 1 delivery, quality, reserve capacity, geographical location, etc.) and
P P
IEh ðC l =C 1-l  1 Þ ¼ ai A C l aj A C 1-l  1 P h ðai ; aj Þ ð20Þ the supplier willingness (communication openness, commitment
>
: otherwise
nl  n1-l  1 to continuous improvement in product and process, etc.). The
underlying idea of this method is to convert the problem into a bi-
criteria one by aggregating all the criteria of each type into one
5.3. Unicriterion global inconsistency of a cluster criterion. The suppliers are then regrouped into four clusters:
suppliers with high capabilities and willingness (best suppliers),
This notion is a global measure that considers simultaneously suppliers with high capabilities and low willingness, suppliers
the unicriterion intra-inconsistency and inter-inconsistency with low capabilities and high willingness, and suppliers with low
degrees of cluster Cl. Formally: capabilities and willingness (worst suppliers). The order char-
acterizing this partition is partial since the second and the third
Definition 5.4. The unicriterion global inconsistency of cluster Cl,
clusters are incomparable.
denoted I h ðC l Þ is defined as the weighted sum of IAh ðC l Þ, IEh ðC l =
However, although this method considers implicitly several
C l þ 1-k Þ and IEh ðC l =C 1-l  1 Þ:
criteria, the segmentation was performed using two criteria that
w1  IAh ðC l Þ þ w2  IEh ðC l =C l þ 1-k Þ þ w3  IEh ðC l =C 1-l  1 Þ cover others. This does not allow analyzing separately the effect of
I h ðC l Þ ¼ ð21Þ
w1 þ w2 þ w3 each criterion of each type on the segmentation. Moreover, when a
supplier is good on some capabilities (or willingness) criteria and
where w1 ¼ nl  ðnl 1Þ=2, w2 ¼ nl  nl þ 1-k and w3 ¼ nl  n1-l  1 .
bad on the others, a loss of information can be induced when
The operator “weighted sum” is considered in this formula aggregating them into one criterion. The objective of this section is
because IAh ðC l Þ, IEh ðC l =C l þ 1-k Þ and IEh ðC l =C 1-l  1 Þ have not the to use a multicriteria ordered clustering approach that considers
same weight since they are the respective averages of each criterion of each type independently and to apply our
nl  ðnl  1Þ=2, nl  nl þ 1-k and nl  n1-l  1 degrees (see their respec- developed measures to analyze the obtained partition on the dif-
tive formulas). Of course, IAh ðC l Þ, IEh ðC l =C l þ 1-k Þ, IEh ðC l =C 1-l  1 Þ ferent criteria. For that purpose, we used De Smet et al.'s method
and I h ðC l Þ vary between 0 and 1: the higher their values, the higher [10] which allows obtaining a total ranking of the clusters. PRO-
the inconsistency of Cl on criterion gh. METHEE was also applied in the preference modeling.
Now, it is possible to deduce the inconsistency profile asso-
ciated to each cluster. 6.2. De Smet et al.'s method
Definition 5.5. The inconsistency profile of cluster Cl, denoted
As mentioned in Section 2.1, De Smet et al.'s method is a
IPðC l Þ, is defined by all the values of the unicriterion global
multicriteria ordered clustering approach allowing to find the best
inconsistency degrees I h ðC l Þ, i.e.:
ordered partition that minimizes the inconsistencies with respect
to the following conditions:
IPðC l Þ ¼ ðI 1 ðC l Þ; I 2 ðC l Þ; …; I s ðC l ÞÞ ð22Þ
 Two actions ai and aj assigned to the same cluster are assumed
to be indifferent or similar. In this case, the preference degrees
IPðC l Þ is useful to analyze the global inconsistency of Cl with π ðai ; aj Þ and π ðaj ; ai Þ should be as low as possible and the
regard to the total order of the clusters on the different criteria. In inconsistency occurs when π ðai ; aj Þ 4 0 and π ðaj ; ai Þ 4 0.
addition, it permits to determine the criteria where the incon-  An action ai is considered to be better than another action aj if ai
sistency level of Cl is high or low. is assigned to a cluster more preferred than the cluster of aj. In
132 M.A. Boujelben / Omega 69 (2017) 126–140

this case, π ðaj ; ai Þ should be as low as possible and the incon- Table 2
sistency occurs when π ðaj ; ai Þ 40. Resulting supplier segmentation.

