You are on page 1of 10

Expert Systems With Applications 114 (2018) 97–106

Contents lists available at ScienceDirect

Expert Systems With Applications


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

Aggregation in the analytic hierarchy process: Why weighted


geometric mean should be used instead of weighted arithmetic mean
Jana Krejčí a, Jan Stoklasa a,b,∗
a
Lappeenranta University of Technology, School of Business and Management, Skinnarilankatu 34, Lappeenranta, Finland
b
Palacký University Olomouc, Department of Applied Economics, Křížkovského 12, Olomouc, Czech Republic

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

Article history: The main focus of this paper is the aggregation of local priorities into global priorities in the Ana-
Received 10 April 2018 lytic Hierarchy Process (AHP) method. We study two most frequently used aggregation approaches - the
Revised 22 June 2018
weighted arithmetic and weighted geometric means - and identify their strengths and weaknesses. We
Accepted 23 June 2018
investigate the focus of the aggregation, the assumptions made on the way, and the effect of different
Available online 17 July 2018
normalizations of local priorities on the resulting global priorities and their ratios. We clearly show the
Keywords: superiority of the weighted geometric mean aggregation over the weighted arithmetic mean aggregation
Analytic hierarchy process in AHP for the purpose of deriving global priorities of alternatives. We also contribute to the literature
Aggregation on rank reversal in AHP. In particular, we show that a change of the normalization condition for the local
Weighted geometric mean priorities of alternatives may result in different ranking when the weighted arithmetic mean aggregation
Weighted arithmetic mean is used for deriving global priorities of alternatives, and we demonstrate that the ranking obtained by
Rank reversal
the weighted geometric mean aggregation is not normalization dependent. Moreover, we prove that the
Normalization of priorities
ratios of global priorities of alternatives obtained by the weighted geometric mean aggregation are in-
variant under the normalization of local priorities of alternatives and weights of criteria. We also propose
three alternative approaches to aggregating preference information contained in local pairwise compari-
son matrices of alternatives into a global consistent pairwise comparison matrix of alternatives and prove
their equivalence.
© 2018 Elsevier Ltd. All rights reserved.

1. Introduction frequent; see, e.g., the literature reviews by Ho (2008); Ho and


Ma (2018). Thus, any flaw in the AHP methodology may lead to
Analytic Hierarchy Process (AHP) is a well-known multi-criteria incorrect decisions with tremendous negative consequences.
decision-making method developed by Saaty (1977, 1980). The AHP is based on structuring the decision problem into a prob-
body of literature and research on AHP is extensive. AHP is contin- lem hierarchy and pairwisely comparing objects in one level of the
uously being used to support decisions in important decision prob- hierarchy with respect to the superior object from the upper level
lems in various fields ranging from engineering and industry ap- of the hierarchy. Based on the pairwise comparisons of objects, lo-
plications (Duman, Tozanli, Kongar, & Gupta, 2017), through social cal priorities of objects are derived and aggregated within the hi-
sciences applications (Jandová, Krejčí, Stoklasa, & Fedrizzi, 2017; erarchy in order to derive global priorities. For the simplicity of
Saaty, 2013) to applications in the medical sector (Nazari, Fal- explanation and without any loss of generality, we will assume a
lah, Kazemipoor, & Salehipour, 2018); for a more comprehensive multi-criteria decision-making problem with a simple 3-level hi-
overview see, e.g., the literature reviews by de F. S. M. Russo and erarchy, i.e. a problem with the goal specified in the first level of
Camanho (2015); Liberatore and Nydick (2008); Subramanian and the hierarchy, n criteria C1 , . . . , Cn relevant to the problem specified
Ramanathan (2012); Vaidya and Kumar (2006). Combinations of in the second level of the hierarchy, and m alternatives A1 , . . . , Am
AHP with other methods in practical decision-making are also specified in the third level of the hierarchy. The generalization of
our findings to hierarchies with more levels is straightforward.
A diagram describing the main stages of AHP is provided in

Corresponding author at: Lappeenranta University of Technology, School of Fig. 1. Note that the last stage, i.e. the construction of a global PCM
Business and Management, Skinnarilankatu 34, Lappeenranta, Finland of alternatives, is usually not done and the global priorities of al-
E-mail addresses: jana.krejci@unitn.it (J. Krejčí), jan.stoklasa@lut.fi, ternatives are used directly to rank the alternatives. In this paper,
jan.stoklasa@upol.cz (J. Stoklasa).

https://doi.org/10.1016/j.eswa.2018.06.060
0957-4174/© 2018 Elsevier Ltd. All rights reserved.
98 J. Krejčí, J. Stoklasa / Expert Systems With Applications 114 (2018) 97–106

Table 1
Local consistent PCMs of alternatives with respect to
criteria C1 and C2 .

PCM M1 of alternatives PCM M2 of alternatives


w.r.t. criterion C1 w.r.t. criterion C2

A1 A2 A3 A1 A2 A3

A1 1 2 8 A1 1 1/3 1/3
A2 1/2 1 4 A2 3 1 1
A3 1/8 1/4 1 A3 3 1 1

Both PCMs M1 and M2 are fully consistent accord-


ing to (1). The non-normalized local-priorities vectors
computed for the alternatives with respect to criteria
C1 and C2 by the GMM are pC1 = (2.520, 1.260, 0.315 )
and pC2 = (0.481, 1.442, 1.442 ), respectively.

 n
Definition 2. (Saaty, 1980) A PCM M = mi j is said to be con-
i, j=1
sistent if it satisfies the multiplicative-transitivity property

mi j = mik · mk j , i, j, k = 1, . . . , n. (1)

For particular examples of consistent PCMs, see, e.g., Table 1.


 n
Proposition 1. (Saaty, 1994) A PCM M = mi j is consistent if
i, j=1
and only if there exists a positive vector w = (w1 , . . . , wn )T such that

mi j = wi /w j , i, j = 1, . . . , n. (2)

Notice that from Proposition 1 the reciprocity property mij =


w
1/mji , i, j = 1, · · · , n, of consistent PCMs follows; mi j = w i = w1j =
j
wi
1
m ji . The reciprocity property is required from inconsistent PCMs as
well. Thus, in practice, the decision maker provides only one of the
PCs mij and mji , i, j ∈ {1, . . . , n}, and the other one is then inserted
automatically into the PCM by using the reciprocity property mi j =
1/m ji .
The priorities w1 , . . . , wn of objects are usually derived from a
PCM by the maximal eigenvector method (EVM) or by the geomet-
Fig. 1. Diagram of AHP stages.
ric mean method (GMM). In both cases, the priorities w1 , . . . , wn
are such that they satisfy (2) in case of a consistent PCM M or
mij ≈ wi /w j , i, j = 1, · · · , n, in case of an inconsistent PCM M. The
EVM was introduced by Saaty (1980) in the original version of AHP.
According to the EVM, the priorities of objects are derived as the
however, we will consider also this optional last step that is im- components of the maximal eigenvector w = (w1 , . . . , wn )T of the
portant for the correct interpretation of the final priorities of alter- PCM M, i.e. as the solutions to the equation
natives. Mw = λw, (3)
The main idea of AHP as introduced by Saaty (1980) is that in-
stead of providing the priorities wi and wj of objects oi and oj , re- where λ is the maximal eigenvalue of the PCM M. According to
spectively, a single number representing the ratio wi /wj is provided the GMM, the priorities of objects are derived as the geometric
by the decision maker. Thus, the decision maker provides a pair- means of the pairwise comparisons in the rows of the PCM M
wise comparison mij , i, j ∈ {1, · · · , n}, of objects oi and oj accord- (Barzilai, d. Cook, & Golany, 1987), i.e., as
ing to their relative priorities belonging to a ratio scale that is sup- 
posed to approximate the ratio wi /wj , i.e., mij ≈ wi /wj . In real-world 
n
wi = n
mi j , i = 1, . . . , n. (4)
applications, pairwise comparisons are usually done by choosing
j=1
linguistic terms from predefined Saaty’s scale (Saaty, 1980). The
pairwise comparisons of objects in one level of the hierarchy with Note 1. The priority vector w = (w1 , . . . , wn )T computed by (4) is
respect to the superior object from the upper level of the hierarchy in fact (up to the multiplication by a scalar) a solution to the prob-
are usually structured into a pairwise comparison matrix (PCM). lem of finding the minimum of the function
 n  n 
n   2
Definition 1. A PCM of n objects is a square matrix M = mi j
i, j=1
f ( w1 , . . . , wn ) = ln(mi j ) − ln wi /w j
whose element mij expresses the relative preference (or impor- i=1 j=1
tance) of object oi over object oj , and as such approximates the in the logarithmic least squares method introduced by
ratio wi /wj , i.e., mij ≈ wi /wj , where w1 , . . . , wn are the possibly un- Crawford and Williams (1985). The approach to the compu-
known priorities of the objects. tation of priorities of objects by the GMM can therefore be
understood as finding priorities of objects such that their ratios
For particular examples of PCMs, see, e.g., Table 1. are as close as possible to the respective elements of the PCM M.
J. Krejčí, J. Stoklasa / Expert Systems With Applications 114 (2018) 97–106 99

