Applied Soft Computing Journal 153 (2024) 111279

Contents lists available at ScienceDirect

Applied Soft Computing

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

Towards explainable TOPSIS: Visual insights into the effects of weights and
aggregations on rankings
Robert Susmaga, Izabela Szczęch ∗, Dariusz Brzezinski
Institute of Computing Science, Poznan University of Technology, Piotrowo 2, 60–965 Poznan, Poland

ARTICLE INFO ABSTRACT

Keywords: Multi-Criteria Decision Analysis (MCDA) is extensively used across diverse industries to assess and rank
TOPSIS alternatives. Among numerous MCDA methods developed to solve real-world ranking problems, TOPSIS
Weighted criteria ranking (Technique for Order Preference by Similarity to Ideal Solution) remains one of the most popular choices in
Interpretability
many application areas. TOPSIS calculates distances between the considered alternatives and two predefined
Visualization
ones, namely the ideal and the anti-ideal, and creates a ranking of the alternatives according to a chosen
Aggregated distance ranking
aggregation of these distances. However, interpreting the inner workings of TOPSIS is difficult, especially
when the number of criteria is large. To this end, recent research has shown that TOPSIS aggregations can be
expressed using the means (M) and standard deviations (SD) of alternatives, creating MSD-space, a tool for
visualizing and explaining aggregations. Even though MSD-space is highly useful, it assumes equally weighted
criteria, making it less applicable to real-world ranking problems. In this paper, we generalize MSD-space
to arbitrary weighted criteria by introducing the concept of WMSD-space defined by what is referred to
as weight-scaled means and standard deviations. We demonstrate that TOPSIS and similar distance-based
aggregation methods can be successfully illustrated in a plane and interpreted even when the criteria are
weighted, regardless of their number. The proposed WMSD-space offers thus a practical method for explaining
TOPSIS rankings in real-world decision problems.

The TOPSIS method has been widely used in many applications, in-
cluding logistics [7], manufacturing [8,9], marketing [10], sustainable
Good weight and measure are heaven’s treasure.
development [11], and engineering [12]; for a much broader survey of
[A proverb] TOPSIS and its applications see [13–15]. Much of the studies on TOP-
SIS focus on normalization and weighting procedures. Indeed, criteria
1. Introduction
weights are an important part of the TOPSIS method, because experts
use weights to encode their view of the significance of particular cri-
Being a sub-discipline of operations research, Multi-Criteria De-
cision Analysis (MCDA) aims to support decision makers in solving teria. However, for weights to have the expected effect, criteria values
problems that involve real-world objects (alternatives) evaluated on need to first be accordingly normalized. This is particularly important
multiple conflicting criteria. Often, this entails selecting the preferred when criteria values are expressed on different scales. The research
objects, assigning them to preference classes, or ranking them; for an of Opricovic and Tzeng [16] has analyzed the impact of different
extended overview of MCDA methods, models, and frameworks, see normalization procedures and different aggregation functions on the
recent reviews [1–5]. Among methods tackling the task of ranking final ranking obtained using the TOPSIS and VIKOR (VIseKriterijumska
many alternatives from the most preferred to the least preferred, a Optimizacija i kompromisno Resenje) methods. Similarly, Zavadskas
commonly chosen approach is TOPSIS (Technique for Order Preference et al. [17] describe the influence of a normalization method on the
by Similarity to Ideal Solution, Hwang and Yoon [6]). It is a popular final TOPSIS rankings. An alternative approach to criterion weighting
method to create rankings given predefined ideal/anti-ideal alterna- is considered by Chakraborty and Yeh [18]. The topic of weights is also
tives. More precisely, TOPSIS calculates distances from the ideal/anti-
an important part of studies on the Relative Ratio method [19], which
ideal to all the alternatives and produces non-negative real values,
estimates differences between alternatives to create a ranking that
which determine a linear pre-order that can be used for ranking.

∗ Corresponding author.
E-mail addresses: [email protected] (R. Susmaga), [email protected] (I. Szczęch), [email protected]
(D. Brzezinski).

https://doi.org/10.1016/j.asoc.2024.111279
Received 15 June 2023; Received in revised form 4 January 2024; Accepted 9 January 2024
Available online 19 January 2024
1568-4946/© 2024 Elsevier B.V. All rights reserved.
R. Susmaga et al. Applied Soft Computing 153 (2024) 111279

balances the distance from the ideal solution and the distance from the • In Section 3, we put forward our new visualization method.
anti-ideal solution. Similar approaches to weight balancing and relative In particular, in Sections 3.1 and 3.2 we show how arbitrary
closeness have been also proposed by Kuo [20] and Abootalebi et al. criteria weights can be re-scaled and used to generalize util-
[21]. With the use of the ROR (Robust Ordinal Regression) method- ity space into weighted utility space. In Section 3.3, we define
ology [22], TOPSIS has also been adapted to incorporate predefined weight-scaled means and standard deviations as equivalents of
relations between alternatives as a form of preferential information alternative means and standard deviations in the weighted utility
from the decision maker [23]. Many other interesting issues relating to space. As a result, we prove the IA-WMSD property (distance to
TOPSIS, including its combinations with other methods, its variations Ideal and Anti-ideal versus Weighted Mean and Standard Devi-
and adaptations, are described in [24–28]. ation property) and, in Section 3.4, we introduce WMSD-space
Last but not least are the attempts aimed at visualizing different (Weighted Mean-Standard Deviation space) that represents alter-
dependencies between processed data, and at visualizing the results natives in two dimensions regardless of the number of analyzed
of ranking methods. In the first group, there are typical projection- criteria and their weights. Finally, in Section 3.5, we visualize
based approaches, in which new features (attributes) are constructed WMSD-space and show how it can be used to express various ag-
in order to produce interpretable scatter-plots of some considered gregation functions using the weight-scaled means and standard
objects, e.g., Multi-dimensional Scaling (MDS), [29–31] or t-distributed deviations of alternatives.
Stochastic Neighbor Embedding (t-SNE), [32,33]. These methods are • In Section 4, we apply the proposed WMSD visualization to two
considered interpretable but are a form of lossy compression into case studies. We show how WMSD-space can be used to explore
at most three new features, making it impossible to reconstruct the the properties of a given dataset, compare the effects of weights
original data. In this sense these methods are imprecise. Additionally, defined by different experts and underline the implications of
there are approaches based on considering all the features simultane- using different aggregations.
ously, such as parallel coordinates or radar charts [34]. These are not • In Sections 5 and 6, we discuss our findings and limitations of
limited, but their interpretability decreases dramatically as the number the proposed method, summarize the paper, and suggest further
of features or alternatives grows. Independently, attempts at visualizing research.
the rankings are simply based on one-dimensional graphs depicting the
ranking positions of alternatives. To the best of our knowledge, apart 2. Preliminaries
from the method presented in this paper, there are no methods capable
The majority of research on the Technique for Order of Preference
of visualizing the precise ranking and the dependencies in the processed
by Similarity to Ideal Solution (TOPSIS) method involves a predefined,
data simultaneously.
finite set of 𝑚 entities (referred to as alternatives) described by a set of
As indicated by the referenced works, the bulk of the research on
𝑛 features (criteria). Consequently, the information can be effectively
TOPSIS is mainly focused on practical use cases and different ways of
represented in an 𝑚 × 𝑛 matrix of values, commonly known as the
performing criteria weighting and normalization. In addition to these
decision matrix. An example of such a decision matrix 𝐗 is illustrated
application-oriented studies, recently, we have formalized the inner
in Fig. 1A, comprising 𝑚 = 4 alternatives (students) characterized by
workings of TOPSIS by describing aggregations using the mean (M) and
𝑛 = 3 criteria (final grades in subjects).
standard deviation (SD) of each alternative. This allowed us to propose
Unlike research papers that focus on specific applications of TOPSIS,
a space for visualizing multi-criteria aggregations called MSD-space
our study will not be confined to a specific set of 𝑚 alternatives. Instead,
(Mean-Standard Deviation space) [35]. Notably, MSD-space is capable
we will delve into the general characteristics of all possible alternatives
of visualizing both data and aggregations, which directly translate
given a set of 𝑛 criteria. Our analysis of all conceivable alternative rep-
to TOPSIS rankings. However, MSD-space assumes equally weighted
resentations is influenced by strategies designed for visually inspecting
criteria, making it less applicable to real-world ranking tasks, which, as
general properties of machine learning metrics [44–47]. To conduct this
a recent analysis of TOPSIS applications shows [13], practically always
analysis, in this section, we will provide the essential definitions needed
involve decision makers defining weights. By weighing the criteria,
to formalize the TOPSIS procedure. Furthermore, we will review the
they introduce preferential information and alter the influence of the conclusions from our previous research on interpreting TOPSIS by
criteria on the final TOPSIS ranking. Oftentimes, a separate MCDA revisiting the definitions of utility space, IA-MSD property (distance to
method is applied to elucidate the criteria weights [36,37]. That is Ideal and Anti-ideal versus Mean and Standard Deviation Property),
why applying weights is crucial for modern TOPSIS variations and and MSD-space (Mean-Standard Deviation space) [35]. The notation
real-world ranking tasks requiring explainability. introduced in the following paragraphs will be used to generalize
Explainability, regarded as methods that allow humans to under- utility space and MSD-space into their weighted counterparts in the
stand and trust the results of algorithms, keeps gaining a lot of at- subsequent sections of this paper.
tention in artificial intelligence [38–40] and multi-criteria decision
support [41–43]. The ability to explain the result of algorithms is 2.1. Formalizing TOPSIS using the utility space
of particular importance in practical applications that impact society.
However, such applications often involve domain experts who incor- TOPSIS [6] is a multi-criteria decision analysis (MCDA) method that
porate their preferential bias in the form of weights imposed on the ranks objects (alternatives) from the best to the worst in terms of their
criteria. Since our existing MSD-space visualizing TOPSIS was designed distance to ideal and anti-ideal points. The descriptions of alternatives
for unweighted criteria, we pose the following research question: Is it with respect to considered attributes is commonly given in the form of
possible to propose a visualization method that generalizes MSD-space to vectors. Among attributes typically used in MCDA, there are criteria,
weighted criteria? characterized by preference-ordered domains.
To answer the above research question, the specific goal of this The main actions performed by TOPSIS can be summarized as:
study is to extend the MSD-space methodology to problems with ar-
bitrarily defined weights of criteria. Moreover, we aim to explain how 1. Prepare the representations of alternatives in terms of crite-
weights affect rankings of alternatives under various TOPSIS aggrega- ria. Apart from forming the decision matrix, this part of the
tions and how the effects of weights provided by multiple experts can procedure may also normalize the criteria and incorporate the
be compared. The detailed contributions of this paper are as follows: user-given weights that actually constitute their preferential in-
formation.
• In Section 2, we formalize the TOPSIS procedure and recall the 2. Determine two reference points, ideal and anti-ideal, and
definitions of utility space, MSD-space, and their properties. verify how far each alternative is from them.

2
R. Susmaga et al. Applied Soft Computing 153 (2024) 111279

Fig. 1. The dataset that will serve as the running example for explaining different representations of objects analyzed in this paper. (A) The original dataset (decision matrix)
describing 𝑚 = 4 students (alternatives) using final grades from 𝑛 = 3 subjects (criteria). (B) The same dataset depicted as a subset of the criteria space, i.e., of all possible alternatives
described by the three criteria describing students. (C) The same alternatives presented as a subset of utility space, the re-scaled equivalent of criteria space. (D) The analyzed
students represented in MSD-space, a space defined by the mean (M) and standard deviation (SD) of the utility space descriptions of the alternatives. (E) Alternatives represented
in weighted utility space, with weights 𝐰 = [0.5, 0.6, 1.0]. (F) Alternatives represented in WMSD-space, a space defined by the weight-scaled mean (WM) and weight-scaled standard
deviation (WSD) of the weighted utility space descriptions of the alternatives.