Cluster Suppliers' numbers

Based on these two assumptions, the authors proposed to
characterize the quality of any given ordered partition by an C1 1, 2, 3, 4, 5, 6, 7, 11, 12, 13, 22, 23, 26
inconsistency matrix. This matrix measures the incompatibility 27, 28, 31, 32, 33, 35, 36, 37, 40, 42
between the preference matrix π describing the pairwise com- C2 14, 17, 18, 24, 25, 29, 38, 41, 43
C3 8, 15, 16, 19, 20, 30, 34, 39
parisons of the alternatives and the order of the clusters P. For- C4 9, 10, 21
mally, the inconsistency matrix between π and P depends on the
preference values and it is defined as follows:
(
0 if ai A C l ; aj A C f ; C l g C f
Iðπ ; PÞij ¼ ð23Þ qualitative criteria (6 capabilities and 6 willingness) listed in
π ðai ; aj Þ Otherwise
Table 1. The evaluation of each supplier on each criterion is
The Iðπ ; PÞ matrix is equal to π except where the preference determined based on 5 assessment levels: 1 very low, 2 low, 3
degrees between actions are compatible with the partition. For average, 4 high and 5 very high. Appendix B gives the evaluations of
those elements, the value is equal to 0. all the suppliers on the different capabilities and willingness cri-
Following this intuition, an exact algorithm was proposed to teria. The criteria weights are determined using the fuzzy AHP
identify the best ordered partition based on a lexicographic com- method [29] as shown in Table 1. In order to apply PROMETHEE,
parison of the inconsistencies matrices. A partition P is better than the level type is used to define the preference function of each
another partition P 0 if and only if Iðπ ; PÞ g L Iðπ ; P 0 Þ (where g L criterion. According to Brans and Mareschal [24], this type is well
denotes the lexicographical order defined on the elements of the suited for qualitative criteria. Moreover, four clusters are con-
inconsistency matrices). This algorithm considers the ordinal sidered to regroup the suppliers: C1 very high suppliers' perfor-
information contained in π from the maximal to the minimal and mance, C2 high suppliers' performance, C3 medium suppliers' per-
starts by testing the highest value and so on. Every time that a new formance and C4 low suppliers' performance.
value π ðai ; aj Þ is considered, we test if putting ai in a better cluster Table 2 gives the obtained ordered partition and the suppliers'
than aj creates a cycle or a path longer than k  1 in the induced numbers of each cluster. As can be seen, 23 suppliers are assigned
partition (where k is the number of clusters): to C1, 9 suppliers are assigned to C2, 8 suppliers are assigned to C3
and 3 suppliers are assigned to C4. This means that 53.48% of the
 If none of these two conditions is verified, the value π ðai ; aj Þ is suppliers have a very high performance.
compatible with the partition and therefore Iðπ ; PÞij ¼ 0.
 If at least one condition is verified, ai cannot be in a better 6.3.1. Analysis of the preference profiles
cluster than aj without creating a cycle and/or a path longer
In order to understand more these results and to analyze the
than k  1 in the induced partition. As a consequence,
characteristics of each cluster on the different criteria, we deter-
Iðπ ; PÞij ¼ π ðai ; aj Þ.
mined the preference profile of each group that allows measuring
its preference quality with regard to the other clusters. Appendix C
The clusters are deduced based on the ranks of the graph repre-
senting all the induced preference relations between the gives the preference profiles of all the groups, i.e., their related
alternatives. unicriterion net flow values on the different criteria (the strict
positive values are in bold). Figs. 3–6 illustrate graphically all these
profiles. The red and blue colors are used respectively to distin-
6.3. Application and results
guish the capabilities and willingness criteria. The values of the
Let us consider the real-case study presented in Rezaei and unicriterion flows, that vary between 1 and 1, are visualized in a
Ortt's work [28]. A manager of a medium-sized broiler company descending order. As can be noticed:
wants to determine the segmentation of 43 suppliers that provide Cluster C1: The suppliers of this cluster are the best for the
the company with the newly hatched chicks (11 suppliers), feed (9 company. The unicriterion net flows of all the criteria are strictly
suppliers), medications (6 suppliers) and the required materials positive. Therefore, these suppliers are preferred to the others on
(17 suppliers). These suppliers are evaluated according to 12 all the capabilities and willingness criteria. However, this pre-
ferential power is not very high. Indeed, all the unicriterion net
flow values do not exceed 0.387.
Table 1 Cluster C2: The suppliers of this group are preferred to the
Criteria and their related weights. others on four willingness criteria: “Reciprocal arrangement”,
“Communication openness”, “Long term relationship” and “Will-
Criterion Weight
ingness to share information”. However, this preferential power on
Capabilities criteria g1 – price 0.0558 these criteria is not high (the values of their respective net flows
g2 – delivery 0.0809 are 0.2081, 0.1397, 0.0577 and 0.024). Moreover, these values are
g3 – quality 0.1003
less than those related to cluster C1 which are equal respectively to
g4 – reserve capacity 0.0805
g5 – geographical location 0.0809 0.2152, 0.2609, 0.2609 and 0.387. For the other criteria, the uni-
g6 – financial position 0.1015 criterion net flow values are strictly negative. They constitute the
weakness of this cluster in comparison with the others.
Willingness criteria g7 – commitment to quality 0.1020
g8 – communication openness 0.0765
Cluster C3: The suppliers of this group are preferred to the
g9 – reciprocal arrangement 0.0755 others on three capabilities criteria: “Reserve capacity”, “Financial
g10 – willingness to share information 0.0755 position” and “Price”. However, this preferential power on these
g11 – supplier's effort in promoting JIT principles 0.1005
criteria is low. Indeed, the values of their respective net flows are
g12 – long term relationship 0.0700
0.1223, 0.0083 and 0.0036:
 M.A. Boujelben / Omega 69 (2017) 126–140 133