Many studies comparing the EVM and the GMM have been up a very good point that “the rank reversal problem in AHP is
conducted. Saaty and Vargas (1984a) showed that when the PCM purely a mathematical artefact, not something that results from a
M is consistent according to (1), then both methods lead to the set of behavioral claims built into the mathematics. Thus, the co-
same results. Further, Crawford and Williams (1985) showed that incidence that these mathematics yield reversal and people exhibit
when M is close to consistency according to (1), then the meth- reversal does not justify rank reversals, even for behavioral argu-
ods provide very similar results. Saaty and Vargas (1984a) com- ments” (Lai (1995), p. 458). Maleki and Zahir (2013) provided a
pared the EVM, the GMM, and the least squares method, and comprehensive literature review on rank reversal in AHP caused
they concluded that the EVM is the only method guaranteeing by adding/deleting an alternative or a criterion.
that mik ≥ mjk for all k = 1, . . . , n implies wi ≥ wj under inconsis- There is another possible cause for rank reversal, which is even
tency. Saaty and Hu (1998) showed an illustrative example in more significant, as it does not require a change in the set of
which the ranking of alternatives obtained by the GMM differs alternatives or criteria to be present. In this paper we show in
from the ranking obtained by the EVM. Based on this example, Section 2 that rank reversal can also be caused by simply chang-
Saaty and Hu (1998) concluded that the EVM is the only valid ing the normalization condition for the local priorities of alterna-
method for deriving priorities from PCMs, in particular then from tives. It is worth noting that the choice of normalization technique
inconsistent PCMs. However, as for example Krejčí (2018) pointed is usually not well justified in practical applications of AHP - most
out, showing that the GMM leads to a ranking different from the frequently (5) is used, less frequently (6). However, if the choice of
one obtained by the EVM does not surely demonstrate that the the normalization technique can influence the final ranking of ob-
EVM provides the correct solution (Krejčí, 2018). Crawford and jects, then its justification is not only highly advised - it becomes
Williams (1985) performed simulations to compare the perfor- a necessity! In Section 2 we show that the problem with rank re-
mance of the EVM and the GMM under different error distribu- versal in global priorities disappears as long as a specific aggrega-
tions and metrics. The simulations suggested better performance of tion method, namely the weighted geometric mean aggregation, is
the GMM in priorities estimation as well as in rank preservation. used.
Other studies favoring the GMM over the EVM have been done, For the simplicity of notation and better clarity of text, let from
e.g., by Barzilai (1997), Blaquero, Carrizosa, and Conde (2006), and now on wk , k = 1, · · · , n, denote the weights of the criteria de-
Dijkstra (2013). rived from the PCM of criteria by the GMM or the EVM and nor-
The priorities w1 , . . . , wn of objects derived from a PCM are usu- malized according to (5) if not specified otherwise. Further, let
C
ally normalized in order to reach uniqueness. The normalization p jk , j = 1, · · · , m, be the local priorities of the alternatives with re-
needs to be done in such a way that it does not change the ratios spect to criterion Ck , k ∈ {1, · · · , n}, derived from the correspond-
wi /w j , i, j = 1, · · · , n, of the priorities, as these play a central role ing local PCM Mk = {mki j }m of alternatives by the GMM or the
i, j=1
in the AHP method (Saaty, 1977; 1980; 1990). Thus, the normaliza-
EVM and arbitrarily normalized if not specified otherwise. The lo-
tion can be done only by multiplying the priorities by a suitable
c·w w cal priorities of alternatives are aggregated within the hierarchy
constant c > 0 (notice that c·w i = w i ).
j j into the global priorities g j , j = 1, · · · , m, of alternatives, based on
Two types of normalization of priorities have been frequently which the alternatives can be ranked.
used in AHP, Further, let us denote by G = {gi j }m i, j=1
the global PCM of al-

n ternatives such that the global pairwise comparison gij of alter-
wi = 1 (5) natives Ai and Aj is computed as gi j = gi /g j , for all i, j = 1, . . . , m.
i=1 Note that such global PCM G = {gi j }m i, j=1
is always consistent ac-
and cording to (1) since gik · gk j = (gi /gk ) · (gk /g j ) = gi /g j = gi j for all
i, j, k = 1, . . . , m.
max wi = 1. (6)
i=1,...,n There are also other possible ways of arriving at a general
global PCM H = {hi j }m of alternatives, e.g. by direct aggregation
Based on the choice of the normalization technique, two versions i, j=1
of AHP are distinguished - the distributive mode and the ideal of the local PCMs M1 , . . . , Mn or by the aggregation of the ratios
mode (see e.g. Saaty & Vargas (1993)). The distributive mode is of local priorities of alternatives. Note that since the elements of
characterized by the normalization of the priorities by using the such global PCM H are not computed as the ratios of the global
normalization condition (5), i.e. so that the sum of all normalized priorities of alternatives, the consistency of H is not guaranteed.
priorities is equal to one. The ideal mode involves the normaliza- Because we investigate the differences between the weighted
tion of the priorities by using the normalization condition (6), i.e. arithmetic mean (WAM) and weighted geometric mean (WGM)
so that the maximum normalized priority is equal to one. The nor- aggregation methods in the computation of the global priorities,
malization (5) was introduced by Saaty (1977, 1980) in the original the following notation will be used whenever we need to clearly
AHP method. Belton and Gear (1983) showed that the normaliza- distinguish between the approaches. The global priorities of al-
tion (5) can lead to rank reversal by adding or deleting an alter- ternatives computed by the WAM, the corresponding global con-
native, and as a solution to this problem they introduced the nor- sistent PCM, and the global PCM of alternatives derived by di-
malization (6). rect aggregation of the local PCMs M1 , . . . , Mn or by the aggrega-
Rank reversal in decision making is in general understood as a tion of the ratios of local priorities of alternatives by the WAM
change in the ranking of alternatives. Saaty and Vargas (1984b) ar- will be denoted by gWAM j
, j = 1, · · · , m, GWAM = {gWAM
ij
}m
i, j=1
, and
gued that rank reversal is desirable as it reflects how a new and H WAM = {hWAM
ij
}m
i, j=1
, respectively. Analogously, the global priorities
important attribute can alter previous preferences. Later, Saaty and of alternatives computed by the WGM, the corresponding global
Vargas (1993) accepted also the normalization (5) and concluded consistent PCM, and the global PCM of alternatives derived by di-
that the ideal mode should be used in decision problems in which rect aggregation of the local PCMs M1 , . . . , Mn or by the aggrega-
the uniqueness of alternatives (sufficient difference of each alterna- tion of the ratios of local priorities of alternatives by the WGM
tive from all the others) is important while the distributive mode will be denoted by gWGM j
, j = 1, · · · , m, GWGM = {gWGM
ij
}m
i, j=1
, and
should be used in decision problems in which the uniqueness is
H WGM = {hWGM
ij
}m
i, j=1
, respectively.
not important. After the study of Belton and Gear (1983) many
In the original AHP method, Saaty (1980) introduced WAM for
other researchers have investigated rank reversal related to adding
aggregating the local priorities of alternatives into the global prior-
or deleting an alternative or a criterion in AHP. Lai (1995) brought
100 J. Krejčí, J. Stoklasa / Expert Systems With Applications 114 (2018) 97–106