3. Rank the alternatives with respect to some aggregation func- interpretable results: the most preferred values are always transformed
tion that combines the distances between the alternatives and to ones, while the least preferred values are always transformed to
the ideal/anti-ideal points. zeros. Second, after proper selection of the min and max values for each
criterion, application of the min–max transformation liberates TOPSIS
Preparing representations of alternatives. TOPSIS starts with encoding from the peril of the so-called rank-reversal problem, to which original
real-world objects (e.g., students described by criteria referring to their TOPSIS is unfortunately predisposed [48].
grades) into a decision matrix 𝐗 (Fig. 1A). The decision matrix is a finite Since the 𝑈 𝑆 is the space of all conceivable representations (images)
subset of the criteria space 𝐶𝑆 (Fig. 1B), where if a criterion  belongs of alternatives, particular decision matrices are simply represented as
to the set of all possible criteria K ( ∈ K), then its domain is a real- finite subsets of 𝑈 𝑆. Aiming at formalizing general dataset-independent
valued interval  = [𝑣𝑚𝑖𝑛 , 𝑣𝑚𝑎𝑥 ]. Since TOPSIS is based on calculating properties, we shall deploy 𝑈 𝑆 in all further considerations.
distances, the bounds of the interval need to be finite. Additionally,
criteria, as attributes with preferentially ordered domains, may differ Determination of the ideal/anti-ideal points and distance calculation. Given
in their preference types (gain or cost), with the least preferred value a set of criteria 𝒦 , |𝒦 | = 𝑛 ≥ 1, the utility space is an 𝑛-dimensional
denoted as 𝑣∗ and the most preferred value as 𝑣∗ . Vectors [𝑣∗1 , 𝑣∗2 , … , 𝑣∗𝑛 ] hypercube [0, 1] × [0, 1] × ⋯ × [0, 1] with 2𝑛 vertices of the form
and [𝑣1∗ , 𝑣2∗ , … , 𝑣𝑛∗ ] will be referred to as the ideal (𝐼) and anti-ideal [𝑧1 , 𝑧2 , … , 𝑧𝑛 ], where 𝑧𝑗 ∈ {0, 1}. Moreover, for each alternative rep-
resentation 𝐸 ∈ 𝐶𝑆 there exists 𝐮 ∈ 𝑈 𝑆 such that 𝐮 is the image of 𝐸
(𝐴) points, respectively.
under the re-scaling transformation—if 𝐸 = [𝑣1 , 𝑣2 , … , 𝑣𝑛 ] ∈ 𝐶𝑆, then
Working on criteria with varying domains and types can make
[1 (𝑣1 ), 2 (𝑣2 ), … , 𝑛 (𝑣𝑛 )] ∈ 𝑈 𝑆. In particular, 𝑈 𝑆 contains vectors
the analysis more troublesome and reduce the meaningfulness of the
𝟏 = [1, 1, … , 1] and 𝟎 = [0, 0, … , 0], which are the respective images of
results, thus a criteria transformation is often applied. In this paper, we
the ideal point and anti-ideal point. In our running example, there are
will use a min–max re-scaling that transforms the criteria space into the
three criteria, thus the points 𝟏 = [1, 1, 1] and 𝟎 = [0, 0, 0] represent in
utility space 𝑈 𝑆 (Fig. 1C) using the function  ∶  → [0, 1]:
𝑈 𝑆 the ideal point and anti-ideal point, respectively (see Fig. 1C).
• given a domain  = [𝑣𝑚𝑖𝑛 , 𝑣𝑚𝑎𝑥 ] = [𝑣∗ , 𝑣∗ ] of a criterion  ∈ K of With the 𝟏 and 𝟎 points at hand, TOPSIS calculates how far each
type gain, the re-scaling function  associated with  is defined alternative is from them. To perform this operation, the Euclidean
𝑣−𝑣 distance measure is used.
as  (𝑣) = 𝑣∗ −𝑣∗ for 𝑣 ∈ ,
∗ Let 𝐚 = [𝑎1 , 𝑎2 , … , 𝑎𝑛 ] and 𝐛 = [𝑏1 , 𝑏2 , … , 𝑏𝑛 ] be vectors (always as-
• given a domain  = [𝑣𝑚𝑖𝑛 , 𝑣𝑚𝑎𝑥 ] = [𝑣∗ , 𝑣∗ ] of a criterion  ∈ K of sumed to be row vectors). The dot product of these vectors is defined as
type cost, the re-scaling function  associated with  is defined ∑
𝑣 −𝑣 ∗ 𝐚 ⋅ 𝐛𝑇 = [𝑎1 , 𝑎2 , … , 𝑎𝑛 ] ⋅ [𝑏1 , 𝑏2 , … , 𝑏𝑛 ]𝑇 = 𝑛𝑗=1 𝑎𝑗 ⋅ 𝑏𝑗 (the second vector is
as  (𝑣) = 𝑣∗ −𝑣 for 𝑣 ∈ . transposed to allow matrix multiplication). This serves to define the Eu-
∗
clidean norm of vector √∑ 𝐚 = [𝑎1 , 𝑎2 , … , 𝑎𝑛 ] as its dot product with itself:
The  (⋅) function is introduced to simplify further TOPSIS process- √
𝑛
ing without the loss of generality and is independent of decision matrix ‖𝐚‖2 = 𝐚 ⋅ 𝐚𝑇 = 2
𝑗=1 𝑎𝑗 . Finally, the Euclidean distance between
normalization that could be performed by a user. The motivations vectors 𝐚 = [𝑎1 , 𝑎2 , … , 𝑎𝑛 ] and 𝐛 = [𝑏1 , 𝑏2 , … , 𝑏𝑛 ] is defined √ as the Eu-
behind introducing the min–max transformation are two-fold. First, clidean norm of their difference: 𝛿2 (𝐚, 𝐛) = ‖ 𝐚 − 𝐛 ‖ = ∑𝑛 (𝑎𝑗 − 𝑏𝑗 )2 .
‖ ‖2 𝑗=1
the counter-domain of the transformation is the [0,1] interval, which Now, the Euclidean distance is used to compute 𝛿2 (𝐮, 𝟎) and 𝛿2 (𝐮, 𝟏) for
combined with the value reversal for cost type criteria produces well each 𝐮 ∈ 𝑈 𝑆.

3
R. Susmaga et al. Applied Soft Computing 153 (2024) 111279

Fig. 2. A depiction of the IA-MSD property in 𝑈 𝑆 and MSD-space for a three-dimensional problem. (A) Vector orthogonality depicted in 𝑈 𝑆. (B) Illustration of the IA-MSD
property in MSD-space. The re-scaled 𝛿201 lengths of vectors 𝐮 and 𝐮 − 𝐮 from panel A correspond to the values of 𝑚𝑒𝑎𝑛(𝐮) and 𝑠𝑡𝑑(𝐮) depicted in MSD-space. (C) Color encoding
of the aggregation function 𝖱(𝐮), with blue representing the least preferred and red the most preferred values.

The maximal Euclidean distance in 𝑈 𝑆, which √ extends between Moreover, in [35] we have shown that the lengths of the above-
vectors 𝟏 and 𝟎, is dependent on 𝑛 and equals 𝑛. For our analyses mentioned vectors can be expressed as follows:
to be 𝑛-independent and thus easily interpretable regardless of 𝑛, we
𝛿 (𝐚,𝐛) 𝛿201 (𝐮, 𝟎) = 𝑚𝑒𝑎𝑛(𝐮), (8)
define a re-scaled Euclidean distance as 𝛿201 (𝐚, 𝐛) = 2√ , ranging
√ 𝑛
𝛿201 (𝐮, 𝟏) = 1 − 𝑚𝑒𝑎𝑛(𝐮), (9)
always between [0, 1] (instead of [0, 𝑛]). The re-scaled distances of an
alternative’s image 𝐮 ∈ 𝑈 𝑆 to the ideal and anti-ideal point will be 𝛿201 (𝐮, 𝐮) = 𝑠𝑡𝑑(𝐮). (10)
denoted as 𝛿201 (𝐮, 𝟏) and 𝛿201 (𝐮, 𝟎), respectively.
These characteristics of 𝑈 𝑆 allowed us to formulate the IA-MSD
Ranking alternatives according to an aggregation. The distances of each property.
alternative’s representation to the reference points are combined with
respect to some chosen aggregation function, the value of which natu- Theorem 1 (IA-MSD Property).
rally forms a ranking of the alternatives (precisely: a linear pre-order). √
𝛿201 (𝐮, 𝟎) = 𝑚𝑒𝑎𝑛(𝐮)2 + 𝑠𝑡𝑑(𝐮)2 , (11)
We shall focus on three aggregations, defined in terms of 𝛿201 (𝐮, 𝟏) √
(‘distance to the ideal in 𝑈 𝑆’) and 𝛿201 (𝐮, 𝟎) (‘distance to the anti-ideal 𝛿201 (𝐮, 𝟏) = (1 − 𝑚𝑒𝑎𝑛(𝐮))2 + 𝑠𝑡𝑑(𝐮)2 . (12)
in 𝑈 𝑆’) as: The IA-MSD property shows that the distances of an alternative to
𝖨(𝐮) = 1 − 𝛿201 (𝐮, 𝟏), (1) the ideal and anti-ideal point are functions of the mean and standard
deviation of the alternative. This interesting dependency between the
𝖠(𝐮) = 𝛿201 (𝐮, 𝟎), (2) distances of alternatives to the predefined ideal (𝐼) and anti-ideal (𝐴)
𝛿201 (𝐮, 𝟎) points on the one hand and 𝑚𝑒𝑎𝑛(𝐮) and 𝑠𝑡𝑑(𝐮) on the other, inspired
𝖱(𝐮) = , (3) us to define MSD-space. This space uses the mean (M) and standard
𝛿201 (𝐮, 𝟏) + 𝛿201 (𝐮, 𝟎)
deviation (SD) of an alternative’s 𝑈 𝑆 representation as its constituents
where 𝐮 ∈ 𝑈 𝑆 is the image of an alternative in 𝑈 𝑆. All three aggrega- (Fig. 2B).
tions generate values belonging to [0, 1]. The 𝖨(𝐮) aggregation is based
solely on the distance to the ideal point, the 𝖠(𝐮) aggregation is based Definition 1 (MSD-space).
on the distance to the anti-ideal point, whereas the relative distance (the
standard aggregation of TOPSIS), denoted as 𝖱(𝐮), takes both previous MSD-space = {[𝑚𝑒𝑎𝑛(𝐮), 𝑠𝑡𝑑(𝐮)] ∣ 𝐮 ∈ 𝑈 𝑆} (13)
distances into account. Using 1 − 𝛿201 (𝐮, 𝟏) instead of a straightforward The MSD-space is a two-dimensional space and, therefore, can be
distance 𝛿201 (𝐮, 𝟏) in the 𝖨(𝐮) aggregation, serves only as a means to visualized in a plane where the mean (M) of an alternative defines its
have all aggregations as functions to be maximized. Although only the position on the 𝑥-axis and the standard deviation (SD) the position on
𝖱(𝐮) aggregation is predominantly deployed in TOPSIS, it is defined the 𝑦-axis. Since MSD-space is a transformation1 of 𝑈 𝑆 which is [0, 1]-
on the basis of 𝖨(𝐮) and 𝖠(𝐮), and as such inherits from them its main bounded, the range of values of M and SD is also bounded. As a result,
properties. For this reason, all three aggregations will be examined in for a given number of criteria 𝑛, there is a limited range of attainable
this paper. means and standard deviations, which forms the shape of MSD-space
(Fig. 2B).
2.2. The IA-MSD property and MSD-space Moreover, the IA-MSD property makes it possible to define all
TOPSIS aggregations in terms of 𝑚𝑒𝑎𝑛(𝐮) and 𝑠𝑡𝑑(𝐮):
√
Given a representation of an alternative in utility space 𝐮 ∈ 𝑈 𝑆, let: 𝖨(𝐮) = 1 − (1 − 𝑚𝑒𝑎𝑛(𝐮))2 + 𝑠𝑡𝑑(𝐮)2 , (14)
∑
𝑛 √
𝑠𝑢𝑚(𝐮) = 𝑢𝑗 , (4) 𝖠(𝐮) = 𝑚𝑒𝑎𝑛(𝐮)2 + 𝑠𝑡𝑑(𝐮)2 , (15)
√
𝑗=1 2
𝑚𝑒𝑎𝑛(𝐮) + 𝑠𝑡𝑑(𝐮)2
𝑠𝑢𝑚(𝐮) 𝖱(𝐮) = √ √ . (16)
𝑚𝑒𝑎𝑛(𝐮) = , (5) (1 − 𝑚𝑒𝑎𝑛(𝐮))2 + 𝑠𝑡𝑑(𝐮)2 + 𝑚𝑒𝑎𝑛(𝐮)2 + 𝑠𝑡𝑑(𝐮)2
𝑛
‖ ‖2 Since all the discussed TOPSIS aggregations are functions of two param-
‖𝐮 − 𝐮‖
‖ ‖2 eters, 𝑚𝑒𝑎𝑛(𝐮) and 𝑠𝑡𝑑(𝐮), one can visualize their values in MSD-space
𝑣𝑎𝑟(𝐮) = , with 𝐮 = [𝑚𝑒𝑎𝑛(𝐮), 𝑚𝑒𝑎𝑛(𝐮), … , 𝑚𝑒𝑎𝑛(𝐮)], (6)
𝑛 using a color map (Fig. 2C). Using such a visualization, one can analyze
√
𝑠𝑡𝑑(𝐮) = 𝑣𝑎𝑟(𝐮). (7) how the preferences expressed by different aggregations change with
varying values of M and SD.
With the above notation, in our previous paper [35], we have used the
fact that for every 𝐮 ∈ 𝑈 𝑆 vectors 𝐮 − 𝟎 and 𝐮 − 𝐮 as well as 𝐮 − 𝐮 and
𝟏 − 𝐮 are orthogonal, and, therefore, one can apply the Pythagorean 1
M is a linear function of elements of 𝐮 ∈ 𝑈 𝑆, while SD is a square root
theorem to relate these vectors (Fig. 2A). of a second order polynomial of elements of 𝐮 ∈ 𝑈 𝑆.