Fig. 2. Discriminating power of all the criteria. Fig. 5. Preference profile of cluster C3.

Fig. 6. Preference profile of cluster C4.

Fig. 3. Preference profile of cluster C1.
 On criteria “Financial position” and “Price”: although these are
strictly positive (0.0083 and 0.0036), they are less than those
related to cluster C1 which are equal respectively to 0.3145 and
0.1348.

For the other criteria, the unicriterion net flow values are strictly
negative. They constitute the weakness of this cluster in compar-
ison with the others.
Cluster C4: The suppliers of this cluster are the worst for the
company. The unicriterion net flows of all the criteria are strictly
negative. Therefore, these suppliers are preferred by the others on
all the capabilities and willingness criteria especially on the fol-
lowing willingness criteria: “Willingness to share information”,
“Long term relationship”, “Communication openness” and “Reci-
procal arrangement”. Indeed, the unicriterion net flow values
related to these four criteria are less than or equal to  0.5.
Based on the preference profiles of all the clusters, it is also
possible to deduce the discriminating power of each criterion gh by
computing the standard deviation of the vector ϕh ðÞ representing
its related unicriterion net flows of all the clusters. Fig. 2 gives in a
Fig. 4. Preference profile of cluster C2.
descending order the discriminating power of all the criteria (the
red color for the capabilities and the blue color for the will-
 On criterion “Reserve capacity”: the value 0.1223 is the highest ingness). It is easy to observe that the highest differentiation
one in comparison to those related to the other clusters on this power is for criterion “Willingness to share information”. This
criterion (even to cluster C1 (0.0826)). means that it discriminates the clusters more than the other
134 M.A. Boujelben / Omega 69 (2017) 126–140

Fig. 7. Similarity profile of cluster C1. Fig. 10. Similarity profile of cluster C4.

Fig. 8. Similarity profile of cluster C2. Fig. 11. Inconsistency profile of cluster C1.

It is also the weakest criterion in terms of preference for clusters C3

and C4.
In addition, it is possible to analyze if two criteria gh and gr are
similar, independent or conflicting in terms of preferences based on
the covariance of their related vectors ϕh ðÞ and ϕr ðÞ. Appendix D
gives the covariance matrix between all the pairs of criteria. Clearly,
most of the obtained values are positive and close to zero. This
implies that the correlation between the criteria is globally positive
and very low. Therefore, the criteria express similar preferences but
with a very low level. As a consequence, the criteria can be con-
sidered as globally independent rather than positively correlated.

6.3.2. Analysis of the similarity profiles

The similarity profile of each group of suppliers was also
determined in order to analyze its homogeneity on all the criteria.
Appendix E gives the similarity profiles of all the groups, i.e., their
corresponding unicriterion global similarity degrees on the dif-
Fig. 9. Similarity profile of cluster C3. ferent criteria. Figs. 7–10 illustrate graphically all these profiles.
The red and blue colors are used respectively for the capabilities
criteria. This can be visualized graphically based on Figs. 3–6. and willingness criteria. The values of the unicriterion global
Indeed, the “Willingness to share information” has the maximal similarity degrees, which vary between 0 and 1, are visualized in a
preferential power for cluster C1 and the fourth one for cluster C2. descending order. As can be seen, the unicriterion global similarity
 M.A. Boujelben / Omega 69 (2017) 126–140 135

Fig. 12. Inconsistency profile of cluster C2. Fig. 14. Inconsistency profile of cluster C4.

unicriterion global inconsistency degrees of all the clusters are

equal on average to 0.1023 for C1, 0.0944 for C2, 0.0929 for C3 and
0.0257 for C4. All the values do not exceed the value 0.2. Thus, the
inconsistency level related to each group of suppliers is low. This
illustrates clearly that the application of De Smet et al.'s method
in this problem gives good results in terms of inconsistency. In
addition, since the unicriterion global inconsistency degrees are
given in a descending order, it is possible to deduce for each
cluster the criteria having more or less impact on the
inconsistency.