ities of alternatives. The global priority gWAM


j
of alternative A j , j ∈ the weights of criteria on the resulting global priorities under the
{1, . . . , m}, is thus computed as WAM and WGM aggregations, and provides evidence that not only
the ratios of global priorities but even the ranking of alternatives

n
C depends on the choice of the normalization technique under the
gWAM
j = wk · p j k , (7)
WAM aggregation. In particular, an example of rank reversal caused
k=1
by different normalization of local priorities under the WAM aggre-
C
where wk is the weight of the criterion Ck and p jk is the local gation is presented, and it is demonstrated that the ratios of global
priority of the alternative Aj with respect to the criterion Ck . Ob- priorities remain unchanged with different normalizations of local
viously, (7) assumes the normalization of criteria weights by (5). and global priorities of alternatives as well as the weights of crite-
Note that, in general, if we assume arbitrarily normalized weights, ria under the WGM aggregation. From this also the independence
n C of ranking of global priorities on the normalization of local priori-
k=1
wk ·p j k
we get gWAM
j
= n
wk
. ties under the WGM aggregation follows. Section 3 presents several
k=1
C C examples of the inability of the WAM aggregation to appropriately
Later, Lootsma (1993) argued that the ratios pi k /p jk of the lo-
reflect the preference information contained in the local PCMs Mk
cal priorities of alternatives should be actually aggregated since and provides theorems demonstrating that the WGM aggregation
according to the original idea of AHP the priorities (both lo- does not suffer from this problem. Finally, in Section 4 we present
cal and global) need to be interpreted in terms of ratios. three different approaches to the aggregation of information in the
Lootsma (1993) proposed to aggregate the ratios of the local pri- AHP method with the GMM for computing the local priorities of
C C
orities pi k and p jk of alternatives Ai and Aj , respectively, with the alternatives and with the WGM aggregation method, and we prove
weights wk of the criteria normalized by (5) by the WGM as their equivalence. We conclude the paper with a short discussion

w k of the relevance of our findings.

n C
pi k
C
. (8)
p jk 2. Effect of the choice of the normalization condition on the
k=1
global-priorities ratios and ranking of alternatives
Note that by (8) we in fact obtain the element hWGM
ij
of the
global PCM HWGM of the alternatives as defined above. In con- Two different normalizations - (5) and (6) - have been used in
C wk n C wk

AHP for local priorities of alternatives (see Section 1), and many re-

n pi k k=1
pi k
formity with the aggregation (8), since C =
 C wk
 , search papers investigating the normalizations and arguing for the
k=1 p jk n
p jk use of one of them have been published. In this section, we will
k=1

Lootsma (1993) proposed to compute the global priority gWGM


j
of focus on more fundamental aspects and consequences of priori-
alternative A j , j ∈ {1, . . . , m}, as ties normalization. Moreover, we will contribute to the research on
rank reversal in AHP. In the literature, the rank reversal was stud-
n 
 wk
C ied in relation to adding/deleting an alternative or a criterion (see
gWGM = p jk . (9)
j the literature review by Maleki & Zahir (2013)). We will show that
k=1
rank reversal can be caused also by the choice of different normal-
We do not restrict ourselves to the normalization of the weights ization techniques even when the sets of alternatives and criteria
of criteria by formula (5) in this paper. For an arbitrary normal- remain unchanged. This drawback occurs namely in connection to
ization of the weights of criteria, the formula for computing the the WAM aggregation. We will also prove that rank reversal caused
global priorities of alternatives from local ones using the WGM will by different normalization is not possible when the WGM is used
therefore be assumed in its general form to aggregate local priorities into global ones.
 First of all, we need to specify why the normalization is done.
n 
 γ wk
n
C Second, if there are more possible normalization techniques, we
gWGM
j = k=1
γ wk
p jk , (10)
need to be able either to select one of them or to conclude that
k=1
either way is possible. In the latter case the choice of a normal-
for any given γ > 0 (γ is called a normalization constant). The ization technique should not affect the results of a given decision
interpretation of the global priorities of alternatives by means problem. We can accept the premise that the ratios of the priori-
of ratios, which is well in line with the multiplicative nature of ties need to be preserved under normalization. This is reasonable
PCMs in AHP, was adopted also by Leskinen and Kangas (2005); since the fundamental information provided repeatedly by the DM
Lootsma (1996); Lootsma and Schuijt (1997); Triantaphyllou (2001). during the specification of the PCMs in AHP is in form of ratios of
In this paper, we will study in detail both the WAM and the the priorities/weights of the compared objects. We can also accept
WGM for aggregating local priorities of alternatives into global pri- the requirement of unique priorities/weights. However, in relation
orities of alternatives in AHP. On four illustrative examples we will to the priorities uniqueness we should again focus on the ratios of
show that the WAM does not reflect the preference information the priorities rather than on their values. Any other requirement
contained in local PCMs of alternatives properly. Moreover, we will on the normalization technique needs more information, more jus-
prove that the WGM aggregation reflects the preference informa- tification, and clear explanation. For example, if we consider the
tion contained in local PCMs of alternatives properly. Finally, we normalization of the weights of criteria, it is reasonable to require
will also propose three alternative approaches based on the GMM (5), which is in accordance with the usual assumptions for the set
and the WGM to aggregating preference information contained in of criteria - its completeness and the mutual independence of the
local PCMs of alternatives into a global consistent PCM of alterna- criteria. Under (5) the normalized weights express each respective
tives and prove their equivalence. criterion’s share of the overall goal (or, in case of a hierarchy with
This paper provides several strong arguments in favor of the more than 3 levels, on the closest higher-level goal). This interpre-
WGM aggregation over the WAM aggregation in AHP and clearly tation in our opinion justifies the choice of (5) over (6) and other
proves that the use of the WGM in AHP avoids many problems (see normalizations preserving weights ratios. Contrarily, if we consider
Section 2 and 3) that are present when the WAM is used to cal- the normalization of the local priorities of alternatives, which are
culate the global priorities. More specifically, Section 2 investigates usually not assumed to constitute a whole, the requirement em-
the effect of normalization of the local priorities of alternatives and bodied by (5) becomes obsolete. Why should we assume that the
J. Krejčí, J. Stoklasa / Expert Systems With Applications 114 (2018) 97–106 101

Table 2
Aggregation of the local priorities computed from M1 and M2 and normalized by (5).