4
R. Susmaga et al. Applied Soft Computing 153 (2024) 111279

Table 1
Notation summary.
Symbol Definition
𝐶𝑆 Criteria space.
𝐗 ⊂ 𝐶𝑆 Decision matrix.
𝐸 ∈ 𝐶𝑆 Alternative.
𝑛 Number of criteria.
𝑚 Number of alternatives in 𝐗.
K Set of criteria.
∈K Criterion.
𝑣 Criterion value.
 = [𝑣𝑚𝑖𝑛 , 𝑣𝑚𝑎𝑥 ] Domain of a criterion.
MSD 𝑣∗ ; 𝑣∗ Least preferred; most preferred criterion value.
 ∶  → [0, 1] Rescaling function that transforms the criteria space into the utility space.
𝑈𝑆 Utility space.
𝐷𝟎𝟏 Diagonal of 𝑈 𝑆.
𝐮 ∈ 𝑈𝑆 Representation (image) of an alternative in utility space (normalized alternative).
𝐮 Vector of means ([𝑚𝑒𝑎𝑛(𝐮), … , 𝑚𝑒𝑎𝑛(𝐮)]).
|⋅| Set cardinality.
‖ ⋅ ‖2 Euclidean norm.
𝛿2 (𝑎, 𝑏) Euclidean distance.
𝛿201 (𝑎, 𝑏) 0-1 re-scaled Euclidean distance.
𝑚𝑒𝑎𝑛(𝐮), M Mean of normalized alternative.
𝑠𝑡𝑑(𝐮), SD Standard deviation of normalized alternative.
𝖨(𝐮), 𝖠(𝐮), 𝖱(𝐮) Ideal (𝖨); anti-ideal (𝖠); and relative distance (𝖱) aggregation for unweighted criteria.
𝐰 Criteria weights.
𝑉𝑆 Weighted utility space.
𝐷𝟎𝐰 Diagonal of 𝑉 𝑆.
𝐯∈𝑉𝑆 Representation (image) of an alternative in weighted utility space (weighted alternative).
WMSD 𝐯 Vector of weighted means (𝐰◦𝐮).
𝑛𝑝 Number of non-zero weighted criteria.
‖𝐰‖2
𝑠 = 𝑚𝑒𝑎𝑛(𝐰) Weight-scaling factor.
𝛿 (𝐚,𝐛)
𝛿𝐰01 (𝐚, 𝐛) = 2 𝑠 Weight-scaled Euclidean distance.
𝑚𝑒𝑎𝑛01 𝐰 (𝐯), WM Weight-scaled mean of a weighted alternative.
𝑠𝑡𝑑𝐰01 (𝐯), WSD Weight-scaled standard deviation of a weighted alternative.
𝖨𝐰 (𝐯); 𝖠𝐰 (𝐯); 𝖱𝐰 (𝐯) Ideal (𝖨); anti-ideal (𝖠); relative distance (𝖱) aggregation for weighted criteria.

3. WMSD-space: A method for visualizing TOPSIS aggregations Next, notice that relation ‘≥’ admits two disjoint sub-cases: ‘=’ and
and rankings ‘>’. Whenever 𝑤𝑖 = 0, then the 𝑖th criterion is in practice ‘zeroed’ and
thus not taken into account in any further considerations. On the other
The concepts of utility space, IA-MSD property, and MSD-space hand, 𝑤𝑖 ≠ 0 means that the 𝑖th criterion is taken into account. This
proposed in [35], and recalled in the previous section, assume that explains the second assumption, namely: 𝑤𝑖 ≠ 0 for at least one 𝑖 (‘at
all of the criteria are equally important. In practice, this is rarely the least one non-zero’). The assumption ensures that the undesired case of
case, as criteria are very often differentiated and thus assigned non- all zero weights cannot occur. If all weights were zero, the descriptions
uniform weights by experts. In this section, we will formalize criteria of all alternatives would be identical and there would be no need to
weighting and show how the utility space 𝑈 𝑆 can be transformed to subject the alternatives to any ranking method.
its weighted counterpart 𝑉 𝑆. We will then define the weight-scaled Finally, we shall consider bounded, and thus finite, weights (‘all
mean (WM) and weight-scaled standard deviation (WSD), which will finite’). Such a constraint may be implemented by ‘normalizing’ the
be used to introduce a weighted version of the MSD-space, called the weights to satisfy max𝑛𝑖=1 𝑤𝑖 = 1 (‘max is 1’). This is achieved by
WMSD-space. To make it easier to follow and compare the introduced dividing all of the weights by their maximum, which is positive thanks
concepts, Table 1 summarizes the mathematical notation used to define to the two previous assumptions. Such an operation does not affect the
‘performance’ of weights, because their relative ratios remain the same
MSD-space in the previous section and WMSD-space presented in this
after any division by a positive number.
section.
As a result, the three assumptions (‘all non-negative’, ‘at least one
non-zero’, and ‘all finite’), are expressed with only two conditions:
3.1. Normalized criteria weights
𝐰 ≥ 𝟎 (‘all non-negative’), (17)
Let 𝐰 = [𝑤1 , 𝑤2 , … , 𝑤𝑛 ] be a vector of real values, acting as criteria 𝑛
max 𝑤𝑖 = 1 (‘max is 1’, i.e. ‘at least one non-zero’ and ‘all finite’).
weights (referred to as weights), where 𝑛 is the number of criteria. 𝑖=1
According to common MCDA practice, these weights are assumed to (18)
be:
It should be stressed that even though the postulated assump-
• all non-negative, tions exclude situations in which all weights are zero, they do not
• at least one non-zero, exclude situations in which some weights are zero. As stated above,
• all finite. in such a situation, the criteria corresponding to zero weights are in
practice eliminated from all further considerations. In result, ‘zeroing’
These assumptions have the following justifications. weights may be viewed as a form of ‘criterion selection’ (only criteria
Non-negativity results from the fact that the weight expresses the corresponding to positive weights are selected).
magnitude of the criterion’s relative importance (greater weight, In the following sections, we will assume that criteria weights
greater importance). In particular, if 𝑤𝑖 > 𝑤𝑗 , then the 𝑖th criterion is adhere to the ‘all non-negative’ and ‘max is 1’ conditions. This can be
expected to have more influence on the final result of the method than easily implemented in practice, as any set of real values satisfying ‘all
the 𝑗th criterion. Therefore, the first considered assumption is: 𝑤𝑖 ≥ 0 non-negative’, ‘at least one non-zero’ and ‘all finite’ can be re-scaled to
for all 𝑖. be ‘all non-negative’ and ‘max is 1’.

5
R. Susmaga et al. Applied Soft Computing 153 (2024) 111279

Fig. 3. Vector orthogonality presented in (A) 𝑈 𝑆 and (B) 𝑉 𝑆, for 𝑛 = 𝑛𝑝 = 2. The weight vector used to transform the presented 𝑈 𝑆 into 𝑉 𝑆 is 𝐰 = [1.0, 0.5]. The diagonal 𝐷𝟎𝟏 is
the blue line segment between vertices 𝟎 and 𝟏 in 𝑈 𝑆. Analogously, 𝐷𝟎𝐰 is the blue line segment between vertices 𝟎 and 𝐰 in 𝑉 𝑆.