7. Conclusion

In this paper, we addressed the problem of multicriteria clus-

tering where the groups are ordered from the best to the worst.
The main contributions of this work are three notions allowing to
interpret and better understand the obtained partition based on
Fig. 13. Inconsistency profile of cluster C3. the PROMETHEE principles. We introduced the notion of pre-
ference profile that permits to measure the preferential quality of
each cluster on the different criteria and to deduce specific inter-
pretations on the criteria (discriminating power, similar or inde-
degrees of all the clusters are equal on average to 0.8657 for C1, pendent or conflicting preferences). Then, we defined the notions
0.8253 for C2, 0.8029 for C3 and 0.8747 for C4. Therefore, the of similarity and inconsistency profiles allowing to analyze these
similarity of each group of suppliers is high. This shows that De two aspects for each cluster on each criterion. In addition, we
Smet et al.'s method provides for this problem a good partition in applied these measures in a supplier segmentation problem and
terms of similarity. Moreover, since the unicriterion global simi- visualized graphically the results. Clearly, these measures provide
larity degrees of each cluster are given in a descending order, it is interesting interpretations for the problem.
possible to determine for each group the criteria where the sup- Of course, there are still many directions for future research.
pliers are more or less similar. Among others, we can mention the problem of elicitation of the
preferential information used in PROMETHEE to build the valued
6.3.3. Analysis of the inconsistency profiles preference model. Another possible line of research is in analyzing
We also analyzed the inconsistency of each cluster on the the sensitivity of these measures to the preferential information of
different criteria with regard to total order of the obtained par- PROMETHEE. Furthermore, an interesting line of research is in the
tition. For that purpose, we determined the inconsistency profile development of multicriteria ordered approach that optimize
of each group of suppliers. Appendix F gives the profiles of all the simultaneously the preferential quality of the partition, the simi-
groups, i.e., their related unicriterion global inconsistency larity and the inconsistency. Finally, it is important to mention the
degrees on the different criteria. Figs. 11–14 illustrate them gra- problem of interacting criteria (i.e., synergy or redundancy
phically. The red and blue colors are used respectively for the between some criteria) in the field of multicriteria ordered clus-
capabilities and willingness criteria. The values of the uni- tering. In such cases, fuzzy measures (e.g. Choquet integral) [30]
criterion global inconsistency degrees, that vary between 0 and 1, can be used to represent these interactions. This is certainly a
are visualized in a descending order. As can be noticed, the motivating line of research that needs more investigation.
136 M.A. Boujelben / Omega 69 (2017) 126–140

Appendix A. Illustrative example

In order to illustrate the proposed notions in this paper, let us consider the following idealized example with 7 alternatives evaluated
according to 2 criteria (equally weighted) that have to be maximized. These alternatives are regrouped as shown in Fig. A1 into three
ordered groups defined successively from the right up corner (best cluster) to the bottom left corner (worst cluster). The preference model
was built using PROMETHEE. The preference functions were assumed to be linear with the following parameters: q1 ¼ q2 ¼ 0 and
p1 ¼ p2 ¼ 5. Tables A1 and A2 give the unicriterion preference degrees of both criteria g1 and g2. In what follows, we are particularly
interested in the preference, similarity and inconsistency profiles of cluster C2.
The preference profile of cluster C2 is defined by the unicriterion net flows of C2 according to g1 and g2, i.e.:
PRPðC 2 Þ ¼ ðϕ1 ðC 2 Þ; ϕ2 ðC 2 ÞÞ
To determine ϕ1 ðC 2 Þ, we should at first compute the following measures:

 The unicriterion positive flow of C2 on g1 ðϕ1þ ðC 2 ÞÞ :

P P
ai A C 1 aj A fC 2 ;C 3 g P 1 ðai ; aj Þ
ϕ1þ ðC 2 Þ ¼ ¼ 0:1866
3n4

Fig. A1. Data of the example – the horizontal and vertical axes represent the evaluations according to g1 and g2. The actions are represented by red asterisks and the clusters
are represented by ellipses.