Aj pCj1 pCj2 gWAM


j
global-priorities gWGM
j
global-priorities
ratios ratios

A1 8/13 1/7 69/182 gWAM


1 /gWAM
2 = 69/67 (8/91)0.5 gWGM
1
/gWGM
2
= (2/3 )0.5
A2 4/13 3/7 67/182 gWAM
1 /gWAM
3 = 69/46 (12/91)0.5 gWGM
1
/gWGM
3
= (8/3 )0.5
A3 1/13 3/7 46/182 g2 /g3
WAM WAM
= 67/46 (3/91)0.5 gWGM
2
/gWGM
3
= 40.5 = 2

Local priorities of the alternatives with respect to each criterion are normalized by (5), i.e.,
3 C1 3 C2
j=1 p j = 1 and j=1 p j = 1. Both g j are computed with w1 = 0.5 and w2 = 0.5.
WAM
and gWGM
j

Table 3
Aggregation of the local priorities computed from M1 and M2 and normalized by (6).

Aj pCj1 pCj2 gWAM


j
global-priorities gWGM
j
global-priorities
ratios ratios

A1 1 1/3 2/3 = 32/48 gWAM


1 /gWAM
2 = 8/9 (1/3)0.5 gWGM
1
/gWGM
2
= (2/3 )0.5
A2 1/2 1 3/4 = 36/48 gWAM
1 /gWAM
3 = 32/27 (1/2)0.5 gWGM
1
/gWGM
3
= (8/3 )0.5
A3 1/8 1 9/16 = 27/48 gWAM
2 /gWAM
3 = 4/3 (1/8)0.5 gWGM
2
/gWGM
3
= 40.5 = 2

Local priorities of the alternatives with respect to each criterion are normalized by (6), i.e.,
max pCj1 = 1 and max pCj2 = 1. Both gWAM
j
and gWGM
j
are computed with w1 = 0.5 and w2 = 0.5.
j=1,2,3 j=1,2,3

sum of the local priorities of alternatives is equal to 1? The per- differs for each normalization condition. Specifically, the alterna-
centual interpretation is meaningless in this case. Both (5) and tives are ranked as A1 A2 A3 under the normalization (5) (see
(6) and any other normalization preserving the ratios of the local Table 2), while the ranking A2 A1 A3 is obtained under the nor-
priorities of alternatives is as good and reasonable as the others malization (6) (see Table 3). Contrarily, the ranking of alternatives
(as long as the same normalization is used consistently for the lo- induced by their global priorities computed using the WGM is the
cal priorities of alternatives on all criteria). same for both normalizations, namely A2 A1 A3 . Notice that the
If there is no good justification for the use of a specific normal- preservation of the ranking under different normalizations for the
ization, then the choice of the normalizing constant is arbitrary. In WGM follows immediately from the observation that the ratios of
such case, the normalization of the local priorities of alternatives the global priorities remain unchanged under different normaliza-
and the weights of criteria should not have an impact on the so- tions. Thus, unless a specific normalization of the local priorities is
lution to the decision problem. It therefore seems to be natural to well justified, the WGM should be used to avoid the dependence
require that also the aggregation of the local priorities will provide of the global-priorities ratios and the ranking of the alternatives on
results (the ratios of the global priorities of alternatives) indepen- an arbitrarily chosen normalization.
dent of the choice of the normalization constant.
The following example shows that the ratios of the global prior- In Example 1 the ratios of the global priorities of alternatives
ities change with different normalizations of local priorities when obtained by the WGM aggregation remained the same with dif-
the WAM aggregation (7) is used, but remain the same when the ferent normalization conditions applied to the local priorities of
WGM aggregation (9) is applied. This demonstrates that the WAM alternatives. The following theorem shows that the ratios of the
might not be the best method for aggregating local priorities into global priorities of alternatives obtained by the WGM aggregation
global ones, particularly when the ratios of the global priorities are are independent of the normalization conditions for the local and
to be interpreted. Also note that the problem persists irrespective global priorities of alternatives as well as for the weights of crite-
of the computation technique for the determination of the local ria. This is valid not only for local consistent PCMs of alternatives
priorities. with respect to criteria but also for inconsistent ones, regardless
of the method used for deriving local priorities of alternatives and
Example 1. Let us consider the PCMs of alternatives A1 , A2 , and weights of criteria.
A3 with respect to criteria C1 and C2 given in Table 1 and equal
importance of both criteria represented by the weights w1 = w2 = Theorem 1. The ratios gWGM /gWGM , i = 1, . . . , m, of the global prior-
i j
0.5.
ities of alternatives obtained by the WGM aggregation method (10) re-
The local priorities of the alternatives with respect to the cri-
main unchanged with any normalization of the local and global pri-
teria calculated by the GMM, the respective global priorities cal-
orities of alternatives and of the weights of criteria, irrespective of the
culated using the WAM (7) and the WGM (9), and the ratios of
method for deriving local priorities of alternatives and weights of cri-
the global priorities are summarized in Table 2 and Table 3. More
teria.
specifically, Table 2 operates under the normalization condition
(5) for the local priorities of alternatives, whereas Table 3 operates C
Proof. Let p jk be the local priority of alternative A j , j ∈ {1, · · · , m},
under the normalization condition (6).
with respect to criterion Ck , k ∈ {1, · · · , n}, derived from the local
Comparing Table 2 and Table 3, we can clearly see that the cor-
PCM Mk of the alternatives with respect to criterion Ck . Further,
responding ratios of the global priorities computed using the WAM
let wk , k = 1, · · · , n, be the weights of the criteria (not necessar-
are different for each normalization of the local priorities of alter-
ily normalized in general). Let α k > 0 be an arbitrary normaliza-
natives. In the case of the WGM, however, the ratios of the global C
priorities remain constant under different normalizations of the lo- tion constant for normalizing the local priorities p jk of alterna-
cal priorities of alternatives. The invariance of the ratios of global tives A j , j = 1, · · · , m, with respect to criterion Ck , k ∈ {1, · · · , n},
priorities obtained by the WGM (9) under normalization is formu- let β > 0 be an arbitrary normalization constant for normalizing
lated and proven later in Theorem 1. the global priorities gWGM j
of alternatives A j , j = 1, · · · , m, and let
Moreover, we can see from the tables that the ranking of alter- γ > 0 be an arbitrary normalization constant for normalizing the
natives induced by their global priorities computed using the WAM weights of criteria.
102 J. Krejčí, J. Stoklasa / Expert Systems With Applications 114 (2018) 97–106

The ratio of the normalized global priorities of two alterna- Table 4


tives Ai and A j , i, j ∈ {1, · · · , m}, obtained by the WGM aggregation Local consistent PCMs of alternatives with respect to criteria C1
and C2 .
C
(10) of the normalized local priorities αk · p jk with the normalized 1 2
PCM M of alternatives PCM M of alternatives
weights γ wk of the criteria can be simplified as follows:
  with respect to criterion C1 with respect to criterion C2
n n  
C γ wk
β k=1
γ wk
k=1 αk · pi k A1 A2 A3 A1 A2 A3
gWGM
i
=   A1 1 2 4 A1 1 3 9
gWGM n n  
C γ wk
A2 1/2 1 2 A2 1/3 1 3
j
β k=1
γ wk
k=1 αk · p jk A3 1/4 1/2 1 A3 1/9 1/3 1


 n  
 k=1 αk · pCi k γ wk between the WAM and the WGM in the aggregation of local prior-
= k=1 γ w
n
k
n  
Ck γ wk ities into global priorities in AHP.
k=1 αk · p j
Example 2. Let us consider the local consistent PCMs of alterna-