3.2. 𝑉 𝑆: The weighted utility space which makes it heavily dependent on 𝐰. To make the maximal distance
in 𝑉 𝑆 independent of at least some characteristics of 𝐰, we define the
While 𝑈 𝑆 has the shape of a hypercube, it changes as soon as it 01 . Given any 𝐯 , 𝐯 ∈ 𝑉 𝑆:
re-scaled weighted Euclidean distance 𝛿𝐰 𝟏 𝟐
becomes non-uniformly weighted, i.e., as soon as non-uniform criteria
01 𝛿2 (𝐯𝟏 , 𝐯𝟐 )
weights are applied. By weighing the criteria, one introduces preferen- 𝛿𝐰 (𝐯𝟏 , 𝐯𝟐 ) = , (21)
𝑠
tial information from the decision maker and alters the influence of the
criteria on the final TOPSIS ranking (as will be later illustrated in case where 𝑠 is a weight-scaling factor
studies in Section 4). Thus, the shape of the weighted version of 𝑈 𝑆, ‖𝐰‖2
𝑠= . (22)
in which all TOPSIS operations are actually performed, generalizes to a 𝑚𝑒𝑎𝑛(𝐰)
hyperrectangle, with the special case of the hypercube obtained for all Observe that the value of 𝑠 always exists (because its denominator is
weights equal to one. Notice that the application of weights is in fact non-zero), and never equals zero (because its nominator is non-zero). It
a linear operation. is because ‖𝐰‖2 > 0 and 𝑚𝑒𝑎𝑛(𝐰) > 0 are guaranteed by 𝐰 ≠ 𝟎 (implied
Given vectors 𝐚 = [𝑎1 , 𝑎2 , … , 𝑎𝑛 ] and 𝐛 = [𝑏1 , 𝑏2 , … , 𝑏𝑛 ], let 𝐚◦𝐛 by the assumptions concerning 𝐰). In particular, for 𝐰 = 𝟏, 𝑠 becomes:
denote their element-wise (Hadamard) product, i.e., 𝐚◦𝐛 = [𝑎1 ⋅ 𝑏1 , 𝑎2 ⋅ √
‖𝐰‖2 ‖𝟏‖ 𝑛 √
𝑏2 , … , 𝑎𝑛 ⋅𝑏𝑛 ]. Now, let 𝐰 = [𝑤1 , 𝑤2 , … , 𝑤𝑛 ] be a vector of weights. Given ‖ ‖2
𝑠= = = = 𝑛. (23)
these weights and an 𝑛-dimensional 𝑈 𝑆 we define 𝑚𝑒𝑎𝑛(𝐰) 𝑚𝑒𝑎𝑛(𝟏) 1

𝑉 𝑆 = {𝐯 ∶ 𝐯 = 𝐰◦𝐮, 𝐮 ∈ 𝑈 𝑆}. (19) which is the value of the divisor in the definition of re-scaled Euclidean
distance 𝛿201 . As a result 𝛿𝐰
01 is a generalization and thus a full analog
The weighted utility space 𝑉 𝑆 (Fig. 3) is the image of 𝑈 𝑆 with (in of the re-scaled Euclidean distance 𝛿201 .
particular): Thanks to 𝑠, given any 𝐰 and any 𝐯𝟏 , 𝐯𝟐 ∈ 𝑉 𝑆:
• 𝟎 ∈ 𝑉 𝑆 being the image of 𝟎 ∈ 𝑈 𝑆, ‖𝐰‖2 ‖𝐰‖
01
𝛿𝐰 (𝐯𝟏 , 𝐯𝟐 ) ∈ [0, ] = [0, ‖𝐰‖ 2 ] = [0, 𝑚𝑒𝑎𝑛(𝐰)], (24)
• 𝐰 ∈ 𝑉 𝑆 being the image of 𝟏 ∈ 𝑈 𝑆. 𝑠 2
𝑚𝑒𝑎𝑛(𝐰)
By the assumptions of 𝐰 (‘max is 1’), if 𝐮 ∈ 𝑈 𝑆, then 𝐯 = 𝐰◦𝐮 ≤ 𝐮. In as opposed to 𝛿2 (𝐯𝟏 , 𝐯𝟐 ) ∈ [0,‖𝐰‖2 ]. Therefore, the maximal re-scaled
result, 𝟏 ∉ 𝑉 𝑆 in general (this only happens when 𝐰 = 𝟏, since then weighted Euclidean distance 𝛿𝐰 01 in 𝑉 𝑆 is independent of at least some
𝑉 𝑆 = 𝑈 𝑆 and 𝟏 ∈ 𝑉 𝑆 is the image of 𝟏 ∈ 𝑈 𝑆). Clearly, 𝑉 𝑆 ⊆ 𝑈 𝑆, with characteristics of 𝐰.
𝑉 𝑆 = 𝑈 𝑆 only for 𝐰 = 𝟏; in all other cases 𝑉 𝑆 ⊂ 𝑈 𝑆. Additionally, Notice that the assumption ‘max is 1’ ensures 0 < 𝑚𝑒𝑎𝑛(𝐰) ≤ 1, so
while 𝑈 𝑆 is an 𝑛-dimensional hypercube, 𝑉 𝑆 is represented by a [0, 𝑚𝑒𝑎𝑛(𝐰)] is a proper interval, additionally satisfying
𝑛𝑝 -dimensional hyperrectangle, where √ [0, 𝑚𝑒𝑎𝑛(𝐰)] ⊆
01 becomes
[0, 1]. It is also clear that for 𝐰 = 𝟏, in which case 𝑠 = 𝑛, 𝛿𝐰
01 (𝐯 , 𝐯 ) = 𝛿 01 (𝐯 , 𝐯 ) = 𝛿2 (𝐯
𝛿201 : 𝛿𝐰
,𝐯 )
√𝟏 𝟐 = 𝛿 01 (𝐯𝟏 , 𝐯𝟐 ).
𝑛𝑝 = |{𝑖 ∶ 𝑤𝑖 > 0}|. (20) 𝟏 𝟐 𝟏 𝟏 𝟐 𝑛 2

The assumption ‘at least one non-zero’ of 𝐰 ensures that 𝑛𝑝 ≥ 1, so in

3.3. The IA-WMSD property in 𝑉 𝑆
general, 1 ≤ 𝑛𝑝 ≤ 𝑛. In the most often satisfied case of 𝑤𝑖 > 0 for all 𝑖,
𝑛𝑝 = 𝑛.
Given two vectors 𝐚 and 𝐛 ≠ 𝟎, let us define [49]:
Similarly to 𝑈 𝑆, two vertices, namely 𝐰 and 𝟎 (images of 𝟏 and 𝟎
𝐚⋅𝐛 𝑇
from 𝑈 𝑆), are of special interest in 𝑉 𝑆, as they constitute the endpoints • vector 𝐚 ↘ 𝐛 = 𝐛, the orthogonal projection of 𝐚 onto 𝐛,
of the segment that will be referred to as the main diagonal of 𝑉 𝑆 and ‖𝐛‖22
denoted as 𝐷𝟎𝐰 (Fig. 3). The diagonal 𝐷𝟎𝐰 is an image of 𝐷𝟎𝟏 = {𝐝 = • vector 𝐚 ↗ 𝐛 = 𝐚 − 𝐚 ↘ 𝐛, the orthogonal rejection of 𝐚 from 𝐛.
𝑑 ⋅ 𝟏 ∣ 𝑑 ∈ [0, 1]} (the diagonal of 𝑈 𝑆), since it contains all vectors
Notice that ‖ ‖
‖𝐛‖2 ≠ 0 is guaranteed by 𝐛 ≠ 𝟎, so the projection vector
{𝐰◦𝐝 ∣ 𝐝 ∈ 𝐷𝟎𝟏 }, in particular 𝐰 (for 𝐝 = 𝟏) and 𝟎 (for 𝐝 = 𝟎). Of course
always exists, and thus the rejection vector always exists. By definition,
𝐷𝟎𝐰 thus equals {𝑑 ⋅ 𝐰 ∣ 𝑑 ∈ [0, 1]}.
vectors 𝐚 ↘ 𝐛 and 𝐚 ↗ 𝐛 are orthogonal, i.e. (𝐚 ↘ 𝐛)(𝐚 ↗ 𝐛)𝑇 = 0.
As was the case with 𝐷𝟎𝟏 ⊆ 𝑈 𝑆, 𝐷𝟎𝐰 satisfies 𝐷𝟎𝐰 ⊆ 𝑉 𝑆, but is
Now, given any 𝐯 = 𝐮◦𝐰 ∈ 𝑉 𝑆, we define:
dependent on 𝑛𝑝 (rather than 𝑛), as 𝐷𝟎𝐰 ⊂ 𝑉 𝑆 for 𝑛𝑝 > 1 and 𝐷𝟎𝐰 = 𝑉 𝑆
for 𝑛𝑝 = 1. ‖𝐯 ↘ 𝐰‖2
𝑚𝑒𝑎𝑛01
𝐰 (𝐯) = , which we refer to as the
The maximal Euclidean distance 𝛿2 in 𝑉 𝑆 is that of 𝐷𝟎𝐰 , which 𝑠
extends between 𝟎 and 𝐰 (Fig. 3). This maximal distance equals ‖𝐰‖2 , 𝑤𝑒𝑖𝑔ℎ𝑡-𝑠𝑐𝑎𝑙𝑒𝑑 𝑚𝑒𝑎𝑛 (WM) of 𝐯, (25)

6
R. Susmaga et al. Applied Soft Computing 153 (2024) 111279

Fig. 4. An illustration of the IA-WMSD property in (A) 𝑉 𝑆 and (B) WMSD-space, for 𝐰 = [1.0, 0.5] and a point 𝐯 = [0.75, 0.25]. The illustration shows how the re-scaled lengths 𝛿𝐰01
of vectors 𝐯 and 𝐯 − 𝐯 are equal to the weight-scaled mean (WM) and standard deviation (WSD), which define the WMSD-space. The diagonal 𝐷𝟎𝐰 is the blue line segment between
vertices 𝟎 and 𝐰.

‖𝐯 ↗ 𝐰‖2
𝑠𝑡𝑑𝐰01 (𝐯) = , which we refer to as the Simultaneously, 𝑠𝑡𝑑𝐰01 (𝐯) satisfies:
𝑠
𝑤𝑒𝑖𝑔ℎ𝑡-𝑠𝑐𝑎𝑙𝑒𝑑 𝑠𝑡𝑎𝑛𝑑𝑎𝑟𝑑 𝑑𝑒𝑣𝑖𝑎𝑡𝑖𝑜𝑛 (WSD) of 𝐯. (26) 𝑠𝑡𝑑𝐰01 (𝐯) = 𝛿𝐰
01 01
(𝐯, 𝐯) ≡ 𝛿𝐰 01
(𝐯 ↘ 𝐰, 𝐯) = 𝛿𝐰 (𝐯 ↗ 𝐰, 𝟎). (33)