Table A1
Unicriterion preference degrees of criterion g1 ðP 1 ð .,. ÞÞ.

a1 a2 a3 a4 a5 a6 a7

a1 – 0 0 0 0 0 0
a2 0.20 – 0 0 0 0 0
a3 0.32 0.12 – 0 0 0 0
a4 0.50 0.30 0.18 – 0 0 0
a5 0.60 0.40 0.28 0.10 – 0 0
a6 0.76 0.56 0.44 0.26 0.16 – 0
a7 1 0.80 0.68 0.50 0.40 0.24 –

Table A2
Unicriterion preference degrees of criterion g2 ðP 2 ð .,. ÞÞ.

a1 a2 a3 a4 a5 a6 a7

a1 0 0.20 0 0 0 0 0
a2 0 – 0 0 0 0 0
a3 0.30 0.50 – 0 0.06 0 0
a4 0.40 0.60 0.10 – 0.16 0 0
a5 0.24 0.44 0 – 0 0 0
a6 0.80 1 0.50 0.40 0.56 – 0.20
a7 0.60 0.80 0.30 0.20 0.36 0 –
 M.A. Boujelben / Omega 69 (2017) 126–140 137

 The unicriterion negative flow of C2 on g1 ðϕ1 ðC 2 ÞÞ :

P P
ai A C 1 aj A fC 2 ;C 3 g P 1 ðaj ; ai Þ
ϕ1 ðC 2 Þ ¼ ¼ 0:2033
3n4

ϕ1 ðC 2 Þ can then be easily deduced as follows:

ϕ1 ðC 2 Þ ¼ ϕ1þ ðC 2 Þ  ϕ1 ðC 2 Þ ¼  0:0167
Similarly, ϕ2 ðC 2 Þ can be computed as described previously (ϕ2 ðC 2 Þ ¼ 0:0133). The preference profile of cluster C2 is therefore:
PRPðC 2 Þ ¼ ð  0:0167; 0:0133Þ
The similarity profile of cluster C2 is defined by the unicriterion global similarity degrees of C2 according to g1 and g2, i.e.:
SPðC 2 Þ ¼ ðS1 ðC 2 Þ; S2 ðC 2 ÞÞ
To determine S1 ðC 2 Þ, we should at first compute the following measures:

 The unicriterion intra-similarity of C2 on g 1 ðSA1 ðC 2 ÞÞ :

To determine this value, we should compute the unicriterion intra-similarity of each pair of alternatives belonging to C2 on g1 using
Eq. (9). SA1 ðC 2 Þ is then deduced as the average of all these degrees (Eq. (10)). This leads to the following value:
SA1 ðC 2 Þ ¼ 0:8133

 The unicriterion outranking inter-similarity of C2 on g1 ðSE1þ ðC 2 ÞÞ :

To determine this value, we should compute the unicriterion outranking inter-similarity of each pair of alternatives belonging to C2 on g1
using Eq. (11). SE1þ ðC 2 Þ is then deduced as the average of all these degrees (Eq. (13)). This leads to the following value:
SE1þ ðC 2 Þ ¼ 0:9067

 The unicriterion outranked inter-similarity of C2 on g1 ðSE1 ðC 2 ÞÞ :

To determine this value, we should compute the unicriterion outranked inter-similarity of each pair of alternatives belonging to C2 on g1
using Eq. (12). SE1 ðC 2 Þ is then deduced as the average of all these degrees (Eq. (14)). This leads to the following value:
SE1 ðC 2 Þ ¼ 0:9067

S1 ðC 2 Þ can then be easily deduced as follows:

SA1 ðC 2 Þ þSE1þ ðC 2 Þ þ SE1 ðC 2 Þ
S1 ðC 2 Þ ¼ ¼ 0:8756
3

Similarly, S2 ðC 2 Þ can be computed as described previously (S2 ðC 2 Þ ¼ 0:9289). The similarity profile of cluster C2 is therefore:
SPðC 2 Þ ¼ ð0:8756; 0:9289Þ
Finally, the inconsistency profile of cluster C2 is defined by the unicriterion global inconsistency degrees of C2 according to g1 and g2,
i.e.:
IPðC 2 Þ ¼ ðI 1 ðC 2 Þ; I 2 ðC 2 ÞÞ
To determine I 1 ðC 2 Þ, we should at first compute the following measures:

 The unicriterion intra-inconsistency of C2 on g1 ðIA1 ðC 2 ÞÞ:

To determine this value, we should compute the unicriterion intra-inconsistency of each pair of alternatives belonging to C2 on g1 using
Eq. (17). IA1 ðC 2 Þ is then deduced as the average of all these degrees (Eq. (18)). This leads to the following value:
IA1 ðC 2 Þ ¼ 0:1867