γ w k
n tives A1 , A2 , and A3 with respect to criteria C1 and C2 given in
 αk pCk
w
n
γ i Table 4 and equal importance of both criteria represented by the
= k =1 k

k=1
αk pCjk weights w1 = w2 = 0.5.
1
From the local PCM M we see that the intensity of preference


n w of A1 over A2 on criterion C1 is the same as the intensity of pref-
 pCk k erence of A2 over A3 , namely m112 = m123 = 2. Similarly, from the
w
n
i
= k=1 k Ck 2
k=1
pj local PCM M we see that the intensity of preference of A1 over A2
on criterion C2 is the same as the intensity of preference of A2 over
  wk A3 , in this case m212 = m223 = 3. Thus, we would expect that also the
n
w
n C
k=1 k
k=1 pi k global intensity of preference of A1 over A2 will be the same as the
=   wk , (11)
global intensity of preference of A2 over A3 , i.e. g12 = g23 . However,
n
w
n C
k=1 k
k=1 p jk we can easily show that the WAM aggregation does not preserve
this equality, under any of the normalizations (5) and (6) of local
which shows that the ratio of the global priorities does not depend
priorities.
on the choice of the normalization constants αk , k = 1, · · · , n, β
In Table 5, the local priorities of alternatives normalized by
and γ . 
(5) are provided, together with the global priorities of alternatives
Many research papers have demonstrated that the WAM aggre- obtained by the WAM (7) and the WGM (9) and their ratios. Analo-
gation (7) may lead to rank reversal when one alternative is re- gously, in Table 6, the local priorities of alternatives normalized by
moved from the set of decision alternatives or when a new alter- (6) are provided, together with the global priorities of alternatives
native (either a copy or a close copy of an existing alternative) obtained by the WAM (7) and the WGM (9) and their ratios.
is added; see the literature review provided by Maleki and Za- We see that gWAM 1 /gWAM
2 = gWAM
2 /gWAM
3 (or in other words
hir (2013). In addition to this, Example 1 further shows that the gWAM
12
 = gWAM ) in Table 5 as well as in Table 6. Contrarily, the
23
WAM aggregation (7) leads to rank reversal even when the set of equality gWGM 1
/gWGM
2
= gWGM
2
/gWGM
3
(or in other words gWGM 12
=
decision alternatives remains unchanged but when the normaliza- WGM
g23 ) is valid for both normalization conditions. Our focus on the
tion condition for the local priorities of alternatives changes. This is priorities ratios gi j = gi /g j is well justified, as apparent from the
an even more serious drawback. Contrarily, with the WGM aggre- literature review in Section 1 and our comments and examples
gation (10) the rank reversal can never occur under different nor- presented in Sections 1 and 2. A reasonable requirement on the
malizations of the local priorities of alternatives and the weights aggregation method would therefore be to reflect the ratios of the
of criteria. This is because the ratios of the global priorities of al- local priorities as well as possible. In our example, where two local
ternatives remain unchanged as proven in Theorem 1. consistent PCMs are considered and where m112 = m123 and m212 =
m223 , it seems only reasonable to expect that g12 = g23 would hold
3. The ability of weighted arithmetic mean and weighted  m
for the global consistent PCM G = gi j , gi j = gi /g j , as well. We
geometric mean to reflect the preference information in local i, j=1
pairwise comparison matrices can clearly see that under the WAM gWAM 1 /gWAM
2 = gWAM
2 /gWAM
3 re-
gardless of the normalization technique used for the local prior-
In Section 2, Example 1, we have already demonstrated that the ities, i.e. gWAM
12
= wWAM
13
. Contrarily, the WGM aggregation reflects
ratios of global priorities depend on the normalization technique the ratios of local priorities well in GWGM , i.e., we get the expected
applied to the local priorities, as long as the WAM is used for ag- result gWGM
1
/gWGM
2
= gWGM
2
/gWGM
3
, or in other words gWGM
12
= gWGM
23
.
gregation. We have also proved the invariance of global-priorities The inability to properly reflect the ratios of local priorities for
ratios under different normalization of the local and global prior- local consistent PCMs seems to be a significant drawback of the
ities of alternatives and the weights of criteria for the WGM ag- WAM aggregation.
gregation. In this section, we will investigate the performance of
The desired ability of the WGM aggregation to preserve the
the WAM and the WGM in the aggregation of local priorities of
equality of pairwise comparisons shown in Example 2 is formu-
alternatives into their global priorities in terms of how well they
lated in the following theorem.
reflect the preference information contained in the local PCMs of
alternatives. Theorem 2. Let M1 , . . . , Mn be local PCMs of alternatives with re-
Shortcomings of the WAM aggregation will be shown on illus- spect to criteria C1 , . . . , Cn , respectively, consistent according to (1),
trative examples. The examples are simplified for the convenience and let w1 , . . . , wn be the weights of criteria. Further, let the intensity
of the reader so that only three alternatives and two criteria are of preference of alternative Ap over alternative Aq be the same as the
considered, the local PCMs of alternatives are consistent according intensity of preference of alternative Ar over alternative As on each
to (1), and the criteria are considered to be equally important. Yet criterion, i.e. mkpq = mkrs for all k = 1, . . . , n. Let the global priorities
these examples will help us highlight some important differences of alternatives be computed by the WGM of local priorities. Then the
J. Krejčí, J. Stoklasa / Expert Systems With Applications 114 (2018) 97–106 103

Table 5
1 2
Aggregation of the local priorities computed from M and M and normalized by (5).

Aj pCj1 pCj2 gWAM


j
global-priorities gWGM
j
global-priorities
ratios ratios

A1 4/7 9/13 115/182 gWAM


1 /gWAM
2 = 115/47 (36/91)0.5 gWGM
1
/gWGM
2
= 60.5
A2 2/7 3/13 47/182 gWAM
1 /gWAM
3 = 115/20 (6/91)0.5 gWGM
1
/gWGM
3
= 360.5
A3 1/7 1/13 20/182 g2 /g3
WAM WAM
= 47/20 (1/91)0.5 gWGM
2
/gWGM
3
= 60.5

Local priorities of the alternatives with respect to each criterion are normalized so that
3 C1 3 C2
j=1 p j = 1 and j=1 p j = 1. Both g j are computed with w1 = 0.5 and w2 =
WAM
and gWGM
j
0.5.

Table 6
1 2
Aggregation of the local priorities computed from M and M and normalized by (6).

Aj pCj1 pCj2 gWAM


j
global-priorities gWGM
j
global-priorities
ratios ratios

A1 1 1 1 gWAM
1 /gWAM
2 = 12/5 10.5 gWGM
1
/gWGM
2
= 60.5
A2 1/2 1/3 5/12 gWAM
1 /gWAM
3 = 72/13 (1/6)0.5 gWGM
1
/gWGM
3
= 360.5
A3 1/4 1/9 13/72 gWAM
2 /gWAM
3 = 30/13 (1/36)0.5 gWGM
2
/gWGM
3
= 60.5

Local priorities of the alternatives with respect to each criterion are normalized so that
max pCj1 = 1 and max pCj2 = 1. Both gWAMj
and gWGM
j
are computed with w1 = 0.5 and
j=1,2,3 j=1,2,3
w2 = 0.5.