Notice that both 𝑚𝑒𝑎𝑛01

and
𝐰 (𝐯) 𝑠𝑡𝑑𝐰01 (𝐯)
always exist, which is guaran- All of the abovementioned considerations allow us to formulate the
teed by the existence of 𝐯 ↘ 𝐰 and 𝐯 ↗ 𝐰 and by the fact that 𝑠 ≠ 0. IA-WMSD property.
Moreover, 𝑚𝑒𝑎𝑛01 01 01
𝐰 (𝟎) = 0 and 𝑚𝑒𝑎𝑛𝐰 (𝐰) = 𝑚𝑒𝑎𝑛(𝐰), whereas 𝑠𝑡𝑑𝐰 (𝟎) = 0
01
and 𝑠𝑡𝑑𝐰 (𝐰) = 0. Theorem 2 (IA-WMSD Property). For every 𝐰 defining 𝐯 ∈ 𝑉 𝑆:
Now, given 𝐯 ∈ 𝑉 𝑆, let us observe how the diagonal 𝐷𝟎𝐰 relates √
01
𝑚𝑒𝑎𝑛01 01 01 𝛿𝐰 (𝐯, 𝟎) = 𝑚𝑒𝑎𝑛01 2 01 2
𝐰 (𝐯) + 𝑠𝑡𝑑𝐰 (𝐯) , (34)
𝐰 (𝐯) and 𝑠𝑡𝑑𝐰 (𝐯): 𝑚𝑒𝑎𝑛𝐰 (𝐯) specifies how far away 𝐯 is from 𝟎
when measured along 𝐷𝟎𝐰 , while 𝑠𝑡𝑑𝐰01 (𝐯) specifies how far away 𝐯 is √
( )2
01
from 𝐷𝟎𝐰 when measured along a direction that is perpendicular to 𝛿𝐰 (𝐯, 𝐰) = 𝑚𝑒𝑎𝑛(𝐰) − 𝑚𝑒𝑎𝑛01
𝐰 (𝐯) + 𝑠𝑡𝑑𝐰01 (𝐯)2 . (35)
it (Fig. 4A). Thus, the diagonal can serve as a ‘tool’ that connects
ideal/anti-ideal distance calculation with the weight-scaled means and Notice that 𝑚𝑒𝑎𝑛(𝐰) in the above may also be expressed as:
standard deviations of the alternatives. More formally, let 𝐯 be defined ‖𝐰‖2 ‖𝐰‖ ‖𝐰‖2
𝑚𝑒𝑎𝑛(𝐰) = ⋅ 𝑚𝑒𝑎𝑛(𝐰) = ‖𝐰‖ 2 = , (36)
as 𝐯 ↘ 𝐰. As visually explained in Fig. 4A, in this case: ‖𝐰‖2 2 𝑠
𝑚𝑒𝑎𝑛(𝐰)
01
𝛿𝐰 (𝐯, 𝟎) = 𝑚𝑒𝑎𝑛01
𝐰 (𝐯), (27) which emphasizes the divisor 𝑠, common to the core definitions of this
01
𝛿𝐰 (𝐯, 𝐰) = 𝑚𝑒𝑎𝑛(𝐰) − 𝑚𝑒𝑎𝑛01
𝐰 (𝐯), (28) paper. It should be also additionally stressed that for 𝐰 = 𝟏 the IA-
01
WMSD property becomes the IA-MSD property. In the next section,
𝛿𝐰 (𝐯, 𝐯) = 𝑠𝑡𝑑𝐰01 (𝐯). (29) we will discuss how the IA-WMSD property can be used to create
𝑛-independent visualizations of weight-based TOPSIS aggregations.
√
Additionally, because 𝐯 = 𝐰◦𝐮 and for 𝐰 = 𝟏 we get 𝑠 = 𝑛, it may
be shown that: 3.4. The WMSD-space
𝐮⋅𝟏𝑇
⋅ ‖𝟏‖
01
‖𝐯 ↘ 𝟏‖
‖
‖ ‖ ‖ ‖
‖2 ‖(𝟏◦𝐮) ↘ 𝟏‖2 ‖𝐮 ↘ 𝟏‖2 ‖𝟏‖22 ‖ ‖2 Analogously to the case of unweighted criteria and the resulting
𝑚𝑒𝑎𝑛𝟏 (𝐯) = √ = √ = √ = √ = MSD-space [35], the relation between the re-scaled weighted distances
𝑛 𝑛 𝑛 𝑛
of an alternative and the predefined reference points allows us to
𝐮⋅𝟏𝑇 𝐮⋅𝟏𝑇
√ propose a new space called WMSD-space that uses 𝑚𝑒𝑎𝑛01
‖𝟏‖ 𝑛 𝐮 ⋅ 𝟏𝑇 𝑠𝑢𝑚(𝐮) 𝐰 (𝐯) (WM) and
= √2 = √ = = = 𝑚𝑒𝑎𝑛(𝐮), (30) 𝑠𝑡𝑑𝐰01 (𝐯) (WSD) as its components.
𝑛 𝑛 𝑛 𝑛
‖𝐯 ↗ 𝟏‖ ‖(𝟏◦𝐮) ↗ 𝟏‖ ‖ ‖
‖ ‖ ‖ ‖2 ‖𝐮 ↗ 𝟏‖2 Definition 2 (WMSD-space).
𝑠𝑡𝑑𝟏01 (𝐯) = √ 2 = √ = √ =
𝑛 𝑛 𝑛
‖ ‖ ‖ ‖ WMSD-space = {[𝑚𝑒𝑎𝑛01 01
𝐰 (𝐯), 𝑠𝑡𝑑𝐰 (𝐯)] ∣ 𝐯 ∈ 𝑉 𝑆} (37)
‖ 𝐮⋅𝟏𝑇 ‖ ‖ 𝑠𝑢𝑚(𝐮) ‖
‖𝐮 − 2 ⋅ 𝟏‖ ‖𝐮 − √ 2 ⋅ 𝟏‖
‖𝐮 − 𝐮 ↘ 𝟏‖ ‖ ‖𝟏‖2 ‖ ‖ ‖ The WMSD-space can be represented in 2D space wherein the
‖ ‖2 ‖ ‖2 ‖ 𝑛 ‖2
= √ = √ = √ = weight-scaled mean (WM) of the alternatives is presented on the 𝑥-
𝑛 𝑛 𝑛 axis and the weight-scaled standard deviation (WSD) of the alternatives
‖ 𝑠𝑢𝑚(𝐮) ‖
‖𝐮 − 𝑛 ⋅ 𝟏‖ ‖ ‖ on the 𝑦-axis. As depicted in Fig. 4, WMSD-space may be treated as
‖ ‖2 ‖𝐮 − 𝑚𝑒𝑎𝑛(𝐮) ⋅ 𝟏‖2
= √ = √ an image of 𝑉 𝑆 under a two-dimensional transformation by functions
𝑛 𝑛 𝑚𝑒𝑎𝑛01 01
√ 𝐰 (𝐯) and 𝑠𝑡𝑑𝐰 (𝐯). It is worth underlining that the IA-WMSD prop-
(𝐮 − 𝑚𝑒𝑎𝑛(𝐮) ⋅ 𝟏)(𝐮 − 𝑚𝑒𝑎𝑛(𝐮) ⋅ 𝟏)𝑇 erty holds in WMSD-space, where it follows the Pythagorean theorem
= √ =
for two right triangles (pink dotted and green striped triangles in
𝑛
√ Fig. 4B). This is a result of WMSD-space being a ‘rotational’ projection
(𝐮 − 𝑚𝑒𝑎𝑛(𝐮) ⋅ 𝟏)(𝐮 − 𝑚𝑒𝑎𝑛(𝐮) ⋅ 𝟏)𝑇 √
= = 𝑣𝑎𝑟(𝐮) = 𝑠𝑡𝑑(𝐮), of 𝑉 𝑆 into two dimensions that retain the IA-WMSD property.
𝑛 Since WMSD-space is actually based on the weighted utility space,
(31) which is bounded, the extreme values of WM and WSD are also
which means that 𝑚𝑒𝑎𝑛01 01 bounded. In other words, for a given set of criteria weights, there is
𝐰 (𝐯) and 𝑠𝑡𝑑𝐰 (𝐯) constitute natural generaliza-
tions of 𝑚𝑒𝑎𝑛(𝐮) and 𝑠𝑡𝑑(𝐮). only a limited range of attainable WM and WSD values.2 As a result, one
As can be noticed in Fig. 4A, 𝑚𝑒𝑎𝑛01𝐰 (𝐯) satisfies:

𝑚𝑒𝑎𝑛01 01 01
𝐰 (𝐯) = 𝛿𝐰 (𝐯, 𝟎) ≡ 𝛿𝐰 (𝐯 ↘ 𝐰, 𝟎). (32) 2
Additional remarks on the boundary are presented in Appendix B.

7
R. Susmaga et al. Applied Soft Computing 153 (2024) 111279

Fig. 5. Visualizations of WMSD-space for the number of criteria (A) 𝑛 = 3, (B) 𝑛 = 4, (C) 𝑛 = 5, each for three different sets of weights depicted by different line types. Notice
that the dotted light gray line on each subplot corresponds to uniform weights and, therefore, the special case of WMSD-space, which is MSD-space. It is also worth noting how
the arithmetic mean of the weight (𝑚𝑒𝑎𝑛(𝐰)) corresponds to the maximal 𝑥-axis coordinate of WMSD-space.

can depict the boundary of WMSD-space, which depends on 𝐰, and thus of the anti-ideal point, as it remains the same, being equal to 𝟎 in
also on the number of criteria 𝑛𝑝 . Fig. 5 presents the shape of WMSD- 𝑈 𝑆 and to 𝐰◦𝟎 = 𝟎 in 𝑉 𝑆. As a result, TOPSIS utilizes 𝐰 instead of
space for different numbers of criteria 𝑛 and weight vectors 𝐰. Owing 𝟏 when computing the ‘distance to the ideal’. This means that new
to the symmetry of 𝑉 𝑆 and its ‘rotational’ projections, the boundary versions of the aggregation functions denoted by 𝖨𝐰 , 𝖠𝐰 and 𝖱𝐰 , must
of the WMSD-space does not depend on the order of the weights of 𝐰, be introduced. When expressed in terms of 𝛿𝐰 01 (𝐯, 𝐰) (‘distance to the
which means that the shape of the WMSD-space remains the same for 01 (𝐯, 𝟎) (‘distance to the anti-ideal in 𝑉 𝑆’), they are
ideal in 𝑉 𝑆’) and 𝛿𝐰
every permutation of a given set of weights. defined as follows:
Looking at Fig. 5 it can be noticed that setting 𝐰 = 𝟏 makes WMSD-
01 (𝐯, 𝐰)
𝛿𝐰
space equivalent to MSD-space. Indeed, the weight-scaling factor 𝑠 = 𝖨𝐰 (𝐯) = 1 − , (38)
‖𝐰‖2 𝑚𝑒𝑎𝑛(𝐰)
𝑚𝑒𝑎𝑛(𝐰)
has been chosen to replicate the relation between 𝑉 𝑆 and 𝑈 𝑆
in the relation between WMSD-space and MSD-space. Let us recall that 𝛿 01 (𝐯, 𝟎)
𝖠𝐰 (𝐯) = 𝐰 , (39)
MSD-space is based on a re-scaled distance measure 𝛿201 , which simply 𝑚𝑒𝑎𝑛(𝐰)
√
divides the Euclidean distance 𝛿2 by 𝑛. By doing so, 𝛿201 is independent 01 (𝐯, 𝟎)
𝛿𝐰
of the number of criteria 𝑛, √ making the maximum distance in MSD- 𝖱𝐰 (𝐯) = . (40)
01
𝛿𝐰 (𝐯, 𝐰) + 𝛿𝐰 01 (𝐯, 𝟎)
space always 1. Notice that 𝑛 is a special case of ‖𝐰‖2 for 𝐰 = 𝟏,
which emphasizes the fact that 𝛿𝐰 01 is a natural generalization of 𝛿 01 for where 𝐮 ∈ 𝑈 𝑆 and 𝐯 = 𝐰◦𝐮 ∈ 𝑉 𝑆 are the images of an alternative
2
the case of 𝐰 ≠ 𝟏. Although making 𝑠 = ‖𝐰‖2 would suffice to ensure from 𝑈 𝑆 and 𝑉 𝑆, respectively. All three aggregations generate values
the IA-WMSD property, additionally scaling by 𝑚𝑒𝑎𝑛(𝐰) makes the sizes belonging to [0, 1].
between WMSD-space and MSD-space follow the relation 𝑚𝑒𝑎𝑛(𝐰) < As was the case with 𝖨(𝐮) (see Section 2), aggregation 𝖨𝐰 (𝐯) fea-
𝑚𝑒𝑎𝑛(𝟏) = 1 for 𝐰 ≠ 𝟏. As a result, in WMSD-space the maximal value 𝐰 𝛿 01 (𝐯,𝐰) 𝐰 𝛿 01 (𝐯,𝐰)
tures a reversal (1 − 𝑚𝑒𝑎𝑛(𝐰) instead of 𝑚𝑒𝑎𝑛(𝐰) ). This modification was
of WM is 𝑚𝑒𝑎𝑛(𝐰) instead of 1. introduced to ensure that all aggregations are interpreted as functions
It is also worth noticing that the number of non-zero criteria and that need to be maximized. The distances from the ideal and anti-ideal
the particular values of their weights (i.e. the size and the values points in 𝖨𝐰 (𝐯) and 𝖠𝐰 (𝐯) have been divided by 𝑚𝑒𝑎𝑛(𝐰) to keep the
of 𝐰) affect the number of vertices of the WMSD-space boundary
values of these aggregations in [0, 1].
(Fig. 5). Finally, as was the case for MSD-space, WMSD-space can
Obviously, using the IA-WMSD property it is possible to express all
always be depicted in two dimensions. We will use this property to
visualize alternatives and values of TOPSIS aggregation functions in the aggregations in terms of 𝑚𝑒𝑎𝑛01 01 01
𝐰 (𝐯) and 𝑠𝑡𝑑𝐰 (𝐯) instead of 𝛿𝐰 (𝐯, 𝐰)
01
and 𝛿𝐰 (𝐯, 𝟎):
WMSD-space.
√( ) 2
3.5. TOPSIS aggregations in WMSD-space 𝑚𝑒𝑎𝑛(𝐰) − 𝑚𝑒𝑎𝑛01
𝐰 (𝐯) + 𝑠𝑡𝑑𝐰01 (𝐯)2
𝖨𝐰 (𝐯) = 1 − , (41)
𝑚𝑒𝑎𝑛(𝐰)
The application of the weights, while changing 𝑈 𝑆 into 𝑉 𝑆, does √
necessarily influence the image of the ideal point, as the image moves 𝑚𝑒𝑎𝑛01 2 01
𝐰 (𝐯) + 𝑠𝑡𝑑𝐰 (𝐯)
2
𝖠𝐰 (𝐯) = , (42)
from 𝟏 in 𝑈 𝑆 to 𝐰 in 𝑉 𝑆. The same does not concern the image 𝑚𝑒𝑎𝑛(𝐰)