 The unicriterion inter-inconsistency of C2 with regard to C3 on

g 1 ðIE1 ðC 2 =C 3 ÞÞ:
P P
ai A C 2 aj A C 3 P 1 ðaj ; ai Þ
IE1 ðC 2 =C 3 Þ ¼ ¼0
3n2

 The unicriterion inter-inconsistency of C2 with regard to C1 on

g 1 ðIE1 ðC 2 =C 1 ÞÞ:
P P
ai A C 2 aj A C 1 P 1 ðai ; aj Þ
IE1 ðC 2 =C 1 Þ ¼ ¼0
3n2
138 M.A. Boujelben / Omega 69 (2017) 126–140

I 1 ðC 2 Þ can then be easily deduced as the weighted sum of IA1 ðC 2 Þ, IE1 ðC 2 =C 3 Þ and IE1 ðC 2 =C 1 Þ using Eq. (21). This leads to the following
value:
I 1 ðC 2 Þ ¼ 0:0373
Similarly, I 2 ðC 2 Þ can be computed as described previously (I 2 ðC 2 Þ ¼ 0:0213). The inconsistency profile of cluster C2 is therefore:
IðC 2 Þ ¼ ð0:0373; 0:0213Þ

Appendix B. Suppliers' evaluations

Supplier no. Criteria

g1 g2 g3 g4 g5 g6 g7 g8 g9 g10 g11 g12

1 3 4 3 3 1 4 4 3 4 4 4 5
2 4 4 5 3 4 2 5 5 3 5 4 4
3 4 4 5 3 3 3 4 4 3 5 3 4
4 4 5 5 3 3 3 5 4 4 4 5 4
5 4 4 4 4 2 2 3 4 3 4 5 4
6 3 5 3 3 2 3 4 3 3 4 5 4
7 3 5 3 4 5 3 4 4 3 4 5 2
8 4 5 3 4 5 4 2 2 2 2 3 2
9 3 2 4 1 1 1 3 2 2 2 2 2
10 3 2 4 1 1 1 3 2 2 2 2 2
11 3 4 4 3 3 3 4 4 5 5 5 4
12 3 5 3 2 5 3 4 4 4 5 5 5
13 4 5 4 4 3 3 3 4 3 5 5 5
14 3 4 3 3 3 1 3 4 3 4 3 4
15 3 3 4 4 3 3 3 2 2 4 5 3
16 3 1 4 4 1 1 3 3 3 4 4 3
17 3 1 3 1 5 1 4 5 4 4 5 4
18 2 1 3 1 3 1 4 5 4 3 4 4
19 3 4 4 3 5 3 4 2 2 2 1 2
20 4 3 4 4 1 3 4 3 3 3 3 3
21 3 3 1 2 2 3 2 1 2 2 3 2
22 3 5 4 2 4 3 3 4 4 4 5 4
23 3 4 3 4 5 3 4 4 4 5 4 4
24 2 1 3 5 1 1 3 4 4 4 4 3
25 3 3 4 4 2 3 4 4 3 3 4 3
26 4 3 4 3 2 3 5 5 4 3 4 4
27 3 4 4 4 3 3 4 4 4 4 3 3
28 3 4 5 4 3 4 4 3 4 4 3 3
29 3 4 3 3 3 2 3 3 4 4 4 3
30 1 4 3 1 1 2 4 3 3 3 5 3
31 3 4 4 3 2 3 5 4 3 4 3 3
32 1 5 5 3 4 3 4 5 4 4 4 3
33 3 4 3 2 4 3 5 4 4 4 4 4
34 4 5 4 4 3 2 2 2 3 2 1 3
35 3 4 5 3 2 3 4 4 4 5 4 4
36 3 3 4 3 3 3 4 3 4 4 4 3
37 4 5 5 3 4 3 4 5 4 4 4 4
38 3 5 4 4 2 2 3 3 4 3 4 4
39 3 4 3 3 3 3 3 2 2 2 3 3
40 4 5 4 3 2 3 4 5 4 4 4 3
41 3 4 3 3 3 3 4 3 4 4 3 3
42 4 3 5 3 3 4 3 4 3 4 5 4
43 2 2 5 1 1 2 4 5 5 5 3 4
 M.A. Boujelben / Omega 69 (2017) 126–140 139

Appendix C. Preference profiles of all the clusters

g1 g2 g3 g4 g5 g6 g7 g8 g9 g 10 g 11 g 12
0 1
C1 0:1348 0:3027 0:1592 0:0826 0:1701 0:3145 0:2645 0:2609 0:2152 0:3870 0:2293 0:2609
C2 BB  0:1895  0:2908  0:1095  0:0629  0:0817  0:3486  0:0501 0:1397 0:2081 0:0240  0:0041 0:0577 C C
B C
C 3 @ 0:0036  0:0223  0:0509 0:1223  0:0196 0:0083  0:2250  0:3634  0:3667  0:3976  0:2045  0:2738 A
C4  0:0417  0:3667  0:2125  0:4417  0:3979  0:3361  0:3611  0:5083  0:5000  0:6167  0:3917  0:5083