global intensity of preference gWGM


pq = gWGM
p /gWGM
q of Ap over Aq is Table 7
Local consistent PCMs of alternatives with respect to criteria C1 and C2
the same as the global intensity of preference gWGM
rs = gWGM
r /gWGM
s of representing directly opposite pairwise comparisons and the global prior-
Ar over As , i.e. gWGM
pq = gWGM
rs in the global consistent PCM GWGM . ities computed by the WAM and the WGM for equal importance of both
criteria.
Proof. Let α k > 0, k = 1, . . . , n, and γ > 0 be some normalization
PCM M 1 of alternatives PCM M 2 of alternatives global priorities
constants for the local priorities and for the weights, respectively.
w.r.t. criterion C1 w.r.t. criterion C2
We know that for local consistent PCMs Mk , k = 1, · · · , n, the
αk pi k
C A1 A2 A3 A1 A2 A3 gWAM
j
gWGM
j
equation C = mki j holds for every i, j = 1, · · · , m, k = 1, · · · , n,
αk p jk A1 1 2 4 A1 1 1/2 1/4 0.625 0.500
C A2 1/2 1 2 A2 2 1 1/2 0.500 0.500
irrespective of the method for deriving the local priorities pi k . A3 1/4 1/2 1 A3 4 2 1 0.625 0.500
Therefore, 1 and M
2 are consistent according to (1) and m
1i j =
 Both local PCMs M

n  
Ck γ wk

2i j , i, j = 1, 2, 3. The vectors of local priorities of the alternatives with
1/m
n

γ wk
αk pi n wk

k=1 C respect to criteria C1 and C2 normalized by (6) are pC1 = (1, 0.5, 0.25 ) and
k=1
n
w
pi k pC2 = (0.25, 0.5, 1 ), respectively.
gWGM =  = k=1 k
ij
n

n  γ wk Ck
pj
k=1
γ wk
αk pCjk k=1
k=1
 receives reverse evaluations on the two criteria, it would be natu-
n

n  wk
= wk
mki j , rally expected that each pair of alternatives is globally evaluated as
k=1

k=1 equally preferred. This means that the global weights of all three
alternatives should be equal and their ratios should hence be equal
for all i, j = 1, . . . , m. Moreover, the equations mkpq = mkrs , k = to 1. However, the aggregation of local priorities (normalized using
1, . . . , n, hold. Thus, (6)) by the WAM results in A2 being considered inferior to A1 and
  A3 (see Table 7, column of gWAM ), without any objective reason.
n 
 wk n 
 wk j
n
n

gWGM = wk
mkpq = wk
mkrs = gWGM , Meanwhile the WGM aggregation interprets the information pre-
pq k=1 k=1 rs
sented in the local PCMs M 1 and M2 in such a way that all the
k=1 k=1
alternatives are considered equivalent in terms of their global pri-
which concludes the proof.  orities (see Table 7). This is much more in line with the actual in-
formation presented in the local PCMs M 1 and M 2 and the fact that
In the following example we will present a case where one
the criteria are equally important. Note that analogous results can
would expect indifference between all the alternatives on the
be obtained for the normalization of local priorities by (5).
global level, but where the WAM aggregation method suggests a
different global ranking.
The ability of the WGM aggregation to properly aggregate di-
Example 3. Let us now investigate the behavior of the WAM and rectly opposite pairwise comparisons of alternatives under equally
the WGM under rather specific conditions. Let us consider the two important criteria shown in Example 3 is formulated in an even
local PCMs of alternatives A1 , A2 , and A3 with respect to criteria more general form in the following theorem.
C1 and C2 given in Table 7 and equal importance of both criteria
represented by the weights w1 = w2 = 0.5. Theorem 3. Let M1 and M2 be local PCMs of alternatives A1 , . . . , Am
Both local PCMs M 1 and M 2 are consistent according to (1). with respect to criteria C1 and C2 , respectively, consistent according
Moreover, notice that the PCMs M 1 and M 2 are direct opposites, to (1), and let the pairwise comparisons of two alternatives Ar and
 1 = 1/m
i.e. m 2 , for all i, j = 1, 2, 3. For example, A1 is 2-times pre- As , r, s ∈ {1, · · · , m}, on the criteria be such that m1rs = 1/m2rs . Fur-
ij ij
ferred over A2 on C1 and A2 is 2-times preferred over A1 on C2 ; ther, let the criteria C1 and C2 be equally important, i.e. w1 = w2 .
analogously for the remaining two pairs of alternatives. Since cri- Then the global priorities of alternatives Ar and As computed by the
teria C1 and C2 are equally important and each pair of alternatives WGM are the same, i.e. gWGM r = gWGM
s .
104 J. Krejčí, J. Stoklasa / Expert Systems With Applications 114 (2018) 97–106

Proof. Let α 1 , α 2 > 0 and γ > 0 be some normalization constants By aggregating the ratios of local priorities of alternatives A1
for the local priorities of alternatives and the weights of criteria, and A2 in Table 3 by means of the WGM (8) we obtain
respectively. We know that for local consistent PCMs M1 and M2 0.5 0.5 1 0.5  2 0.5
αk pi k
C 
2
p1k
C
1
the equation = mki j , k = 1, 2, holds irrespective of the method hWGM
12 = C
= 1
· 3
=
C
αk p jk p2k 1 3
k=1 2
C
for deriving the local priorities pi k . Thus, we obtain and
  γ wk 0.5 1 0.5 0.5  3 0.5
2 2 
2
p2k
C
1
gWGM
k=1
γ wk
k=1 αk pCr k hWGM
21 = = 2
· = .
r
=    k=1
C
p1k 1 1
3
2
gWGM 2 2 Ck γ wk
s k=1
γ wk
k=1 αk ps 1
Obviously, hWGM = , and analogously for the other two pairs
12 hWGM
 γ w k
21


2 Ck of alternatives. Thus, the global PCM HWGM of alternatives keeps
2
γ wk pr
= k=1
C
the reciprocity.
k=1
ps k
Note 2. In Example 4 we showed on a particular problem with two
 criteria and three alternatives that the WAM aggregation of ratios
2 
 γ wk
2
= k=1
γ wk
mkrs does not preserve the reciprocity. This is true also for general deci-
k=1
sion problems with n criteria, m alternatives, and arbitrary weights
 of the criteria. In particular
2
γ wk
γ w1  γ w2 ⎡ ⎤
= k=1 m1rs · m2rs n pi k
C

⎢ k=1 wk C

 γ w1 ⎢
p jk
⎥ = ⎡ 1
n
C ⎤. (12)
( w1 =w2 )
=
2γ w1
m1rs · mrs 2 ⎣ k=1 wk
⎦ n p k
j
wk
 
⎢ k=1 pCi k ⎥
⎣ nk=1 wk ⎦
m1rs = 1
2 √
2γ w1
1γ w1 = 1,
mrs
=

i.e., gWGM
r = gWGM
s , which concludes the proof.  The following theorem demonstrates that the reciprocity is pre-
served when the ratios of local priorities are aggregated by the
Notice that Theorem 3 is much more general than the par- WGM (8).
ticular case examined in Example 3. Namely, in Example 3 the
special property m1i j = 1/m2i j was valid for each pair of alterna- Theorem 4. The WGM aggregation (8) of ratios of local priorities pre-
serves the reciprocity.
tives Ai and A j , i, j = 1, 2, 3. By contrast, the special property = m1rs
1/m2rs is required only for one particular pair of alternatives in Proof.
Theorem 3 (i.e., the remaining pairwise comparisons in the local

⎛ ⎞ nwk
consistent PCMs M1 and M2 can be arbitrary), resulting in equal n wk
wk

n 
Ck n ⎜
 ⎟ k=1
global priorities of these two alternatives. w
p 1
⎜ ⎟

i
k
= =
There seems to be at least one more good reason why to prefer
k=1
C
pk ⎝ pCk ⎠
k=1 j k=1 j
the WGM over the WAM. The WAM aggregation does not preserve pi k
C

the reciprocity condition when applied to the aggregation of the


ratios of local priorities. However, the reciprocity is a very basic 
n
1 1
requirement on the PCMs in AHP.  nwk =  wk (13)
k=1 C
p jk wk n

wk

n C
p jk
k=1
C k=1 C
Example 4. Let us consider the local PCMs of alternatives with re- pi k k=1 pi k
spect to criteria C1 and C2 given in Table 1 and equal weights of
the criteria w1 = w2 = 0.5. The local priorities of alternatives nor- 
malized by (6) are given in Table 3. In the following section, we will study the properties of the
By aggregating the ratios of local priorities of alternatives in WGM aggregation method in more detail, in particular in combi-
Table 3 by means of the WAM we obtain the elements hWAM ij
, i, j = nation with the GMM for deriving local priorities of alternatives.
1, 2, 3, of the global PCM H WAM = {hWAM
ij
}3i, j=1 (recall the difference We will show that using the GMM to calculate local priorities of
between the global consistent PCM G and the general global PCM alternatives and the WGM to aggregate these local priorities into
H as discussed in Section 1). In particular, for alternatives A1 and global priorities of alternatives, several alternative procedures can
A2 we obtain be used to obtain the ratios of global priorities.