8
R. Susmaga et al. Applied Soft Computing 153 (2024) 111279

Fig. 6. An illustrative point 𝐯 = [0.75, 0.25] depicted in WMSD-space defined by 𝐰 = [1.0, 0.5] for aggregations (A) 𝖨𝐰 (𝐯), (B) 𝖠𝐰 (𝐯) and (C) 𝖱𝐰 (𝐯). Color encodes the aggregation
value, with blue representing the least preferred and red the most preferred values.

Table 2
The relation between the mean and standard deviation as the property of the analyzed aggregation functions.
On the left, properties of aggregation functions in MSD-space, as presented in [35]. On the right, the
properties of aggregation functions in WMSD-space. Notably, the configuration of gain/cost table entries is
the same on the left and on the right.
MSD-space WMSD-space
Aggregation
𝑚𝑒𝑎𝑛(𝐮) 𝑠𝑡𝑑(𝐮) 𝑚𝑒𝑎𝑛01
𝐰
(𝐯) 𝑠𝑡𝑑𝐰01 (𝐯)
𝖨 gain cost gain cost
𝖠 gain gain gain gain
𝑚𝑒𝑎𝑛(𝐰)
𝑚𝑒𝑎𝑛(𝐮) < 0.5: gain 𝑚𝑒𝑎𝑛01
𝐰
(𝐯) < 2
: gain
𝑚𝑒𝑎𝑛(𝐰)
𝖱 gain 𝑚𝑒𝑎𝑛(𝐮) = 0.5: neutrality gain 𝑚𝑒𝑎𝑛01
𝐰
(𝐯) = 2
: neutrality
𝑚𝑒𝑎𝑛(𝐰)
𝑚𝑒𝑎𝑛(𝐮) > 0.5: cost 𝑚𝑒𝑎𝑛01
𝐰
(𝐯) > 2
: cost

of countries described in terms of factors constituting the Index of

Economic Freedom. The goal of the case studies is to visualize the
𝖱𝐰 (𝐯) =
√ alternatives within the MSD-space and WMSD-space, present the impact
𝑚𝑒𝑎𝑛01 2 01 2 of introducing weights, and discuss how the two spaces depict the
𝐰 (𝐯) + 𝑠𝑡𝑑𝐰 (𝐯)
√ √( )2
. (43) relations between each alternative’s properties and their aggregation
𝑚𝑒𝑎𝑛01 2 01 2 𝑚𝑒𝑎𝑛(𝐰) − 𝑚𝑒𝑎𝑛01 + 𝑠𝑡𝑑𝐰01 (𝐯)2 values. The datasets and code for reproducing the results presented in
𝐰 (𝐯) + 𝑠𝑡𝑑𝐰 (𝐯) + 𝐰 (𝐯)
this section are available at GitHub: https://github.com/dabrze/topsis-
Notice that each aggregation is a function of a single vector argu- msd-improvement-actions.
ment 𝐯 ∈ 𝑉 𝑆, and thus, in practice, a function of the 𝑛 real-valued In the next subsections, the following notation shall be used to
elements of this vector. However, the same IA-WMSD property makes present alternative rankings: 𝐗𝑖 ≻𝑎𝑔𝑔 𝐗𝑗 : alternative 𝐗𝑖 is preferred
it possible to render these aggregations in a plane as a function of just over 𝐗𝑗 under every aggregation from 𝑎𝑔𝑔; 𝐗𝑖 ∼𝑎𝑔𝑔 𝐗𝑗 : 𝐗𝑖 and 𝐗𝑗 are
two real-valued arguments: 𝑚𝑒𝑎𝑛01 01
𝐰 (𝐯) and 𝑠𝑡𝑑𝐰 (𝐯). This demonstrates indifferent under 𝑎𝑔𝑔; 𝐗𝑖 ≺𝑎𝑔𝑔 𝐗𝑗 : 𝐗𝑗 is preferred over 𝐗𝑖 under 𝑎𝑔𝑔.
the introduced underlying dimensionality reduction. For step-by-step Moreover, for an alternative 𝐗𝑖 we will use 𝑚𝑒𝑎𝑛(𝐗𝑖 ) as a shorthand
examples of how the equations presented in the above sections can be for 𝑚𝑒𝑎𝑛(𝑈 𝑆(𝐗𝑖 )) and 𝑚𝑒𝑎𝑛01 01
𝐰 (𝐗𝑖 ) as a shorthand for 𝑚𝑒𝑎𝑛𝐰 (𝑉 𝑆(𝐗𝑖 )).
used to transform an alternative from the criteria space to WMSD-space 01
Analogously, 𝑠𝑡𝑑(𝐗𝑖 ) and 𝑠𝑡𝑑𝐰 (𝐗𝑖 ) will also denote re-scaled standard
see Appendix A. deviations calculated for images of the alternative in 𝑈 𝑆 and 𝑉 𝑆
Now, when visualizing WMSD-space it is helpful to show alterna- respectively.
tives against the values of TOPSIS aggregations (𝖨𝐰 (𝐯), 𝖠𝐰 (𝐯) and 𝖱𝐰 (𝐯)),
which can be color-coded as was done for MSD-space [35]. 4.1. Student grades
As seen in Fig. 6, coloring WMSD-space in a way that represents the
values of the aggregation functions reveals the interplay of 𝑚𝑒𝑎𝑛01
𝐰 (𝐯) and The first dataset contains 15 alternatives, i.e., students described by
𝑠𝑡𝑑𝐰01 (𝐯). The WM-WSD interplay for different aggregation functions in three criteria which are the average grades obtained by these students
WMSD-space resembles that described for MSD-space. This is to be in Maths, Biology, and Art. The domains of the criteria are [0, 100]
expected as MSD-space constitutes a special case of WMSD-space where for Maths, [1, 6] for Biology, and [1, 6] for Art. The alternatives are
𝐰 = 𝟏. Table 2, shows which aggregation functions act like type cost or presented in Table 3. The rankings of the alternatives are considered
type gain criteria depending on 𝑚𝑒𝑎𝑛01 01
𝐰 (𝐯) and 𝑠𝑡𝑑𝐰 (𝐯), and compares in two scenarios: when all criteria are of equal importance (i.e., 𝐰 =
WMSD-space with MSD-space. [1.0, 1.0, 1.0]) and when 𝐰 = [0.5, 0.6, 1.0]. The description of alternatives
As can be seen in Table 2, MSD-space and WMSD-space share in terms of 𝑈 𝑆 (equal weights), 𝑉 𝑆 (unequal weights), MSD-space,
the same configuration of gain/cost characteristics of the means and WMSD-space and the three aggregations are presented in Table 4. The
standard deviations across the spaces. It is also worth noticing that alternatives have been chosen to cover different areas of the MSD-
the mean weight 𝑚𝑒𝑎𝑛(𝐰) plays a role in the properties of aggregation space and WMSD-space (see Fig. 7) and to represent some of their
𝖱𝐰 (𝐯). Finally, it can be observed that WMSD-space is a generalization characteristic points, e.g. the worst possible alternative (𝐒10 ) or best
of MSD-space to weighted criteria (𝐯 = 𝐰◦𝐮). For the special case of 𝐰 = possible alternative (𝐒12 ).
𝟏, WMSD-space is equivalent to MSD-space (in particular, 𝑚𝑒𝑎𝑛(𝟏)2
= 0.5). The visualizations in Fig. 7 confront the shapes of MSD-space and
In the following section, we will discuss how these properties apply to WMSD-space. The latter can be regarded as a natural generalization
practical ranking problems. of the MSD-space that incorporates different weights assigned to the
criteria. For 𝐰 ≠ 𝟏, the WMSD-space is characterized by a potentially
4. Case studies larger number of vertices and a possibly smaller range of mean values
than the MSD-space. Naturally, preference information given by the de-
In this section, we present two case studies conducted on a dataset cision maker in the form of weights influences not only the shape of the
of students described in terms of school grades and on a dataset space but also the position of the alternatives within those spaces. For