Appendix D. Covariance matrix

g1 g2 g3 g4 g5 g6 g7 g8 g9 g 10 g 11 g 12
0 1
g1 – 0:0341 0:0153 0:0116 0:0158 0:0380 0:0169 0:0059  0:0004 0:0193 0:0129 0:0112
g2 B B 0:0341 – 0:0463 0:0581 0:0635 0:0952 0:0657 0:0628 0:0515 0:0961 0:0613
C
0:0703 C
B C
g3 B B 0:0153 0:0463 – 0:0306 0:0347 0:0466 0:0386 0:0425 0:0376 0:0593 0:0369 0:0441 C
C
g4 B B 0:0116 0:0581 0:0306 – 0:0560 0:0568 0:0430 0:0525 0:0492 0:0697 0:0465
C
0:0581 C
B C
g5 B B 0:0158 0:0635 0:0347 0:0560 – 0:0619 0:0555 0:0679 0:0631 0:0891 0:0562 0:0697 C
C
B C
g 6 B 0:0380 0:0952 0:0466 0:0568 0:0619 – 0:0622 0:0528 0:0400 0:0876 0:0566 0:0629 C
B C
g7 B B 0:0169 0:0657 0:0386 0:0430 0:0555 0:0622 – 0:0945 0:0894 0:1191 0:0713 0:0894 C
C
B C
g 8 B 0:0059 0:0628 0:0425 0:0525 0:0679 0:0528 0:0945 – 0:1401 0:1637 0:0963 0:1265 C
B C
g9 B B  0:0004 0:0515 0:0376 0:0492 0:0631 0:0400 0:0894 0:1401 – 0:1585 0:0927 0:1238 C
C
B C
g 10 B 0:0193 0:0961 0:0593 0:0697 0:0891 0:0876 0:1191 0:1637 0:1585 – 0:1185 0:1516 C
B C
g 11 B
@ 0:0129 0:0613 0:0369 0:0465 0:0562 0:0566 0:0713 0:0963 0:0927 0:1185 – 0:0906 C
A
g 12 0:0112 0:0703 0:0441 0:0581 0:0697 0:0629 0:0894 0:1265 0:1238 0:1516 0:0906 –

Appendix E. Similarity profiles of all the clusters

g1 g2 g3 g4 g5 g6 g7 g8 g9 g 10 g 11 g 12
0 1
C1 0:8507 0:8775 0:8538 0:8933 0:7931 0:9104 0:8542 0:8906 0:8753 0:8823 0:8669 0:8296
C2 BB 0:8889 0:6944 0:8796 0:7130 0:7685 0:7901 0:8765 0:8333 0:8642 0:8395 0:8796 0:8765 C
C
B C
C 3 @ 0:7698 0:7560 0:9107 0:8274 0:6786 0:7698 0:7857 0:9107 0:8730 0:7778 0:6964 0:9048 A
C4 1 0:8889 0:6667 0:8889 0:8889 0:7037 0:8519 0:8889 1 1 0:8889 1

Appendix F. Inconsistency profiles of all the clusters

g1 g2 g3 g4 g5 g6 g7 g8 g9 g 10 g 11 g 12
0 1
C1 0:1155 0:0894 0:1041 0:1262 0:1760 0:0645 0:0926 0:0856 0:1029 0:0739 0:0982 0:1141
C2 BB 0:0877 0:1425 0:0738 0:1674 0:1396 0:1179 0:0507 0:0833 0:0975 0:0526 0:0651 0:0624 C
C
B C
C 3 @ 0:1504 0:1445 0:0657 0:1688 0:1859 0:1461 0:0595 0:0122 0:0173 0:0390 0:1006 0:0195 A
C 4 0:0569 0:0386 0:0732 0:0122 0:0163 0:0569 0:0163 0:0041 0 0 0:0203 0