2 C
p1k 1 1
7 4. Alternative approaches to the calculation of ratios of global
3
hWAM
12 = 0.5 · = 0.5 · + 0.5 · = ,
C
p2k
1 1 6 priorities using the weighted geometric mean aggregation
k=1 2

and In this section, we suggest three alternative procedures for


 obtaining the ratios of the global priorities of alternatives. Note

2 C
p2k 1
1 7 that we do not require consistency (1) neither for the local PCMs
hWAM
21 = 0.5 · = 0.5 · 2
+ 0.5 · = .
C
p1k 1 1 4 Mk , k = 1, . . . , n, of the alternatives with respect to criteria nor for
k=1 3
the PCM of the criteria. Without any loss of generality we will as-
Obviously, hWAM = 1
, which means that the reciprocity of the sume that the weights w1 , . . . , wn of criteria C1 , . . . , Cn are normal-
12 hWAM
21 ized by (5) to maintain the notation as simple as possible. Recall
global PCM HWAM is violated. Analogously for the other two pairs
that we have already proven that the normalization constants can
of alternatives.
J. Krejčí, J. Stoklasa / Expert Systems With Applications 114 (2018) 97–106 105

be chosen arbitrarily as long as the WGM is used for the aggrega-


 wk

n
m
tion.
m
mkis
k=1 s=1
We define three procedures for aggregating preference informa- =  w k , (15)
tion in AHP with the GMM for deriving local priorities of alterna-
n
m

tives and with the WGM aggregation as follows:


m
mkjs
k=1 s=1
Procedure 1
which equals to (14). Thus, H WGM = GWGM .
1. For each criterion Ck , k = 1, · · · , n, compute the local priori- Now, let us show the equivalence of Procedure 1 and Procedure
C
ties p jk , j = 1, · · · , m, of alternatives from the local PCM Mk = 3. According to Procedure 3 and by further adjusting, we obtain for
{mkrs }m by the GMM (4). each two alternatives Ai , A j , i, j = 1, · · · , m:
r,s=1
2. For each alternative A j , j = 1, · · · , m, aggregate the local priori-  
C
m
n  wk  wk
ties p jk , k = 1, · · · , n, into the global priority gWGM of the alter- m
mkis
m
n
j m
mkis
native by the WGM (9). s=1 k=1 s=1 k=1
3. For each pair of alternatives Ai , A j , i, j = 1, · · · , m, compute
WGM
gi j =  = m n
  (16)
the ratio gWGM /gWGM and construct the global consistent PCM m

m
n  k wk m
mkjs
wk
i j m js s=1 k=1
GWGM = {gWGM
ij
}m , gWGM = gWGM
i, j=1 i j i
/gWGM
j
. s=1 k=1

Procedure 2 # w k  wk

n
m
n
m
m
mkis m
mkis
1. For each criterion Ck , k = 1, · · · , n, compute the local priori- k=1 s=1 k=1 s=1
C =
n m  w k =  w k , (17)
ties p jk , j = 1, · · · , m, of alternatives from the local PCM Mk =
n
m
m
mkjs m
mkjs
{mkrs }m
r,s=1
by the GMM (4). k=1 s=1
k=1 s=1
2. For each two alternatives Ai , A j , i, j = 1, · · · , m, compute the ra-
C C WGM
tio pi k /p jk of their local priorities with respect to each criterion which equals to (14). Thus, G = GWGM . 
Ck , k = 1, · · · , n.
Notice that the GMM (4) is used for deriving the local priori-
3. For each two alternatives Ai , A j , i, j = 1, · · · , m, aggregate the
C C
ties of alternatives from local PCMs in Procedures 1 and 2, while
ratios pi k /p jk , k = 1, · · · , n, by the WGM (8) to obtain hWGM
ij in Procedure 3 it is used for deriving the global priorities of alter-
and construct the global PCM H WGM = {hWGM
ij
}m
i, j=1
. natives from the global PCM. Procedures 1 and 2 lead to the same
results also when the GMM for the computation of local priori-
Procedure 3 ties of alternatives is replaced by the EVM (3). This follows directly
from the Eq. (8) that is independent of the method for computing
1. For each two alternatives Ai , A j , i, j = 1, · · · , m, aggregate the local priorities of alternatives.
pairwise comparisons mkij , k = 1, · · · , n, by the WGM (9) with
the weights w1 , . . . , wn to obtain the elements of the global 5. Conclusions
WGM WGM
PCM H = { hi j }m
i, j=1
. Note that such global PCM H is not
guaranteed to be consistent. The main message of this paper is clear: Aggregation of local pri-
WGM
2. Compute the global priorities hj , j = 1, · · · , m, of alterna- orities of alternatives into global priorities in AHP should not be done
by the weighted arithmetic mean method. Instead, the weighted geo-
tives from the global PCM H by the GMM (4).
metric mean method should be used. We showed on several exam-
3. For each two alternatives Ai , A j , i, j = 1, · · · , m, compute the
WGM WGM ples the failure of the WAM method to properly reflect the prefer-
ratios hi /h j and construct the global consistent PCM ence information contained in local PCMs of alternatives. Contrar-
WGM WGM WGM
G = {gWGM
ij
}m , gWGM
i, j=1 i j
= hi /h j . ily, the WGM method proved to reflect the preference information
contained in local PCMs properly.
The following theorem demonstrates that all three procedures When the WAM is applied, the ratios of the global priorities
mentioned above lead to the same results. of alternatives are dependent on the choice of the normalization
technique for local priorities of alternatives. Even worse, with the
Theorem 5. The Procedures 1, 2 and 3 provide the same ratios of use of the WAM aggregation method, a rank reversal may occur
the global priorities of alternatives, i.e. they lead to the same global depending solely on the choice of the normalization technique.
WGM
consistent PCM of alternatives (GWGM = H WGM = G ). Moreover, when the WAM is applied for the aggregation of the
ratios of local priorities, the reciprocity condition is violated. The
Proof. First, let us show the equivalence of Procedure 1 and Pro- WAM exhibits also other instances of strange behavior, such as the
cedure 2. According to Procedure 1 we obtain for each two alter- inability to preserve equality of pairwise comparisons or producing
natives Ai , A j , i, j = 1, · · · , m: an unexpected global ranking from directly opposite preference in-
 wk formation, where indifference between the alternatives on a global

n
m
level would be expected by common sense. We showed the occur-
m
mkis
k=1 s=1 rence of these problems with the WAM aggregation on examples of
gWGM
ij =  wk . (14)
completely consistent PCMs to rule out the inconsistency of pair-