9
R. Susmaga et al. Applied Soft Computing 153 (2024) 111279

Table 3 how they influence the final rankings under different aggregations.
Descriptions of alternatives for the first case study; students described
The interplay of 𝑚𝑒𝑎𝑛01 01
𝐰 (𝐯) and 𝑠𝑡𝑑𝐰 (𝐯) in the context of preferences,
by their average grades from Maths [0 − 100], Biology [1 − 6], and Art
[1 − 6].
as formalized in Table 2, is illustrated in the WMSD-space by the
Math Bio Art
color reflecting the aggregation value imposed for each point of the
space. Getting a higher ranking position requires an increase in the
𝐒1 29.11 2.46 2.46
𝐒2 49.37 3.53 3.47
aggregation value, which is reflected by a change of the alternative’s
𝐒3 70.89 4.54 4.54 color towards dark red. This can naturally be achieved when the
𝐒4 40.51 3.53 1.89 alternative obtains more desirable values on the criteria. In our example
𝐒5 35.44 4.80 3.22 from Table 3, this means that a student should get better marks in some
𝐒6 59.49 3.47 5.11
subjects while not worsening them in any other subject. This would
𝐒7 44.30 4.80 1.38
𝐒8 93.67 5.05 2.39 result in the increase of 𝑚𝑒𝑎𝑛01 𝐰 (𝐯), which can, however, be hard to
𝐒9 55.70 2.20 5.62 achieve, or in some cases even impossible. However, the preference
𝐒10 0.00 1.00 1.00 related interplay in the WMSD-space between 𝑚𝑒𝑎𝑛01 01
𝐰 (𝐯) and 𝑠𝑡𝑑𝐰 (𝐯)
𝐒11 0.00 5.56 1.13 shows other ways to influence the ranking even without increasing
𝐒12 100.00 6.00 6.00
𝑚𝑒𝑎𝑛01 𝐰 (𝐯).
𝐒13 100.00 1.44 5.87
𝐒14 70.71 4.84 3.22 First, let us focus on three alternatives characterized by the same
𝐒15 89.90 3.32 4.79 𝑚𝑒𝑎𝑛01 𝐰 (𝐯) = 0.5: 𝐒3 , 𝐒6 and 𝐒9 (analogous discussion is valid for, e.g., 𝐒1 ,
𝐒4 and 𝐒7 ). The ranking under the 𝖨𝗐 (𝐯) aggregation is the following:
𝐒3 ≻𝖨𝗐 (𝐯) 𝐒6 ≻𝖨𝗐 (𝐯) 𝐒9 , as opposed to 𝖠𝗐 (𝐯), where 𝐒9 ≻𝖠𝗐 (𝐯) 𝐒6 ≻𝖠𝗐 (𝐯) 𝐒3 .
Clearly, a change in the 𝑠𝑡𝑑𝐰01 (𝐯), with no change of 𝑚𝑒𝑎𝑛01 𝐰 (𝐯), is enough
example, the positions of 𝐒8 and 𝐒9 change quite drastically, as the two
to influence the rankings. Under aggregation 𝖨𝗐 (𝐯) less variant values of
alternatives almost swap their positions: they are both characterized by
the criteria are preferred, as 𝑠𝑡𝑑𝐰01 (𝐯) is of cost-type for this aggregation.
a very similar 𝑠𝑡𝑑(𝐮) and 𝑠𝑡𝑑𝐰01 (𝐯), but 𝑚𝑒𝑎𝑛(𝐒8 ) = 0.68 > 𝑚𝑒𝑎𝑛(𝐒9 ) = 0.57
In contrast, under the 𝖠𝗐 (𝐯) aggregation an increase of 𝑠𝑡𝑑𝐰01 (𝐯) results
whereas 𝑚𝑒𝑎𝑛01 01
𝐰 (𝐒8 ) = 0.35 < 𝑚𝑒𝑎𝑛𝐰 (𝐒9 ) = 0.50. As expected, the images in the increase of the aggregation function. Aggregation 𝖱𝗐 (𝐯) on the
of the best possible alternative (𝟏) and the worst possible one (𝟎) have 𝑚𝑒𝑎𝑛(𝐰)
other hand, resembles aggregation 𝖠𝗐 (𝐯) when 𝑚𝑒𝑎𝑛01 𝐰 (𝐯) < 2
fixed relative positions no matter what the values of weights are, in 𝑚𝑒𝑎𝑛(𝐰)
and aggregation 𝖨𝗐 (𝐯) when 𝑚𝑒𝑎𝑛01 𝐰 (𝐯) > . In the very middle
the sense that they are always situated, respectively, in the rightmost 𝑚𝑒𝑎𝑛(𝐰)
2

and leftmost vertices of MSD-space and WMSD-space (see 𝐒12 and 𝐒10 of WMSD-space, i.e. when 𝑚𝑒𝑎𝑛01 𝐰 (𝐯) = 2
, the change of 𝑠𝑡𝑑𝐰01 (𝐯)
has no effect on the ranking at all. This brings us to the conclusion
in Fig. 7).
that the isolines in the WMSD-space can guide the decision maker as to
The change of alternative position imposed by the incorporation of
what actions need to be taken in order to influence the ranking without
weights can also be observed for alternatives that are characterized by
changing the 𝑚𝑒𝑎𝑛01 𝐰 (𝐯) of the alternative.
different vectors in 𝑈 𝑆, but by a single point in MSD-space. To illustrate
Now, let us focus on another set of alternatives: 𝐒4 , 𝐒5 and 𝐒6
this case, let us consider 𝐒6 and 𝐒14 . As shown in Table 4 and Fig. 7
(analogous discussion is valid for, e.g., 𝐒10 , 𝐒1 , 𝐒2 , 𝐒3 and 𝐒12 ). They
vectors [0.59, 0.49, 0.82] (the 𝑈 𝑆 image of 𝐒6 ) and [0.71, 0.77, 0.44] (the
are characterized by the same 𝑠𝑡𝑑𝐰01 (𝐯), but different 𝑚𝑒𝑎𝑛01 𝐰 (𝐯). Under
𝑈 𝑆 image of 𝐒14 ) are characterized by 𝑚𝑒𝑎𝑛(𝐒𝟔 ) = 𝑚𝑒𝑎𝑛(𝐒𝟏𝟒 ) = 0.64
all the three considered aggregations, the alternatives are ranked the
and 𝑠𝑡𝑑(𝐒𝟔 ) = 𝑠𝑡𝑑(𝐒𝟏𝟒 ) = 0.14. As a result, those vectors share the
same: 𝐒6 ≻𝖨𝗐 (𝐯),𝖠𝗐 (𝐯),𝖱𝗐 (𝐯) 𝐒5 ≻𝖨𝗐 (𝐯),𝖠𝗐 (𝐯),𝖱𝗐 (𝐯) 𝐒4 . This results from the
same point in MSD-space and are thus identically evaluated by the
fact that 𝑚𝑒𝑎𝑛01 𝐰 (𝐯) is of type gain for all the considered aggregations.
aggregations 𝖨(𝐮), 𝖠(𝐮) and 𝖱(𝐮). However, the two alternatives do not
Thus, moving to the right in the WMSD-space (i.e., keeping the same
share the same point in WMSD-space, as their positions in WMSD-
𝑠𝑡𝑑𝐰01 (𝐯) and only increasing 𝑚𝑒𝑎𝑛01 𝐰 (𝐯)) always increases the aggregation
space are influenced by the weights. The weights affect the vectors in
functions.
𝑈 𝑆, which are different for 𝐒6 and 𝐒14 . As a result, the 𝑚𝑒𝑎𝑛01 𝐰 (𝐒6 ) = Last but not least, the alternative’s rating can be influenced by
0.50 ≠ 𝑚𝑒𝑎𝑛01 𝐰 (𝐒14 ) = 0.39, which implies that the two alternatives are a simultaneous change in 𝑚𝑒𝑎𝑛01 01
𝐰 (𝐯) and 𝑠𝑡𝑑𝐰 (𝐯). A rather hard to
not identically evaluated by the considered aggregations (even though
predict compensation of those two is clearly visible in the WMSD-space,
𝑠𝑡𝑑𝐰01 (𝐒6 ) = 𝑠𝑡𝑑𝐰01 (𝐒14 ) = 0.10). In particular, 𝐒6 ≻𝖨𝗐 (𝐯),𝖠𝗐 (𝐯),𝖱𝗐 (𝐯) 𝐒14 as the change in rating is equivalent to ‘switching between’ isolines.
(see Table 4). Similarly, alternatives that are characterized by the Let us consider 𝐒9 and 𝐒13 (analogous discussion is valid for 𝐒7 and
same point in WMSD-space do not share, in general, the same point 𝐒11 ). Observe that the rankings of those two alternatives are different
in MSD-space. In particular, 𝐒6 and 𝐒15 are characterized by different under different aggregations: 𝐒9 ≻𝖨𝗐 (𝐯) 𝐒13 , but 𝐒13 ≻𝖱𝗐 (𝐯) 𝐒9 . This
vectors in 𝑈 𝑆, namely [0.59, 0.49, 0.82] for 𝐒6 and [0.90, 0.46, 0.76] for change results from the fact that the isolines for the 𝖱𝗐 (𝐯) aggregation
𝐒15 , but are characterized by 𝑚𝑒𝑎𝑛01 01
𝐰 (𝐒6 ) = 𝑚𝑒𝑎𝑛𝐰 (𝐒15 ) = 0.50 and ‘straighten up’ while moving towards the middle values of 𝑚𝑒𝑎𝑛01 𝐰 (𝐯),
𝑠𝑡𝑑𝐰01 (𝐒6 ) = 𝑠𝑡𝑑𝐰01 (𝐒15 ) = 0.10, which puts them in the very same position while the isolines for the 𝖨𝗐 (𝐯) aggregation keep the same concentric
in WMSD-space. This however, does not imply sharing the same point character. The isolines are thus a visual representation of the trade-offs
in MSD-space, since this position is influenced by the vector of weights. between 𝑚𝑒𝑎𝑛01 01
𝐰 (𝐯) and 𝑠𝑡𝑑𝐰 (𝐯) for different aggregations.
Consequently, 𝐒6 and 𝐒15 are evaluated identically by 𝖨𝗐 (𝐯), 𝖠𝗐 (𝐯) and
𝖱𝗐 (𝐯), but differently by 𝖨(𝐮), 𝖠(𝐮) and 𝖱(𝐮) (Fig. 7). 4.2. Index of economic freedom
Although the MSD-space and WMSD-space share the same character
of the isolines under particular aggregation, the change of the alterna- The second case study is based on publicly available data from
tives’ position across those spaces caused by the weights influences the the Index of Economic Freedom [50], which covers 12 freedoms—
final ratings of the alternatives. Let us look again at 𝐒8 and 𝐒9 . It is clear from property rights to tax burdens—in 184 countries. The data has
that because of weights, their rankings are reversed: 𝐒8 ≻𝖨(𝐮),𝖠(𝐮),𝖱(𝐮) 𝐒9 been annually collected for almost 30 years now by The Heritage
versus 𝐒9 ≻𝖨𝗐 (𝐯),𝖠𝗐 (𝐯),𝖱𝗐 (𝐯) 𝐒8 (Table 4). Alternative 𝐒8 is much better Foundation [51] and served as the basis for many case studies and
than 𝐒9 on the first two criteria, but their importance was diminished analyses [52–56]. In particular, our case study is based on the data
by the weights being 0.5 and 0.6, causing 𝐒9 to climb higher in those gathered for the 25th anniversary of the Index in 2019, which were
rankings that take weights into account. used by de Lima Silva and de Almeida Filho [55].
The above analysis compared the MSD-space and WMSD-space from Economic freedom is understood as the right of every human to
the viewpoint of criteria weights. Let us now illustrate some trade-offs control their own labor and property. Within the Index, 12 factors are
between the values of 𝑚𝑒𝑎𝑛01 01
𝐰 (𝐯) and 𝑠𝑡𝑑𝐰 (𝐯) in WMSD-space, and show measured and grouped into four categories: Rule of Law, Government