References
[1] Rokach L, Maimon O. Clustering methods. In: Maimon O, Rokach L, editors. [8] Zopounidis C, Doumpos M. Business failure prediction using the UTADIS
Data mining and knowledge discovery handbook. US: Springer; 2005. multicriteria analysis method. The Journal of the Operational Research Society
p. 321–52. 1999;50:1138–48.
[2] Ishizaka A, Nemery P. Multi-criteria decision analysis: methods and software. [9] Meyer P, Olteanu AL. Formalizing and solving the problem of clustering in
Chichester: Wiley; 2013. MCDA. European Journal of Operational Research 2013;227:494–502.
[3] Zopounidis C, Doumpos M. Multicriteria classification and sorting methods: a [10] De Smet Y, Nemery P, Selvaraj R. An exact algorithm for the multicriteria
ordered clustering problem. Omega 2012;40:861–9.
literature review. European Journal of Operational Research 2002;138:229–46.
[11] De Smet Y, Guzman L. Towards multicriteria clustering: an extension of the k-
[4] Yu W. ELECTRE TRI: aspects méthodologiques et manuel d'utilisation. Docu-
means algorithm. European Journal of Operational Research 2004;2:390–8.
ment du Lamsade no. 74. Université Paris-Dauphine; 1992.
[12] Eppe S, DeSmet Y. On the use of valued action profiles for relational multi-
[5] Ishizaka A, Nemery P. Assigning machines to incomparable maintenance
criteria clustering. International Journal of Multicriteria Decision Making
strategies with ELECTRE-SORT. Omega 2014;47:45–59.
2014;4:201–33.
[6] Belacel N. Multicriteria assignment method PROAFTN: methodology and
[13] Cailloux O, Lamboray C, Nemery P. A taxonomy of clustering procedures. In:
medical applications. European Journal of Operational Research Proceedings of the 66th meeting of the European working group on MCDA;
2000;125:175–83. 2007.
[7] Doumpos M, Zopounidis C. A multicriteria classification approach based on [14] Behzadian M, Kazemzadeh RB, Albadvi A, Aghdasi M. PROMETHEE: a com-
pairwise comparisons. European Journal of Operational Research prehensive literature review on methodologies and applications. European
2004;158:378–89. Journal of Operational Research 2010;200:198–215.
140 M.A. Boujelben / Omega 69 (2017) 126–140

[15] Rocha C, Dias LC, Dimas I. Multicriteria classification with unknown cate- [23] Halkidi M, Batistakis Y, Vazirgiannis M. On clustering validation techniques.
gories: a clustering–sorting approach and an application to conflict manage- Journal of Intelligent Information Systems 2001:107–45.
ment. Journal of Multi-Criteria Decision Analysis 2013;20:13–27. [24] Brans JP, Mareschal B. PROMETHEE-GAIA: une méthodologie d'aide à la
[16] Rocha C, Dias LC. MPOC: an agglomerative algorithm for multicriteria partially décision en présence de critères multiples, Ellipses, Paris; 2002.
ordered clustering. 4OR 2013;11:253–73. [25] Karsak EE, Dursun M. An integrated fuzzy MCDM approach for supplier eva-
[17] Valls A, Torra V. Using classification as an aggregation tool in MCDM. Fuzzy luation and selection. Computers and Industrial Engineering 2015;82:82–93.
Sets and Systems 2000;115:159–68. [26] Kraljic P. Purchasing must become supply management. Harvard Business
[18] De Smet Y, Gilbart F. A cluster definition method for country risk problems. Review 1983;(September/October):109–17.
Technical Report TR/SMG/2001-013, SMG. Université Libre de Bruxelles; 2001. [27] Day M, Magnan GM, Moeller MM. Evaluating the bases of supplier segmen-
[19] Fernandez E, Navarro J, Bernal S. Handling multicriteria preferences in cluster tation: a review and taxonomy. Industrial Marketing Management
analysis. European Journal of Operational Research 2010;202:819–27. 2010;39:625–39.
[20] Boujelben MA, De Smet Y. A multicriteria ordered clustering algorithm to [28] Rezaei J, Ortt R. Multi-criteria supplier segmentation using a fuzzy preference
determine precise or disjunctive partitions. International Journal of Multi- relations based AHP. European Journal of Operational Research 2013;225:75–
criteria Decision Making 2016;6:157–87. 84.
[21] Boujelben MA, De Smet Y, Frikha A, Chabchoub H. DISSET: a disjunctive [29] Wang TC, Chen YH. Applying fuzzy linguistic preference relations to the
sorting method based on evidence theory. Foundations of Computing and improvement of consistency of fuzzy AHP. Information Sciences
Decision Sciences 2007;32:253–74. 2008;178:3755–65.
[22] Figueira J, Mousseau V, Roy B. ELECTRE methods. Multiple criteria decision [30] Torra V, Narukawa Y. Modeling decisions: information fusion and aggregation
analysis: state of the art surveys. Boston, Dordrecht, London: Springer-Verlag; operators. Berlin, Heidelberg: Springer-Verlag; 2007.
2005. p. 133–62.