n
m
m
mkjs wise comparisons as the possible cause of these problems.
k=1 s=1 We demonstrated that the WGM aggregation does not suffer
According to Procedure 2 and by further adjusting, we obtain from the problems identified for the WAM aggregation. Most im-
portantly, the results to decision problems do not dependent on
⎛m ⎞wk  wk

m the choice of the normalization technique. In fact, the normaliza-
n ⎜
m
mkis m
mkis
 ⎟ 
n tion is no longer needed in AHP with the WGM aggregation. We
hWGM = ⎜  s=1 ⎟ = 
s=1
w k also showed that when the WGM is used for the aggregation of
ij ⎝ m ⎠
m
k=1 k=1
m
mkjs m
mkjs
s=1 s=1
106 J. Krejčí, J. Stoklasa / Expert Systems With Applications 114 (2018) 97–106

local priorities into global priorities of alternatives, then the ra- Duman, G., Tozanli, O., Kongar, E., & Gupta, S. M. (2017). A holistic approach for
tios of global priorities can equivalently be obtained by aggregating performance evaluation using quantitative and qualitative data: A food industry
case study. Expert Systems with Applications, 81, 410–422.
the ratios of local priorities. Moreover, when the GMM is used to de F. S. M. Russo, R., & Camanho, R. (2015). Criteria in ahp: a systematic re-
compute local priorities, then the following procedure provides the view of literature. Information Technology and Quantitative management, 55(1),
same results: first, aggregating the local PCMs with respect to each 1123–1132.
Ho, W. (2008). Integrated analytic hierarchy process and its applications - a litera-
criterion into a global PCM of alternatives by the WGM and, after- ture review. European Journal of Operational Research, 186(1), 211–228.
wards, computing the ratios of global priorities derived from this Ho, W., & Ma, X. (2018). The state-of-the-art integrations and applications of the
global PCM by the GMM. analytic hierarchy process. European Journal of Operational Research, 267(2),
399–414.
Given the extensive body of literature and research on AHP and
Jandová, V., Krejčí, J., Stoklasa, J., & Fedrizzi, M. (2017). Computing interval
given its numerous applications in various fields ranging from en- weights from incomplete pairwise comparison matrices of large dimension -
gineering and industry applications through social sciences appli- a weak consistency based approach. IEEE Transactions on Fuzzy Systems, 25(6),
1714–1728.
cations to applications in the medical sector, the relevance of the
Krejčí, J. (2018). Pairwise comparison matrices and their fuzzy extension: Multi-cri-
presented results is very high. The WAM aggregation can lead to teria decision making with a new fuzzy approach. Studies in Fuzziness and Soft
problems; unexplained changes in the ratios of global priorities Computing: 366. Cham:Springer.
and rank reversal even under constant sets of alternatives and cri- Lai, S. K. (1995). A preference-based interpretation of AHP. Omega, 23(4), 453–462.
Leskinen, P., & Kangas, J. (2005). Rank reversals in multi-criteria decision analysis
teria may occur. Contrarily, the WGM offers aggregation without with statistical modelling of ratio-scale pairwise comparisons. The Journal of the
the risk of unanticipated rank reversal and without the need to Operational Research Society, 56(7), 855–861.
justify the choice of a particular normalization technique to be Liberatore, M. J., & Nydick, R. L. (2008). The analytic hierarchy process in medical
and health care decision making: A literature review. European Journal of Oper-
used. ational Research, 189(1), 194–207.
In light of the findings presented in this paper we strongly dis- Lootsma, F. A. (1993). Scale sensitivity in the multiplicative AHP and SMART. Journal
courage the use of the WAM aggregation in AHP and advocate the of Multi-Criteria Decision Analysis, 2(1), 87–110.
Lootsma, F. A. (1996). A model for the relative importance of the criteria in the
WGM as the most appropriate method for the aggregation of infor- multiplicative AHP and SMART. European Journal of Operational Research, 94(1),
mation in AHP. 467–476.
Lootsma, F. A., & Schuijt, H. (1997). The multiplicative AHP, SMART and ELECTRE in
a common context. Journal of Multi-Criteria Decision Analysis, 6(1), 185–196.
Declarations of interest
Maleki, H., & Zahir, S. (2013). A comprehensive literature review of the rank reversal
phenomenon in the analytic hierarchy process. Journal of Multi-Criteria Decision
none. Analysis, 20(1), 141–155.
Nazari, S., Fallah, M., Kazemipoor, H., & Salehipour, A. (2018). A fuzzy inference-
fuzzy analytic hierarchy process-based clinical decision support system for di-
Acknowledgements agnosis of heart diseases. Expert Systems with Applications, 95, 261–271.
Saaty, T. L. (1977). A scaling method for priorities in hierarchical structures. Journal
We would like to acknowledge the funding received for this of Mathematical Psychology, 15(3), 234–281.
Saaty, T. L. (1980). The Analytic Hierarchy Process. New York: McGraw Hill.
project from the Finnish Strategic Research Council, grant num- Saaty, T. L. (1990). How to make a decision: The analytic hierarchy process. European
ber 313396 / MFG40 - Manufacturing 4.0 and from the inter- Journal of Operational Research, 48(1), 9–26.
nal grant agency of Palacký University Olomouc, grant number Saaty, T. L. (1994). Fundamentals of decision making and priority theory with the AHP.
Pittsburgh: RWS Publications.
IGA_F F _2018_002. Saaty, T. L. (2013). Decision making with the analytic network process: economic, po-
litical, social and technological applications with benefits, opportunities, costs and
References risks (2nd edition). New York: Springer.
Saaty, T. L., & Hu, G. (1998). Ranking by Eigenvector versus other methods in the
Barzilai, J. (1997). Deriving weights from pairwise comparison matrices. The Journal Analytic Hierarchy Process. Applied Mathematics Letters, 11(4), 121–125.
of the Operational Research Society, 48(12), 1226–1232. Saaty, T. L., & Vargas, L. G. (1984a). Comparison of eigenvalue, logarithmic least
Barzilai, J., d. Cook, W., & Golany, G. (1987). Consistent weights for judgements squares and least squares methods in estimating ratios. Mathematical Modelling,
matrices of the relative importance of alternatives. Operations Research Letters, 5(5), 309–324.
6(3), 131–134. Saaty, T. L., & Vargas, L. G. (1984b). The legitimacy of rank reversal. Omega, 12(5),
Belton, V., & Gear, T. (1983). On a short-coming of saaty’s method of analytic hier- 513–516.
archies. Omega, 11(1), 228–230. Saaty, T. L., & Vargas, L. G. (1993). Experiments on rank preservation and reversal in
Blaquero, R., Carrizosa, E., & Conde, E. (2006). Inferring efficient weights from pair- relative measurement. Mathematical and Computer Modelling, 17(4–5), 3–18.
wise comparison matrices. Mathematical Methods of Operations Research, 64(2), Subramanian, N., & Ramanathan, R. (2012). A review of applications of analytic hi-
271–284. erarchy process in operations management. International Journal of Production
Crawford, G., & Williams, C. (1985). A note on the analysis of subjective judgment Economics, 128(1), 215–241.
matrices. Journal of Mathematical Psychology, 29(2), 387–405. Triantaphyllou, E. (2001). Two new cases of rank reversals when the AHP and some
Dijkstra, T. K. (2013). On the extraction of weights from pairwise comparison matri- of its additive variants are used that do not occur with the multiplicative AHP.
ces. Central European Journal of Operational Research, 21(1), 103–123. Journal of Multi-Criteria Decision Analysis, 10(1), 11–25.
Vaidya, O. S., & Kumar, S. (2006). Analytic hierarchy process: An overview of appli-
cations. European Journal of Operational Research, 169(1), 1–29.

You might also like