10
R. Susmaga et al. Applied Soft Computing 153 (2024) 111279

Table 4
Descriptions of students in terms of 𝑈 𝑆, 𝑉 𝑆 given 𝐰 = [0.5, 0.6, 1.0], MSD-space, WMSD-space and the three aggregations. The ranking positions of alternatives for particular
aggregations are presented as superscripts (𝚙𝚘𝚜) .
𝑈𝑆 𝑉𝑆 MSD WMSD Aggregations
𝑈1 𝑈2 𝑈3 𝑉1 𝑉2 𝑉3 M SD WM WSD 𝖨(𝐮) 𝖠(𝐮) 𝖱(𝐮) 𝖨𝐰 (𝐯) 𝖠𝐰 (𝐯) 𝖱𝐰 (𝐯)
𝐒1 0.29 0.29 0.29 0.15 0.17 0.29 0.29 0.00 0.20 0.00 0.29(𝟷𝟹) 0.29(𝟷𝟺) 0.29(𝟷𝟺) 0.29(𝟷𝟷) 0.29(𝟷𝟺) 0.29(𝟷𝟺)
𝐒2 0.49 0.51 0.49 0.25 0.30 0.49 0.50 0.01 0.35 0.00 0.50(𝟽) 0.50(𝟷𝟸) 0.50(𝟷𝟶) 0.50(𝟾) 0.50(𝟷𝟶) 0.50(𝟾)
𝐒3 0.71 0.71 0.71 0.35 0.43 0.71 0.71 0.00 0.50 0.00 0.71(𝟸) 0.71(𝟻) 0.71(𝟸) 0.71(𝟸) 0.71(𝟼) 0.71(𝟸)
𝐒4 0.41 0.51 0.18 0.20 0.30 0.18 0.36 0.14 0.20 0.10 0.35(𝟷𝟸) 0.39(𝟷𝟹) 0.37(𝟷𝟹) 0.27(𝟷𝟸) 0.32(𝟷𝟹) 0.31(𝟷𝟹)
𝐒5 0.35 0.76 0.44 0.18 0.46 0.44 0.52 0.17 0.35 0.10 0.49(𝟾) 0.55(𝟿) 0.52(𝟿) 0.48(𝟿) 0.52(𝟿) 0.50(𝟾)
𝐒6 0.59 0.49 0.82 0.30 0.30 0.82 0.64 0.14 0.50 0.10 0.61(𝟺) 0.65(𝟼) 0.63(𝟺) 0.68(𝟹) 0.73(𝟺) 0.69(𝟹)
𝐒7 0.44 0.76 0.08 0.22 0.46 0.08 0.43 0.28 0.20 0.20 0.36(𝟷𝟷) 0.51(𝟷𝟷) 0.44(𝟷𝟷) 0.23(𝟷𝟹) 0.40(𝟷𝟸) 0.34(𝟷𝟷)
𝐒8 0.94 0.81 0.28 0.47 0.49 0.28 0.68 0.29 0.35 0.20 0.57(𝟼) 0.73(𝟹) 0.63(𝟺) 0.42(𝟷𝟶) 0.58(𝟽) 0.50(𝟾)
𝐒9 0.56 0.24 0.92 0.28 0.14 0.92 0.57 0.28 0.50 0.20 0.49(𝟾) 0.64(𝟾) 0.56(𝟾) 0.60(𝟻) 0.77(𝟹) 0.66(𝟼)
𝐒10 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00(𝟷𝟻) 0.00(𝟷𝟻) 0.00(𝟷𝟻) 0.00(𝟷𝟻) 0.00(𝟷𝟻) 0.00(𝟷𝟻)
𝐒11 0.00 0.91 0.03 0.00 0.55 0.03 0.31 0.42 0.15 0.26 0.19(𝟷𝟺) 0.53(𝟷𝟶) 0.39(𝟷𝟸) 0.14(𝟷𝟺) 0.43(𝟷𝟷) 0.33(𝟷𝟸)
𝐒12 1.00 1.00 1.00 0.50 0.60 1.00 1.00 0.00 0.70 0.00 1.00(𝟷) 1.00(𝟷) 1.00(𝟷) 1.00(𝟷) 1.00(𝟷) 1.00(𝟷)
𝐒13 1.00 0.09 0.97 0.50 0.05 0.97 0.69 0.42 0.55 0.26 0.47(𝟷𝟶) 0.81(𝟸) 0.61(𝟽) 0.57(𝟼) 0.86(𝟸) 0.67(𝟻)
𝐒14 0.71 0.77 0.44 0.35 0.46 0.44 0.64 0.14 0.39 0.10 0.61(𝟺) 0.65(𝟼) 0.63(𝟺) 0.53(𝟽) 0.58(𝟽) 0.55(𝟽)
𝐒15 0.90 0.46 0.76 0.45 0.28 0.76 0.71 0.18 0.50 0.10 0.66(𝟹) 0.73(𝟹) 0.68(𝟹) 0.68(𝟹) 0.73(𝟺) 0.69(𝟹)

Fig. 7. Students depicted in MSD-space (left) and WMSD-space defined for 𝐰 = [0.5, 0.6, 1.0] (right) for three different aggregation functions: 𝖨(𝐮) (top), 𝖠(𝐮) (middle) and 𝖱(𝐮)
(bottom). Color encodes the aggregation value, with blue representing the least preferred and red the most preferred values. Note that the WMSD sub-figures on the right are on
a different scale than those on the left; we have scaled up the WMSD sub-figures to increase their readability.

Size, Regulatory Efficiency, Open Markets. There are three factors per are available in [57]. For the purpose of this case study, we have limited
category, each factor is graded on a 0–100 scale of type gain. Details on the Index only to the 12 countries of South America and aggregated the
how the values of factors are determined for the considered countries criteria by taking the mean of factors forming each category (Table 5).

11
R. Susmaga et al. Applied Soft Computing 153 (2024) 111279

Fig. 8. The visualization of preference as expressed by 𝖱𝗐 (𝐯) aggregation in WMSD-spaces defined for four different weight vectors. The color map reflects the preference: dark
blue—the least preferred, dark red—the most preferred. Note that the sub-figures are on different scales; we have scaled up the WMSD sub-figures to increase their readability.

Table 5 permutations of each other to illustrate that the order of weights does
Descriptions of alternatives for the second case study; South American countries
not affect the shape of WMSD-space. Finally, the last weight vector
described by four criteria of type gain: Rule of Law, Government Size, Regulatory
Efficiency, Open Markets. Each criterion is formed as a mean of three factors obtained
completely eliminates the influence of the Open Markets criterion on
from the Index of Economic Freedom [50]. the country rankings by setting its weight to zero; this makes 𝑛𝑝 = 3 <
ID Country Rule of Law Gov. Size Reg. Eff. Open Markets 𝑛 = 4.
ARG Argentina 41.93 50.60 54.50 61.67 Fig. 8 shows the countries of South America superimposed on a
BOL Bolivia 17.50 49.77 60.17 41.80 gradient depicting the preference expressed by 𝖱𝐰 (𝐯) in WMSD-space.
BRA Brazil 45.70 43.87 61.77 56.33 The final country rankings are gathered in Table 6.
CHL Chile 62.43 82.43 75.37 81.27
As can be noticed in Fig. 8, the weights have a clear influence both
COL Colombia 42.33 76.17 75.17 75.33
ECU Ecuador 27.13 54.87 58.60 47.13
on the shape of WMSD-spaces and the rankings of the alternatives. First,
GUY Guyana 39.27 71.33 66.07 50.60 observe that 𝑚𝑒𝑎𝑛(𝐰𝟐 ) = 𝑚𝑒𝑎𝑛(𝐰𝟑 ) = 𝑚𝑒𝑎𝑛(𝐰𝟒 ) = 0.5, resulting in the
PRY Paraguay 31.67 90.50 54.50 70.53 same WM range (x-axis) in Figs. 8B, C and D. Additionally, vector 𝐰𝟑
PER Peru 40.63 85.07 71.73 73.80 is simply a permutation of 𝐰𝟐 , thus the whole shape of the respective
SUR Suriname 35.60 52.57 59.27 44.87
WMSD-spaces is exactly the same (see Fig. 8B and C). Nonetheless,
URY Uruguay 65.47 71.53 73.03 64.53
VEN Venezuela 09.53 50.13 20.63 23.33 the positions of the alternatives within those shapes differ, leading to
different rankings. The weights also naturally influence the number of
vertices in the WMSD-spaces, which is, in particular, reflected by a
smaller number of vertices when some criteria are given weights equal
To ensure reproducibility, the raw data from the Heritage Foundation, to zero (compare Fig. 8B and D).
their transformations, and final rankings under different aggregations Looking at the rankings under different weight vectors (Table 6),
and weights are available in the online supplementary materials. one notices the inevitable changes in country ratings imposed by incor-
The case study focuses on the 𝖱𝐰 (𝐯) aggregation, since it is the one porating weights. For example, Uruguay shifts from the second position
most commonly used in practice. The following four sets of weights under 𝐰𝟏 or 𝐰𝟒 to as far as the fifth position under 𝐰𝟐 . Interestingly,
are considered as examples of preference information given by different for particular weight vectors, some countries are almost indiscernible as
decision-makers: their values of the 𝖱𝐰 (𝐯) aggregation differ only slightly, e.g., Argentina,
• 𝐰𝟏 = [1.00, 1.00, 1.00, 1.00], Ecuador, and Suriname under 𝐰𝟑 . The WMSD visualizations (Fig. 8)
• 𝐰𝟐 = [0.25, 1.00, 0.25, 0.50], explain in terms of means and standard deviations why the ranking
• 𝐰𝟑 = [0.50, 1.00, 0.25, 0.25], positions of those countries are (almost) the same. Observe that the
considered countries are located in the green region, i.e., very close to
• 𝐰𝟒 = [1.00, 0.66, 0.33, 0.00].
the middle of the WM range (x-axis), which happens to be the region
These sets of weights were not chosen with any domain-specific ra- where 𝑠𝑡𝑑𝐰01 (𝐯) hardly influences the rankings (Table 2). Thus, despite
tionale but to demonstrate different situations arising when working occupying different points in the WMSD-space, the countries are ranked
with weights. In particular, the first weight vector expresses equal almost equally.
importance of all criteria, allowing for the visualization of the alter- Since the Index of Economic Freedom is updated annually, it would
natives in MSD-space, as opposed to the other sets of weights, which also be interesting to visually compare the data from 2019 and 2023.
require WMSD-space. The second and third vectors were chosen to be This can be easily done using WMSD-space. In Fig. 9, we show such a

12
R. Susmaga et al. Applied Soft Computing 153 (2024) 111279

Table 6
Rankings of South American countries resulting from 𝖱𝐰 (𝐯) aggregation under different weight vectors.
𝐰𝟏 𝐰𝟐 𝐰𝟑 𝐰𝟒
Rank
[1.00, 1.00, 1.00, 1.00] [0.25, 1.00, 0.25, 0.50] [0.50, 1.00, 0.25, 0.25] [1.00, 0.66, 0.33, 0.00]
1 CHL 0.746 CHL 0.806 CHL 0.775 CHL 0.684
2 URY 0.685 PER 0.787 PER 0.725 URY 0.677
3 PER 0.659 PRY 0.785 PRY 0.713 PER 0.548
4 COL 0.658 COL 0.738 URY 0.701 COL 0.539
5 PRY 0.599 URY 0.700 COL 0.685 GUY 0.503
6 GUY 0.564 GUY 0.652 GUY 0.634 PRY 0.501
7 ARG 0.521 ARG 0.524 ARG 0.497 BRA 0.464
8 BRA 0.519 ECU 0.523 ECU 0.497 ARG 0.453
9 SUR 0.481 SUR 0.507 SUR 0.494 SUR 0.424
10 ECU 0.471 BOL 0.474 BRA 0.456 ECU 0.382
11 BOL 0.430 BRA 0.471 BOL 0.444 BOL 0.321
12 VEN 0.283 VEN 0.426 VEN 0.412 VEN 0.262

Fig. 9. The visualization in WMSD-space of the change of the countries’ positions between 2019 (solid circles) and 2023 (empty circles) under 𝖱𝗐 (𝐯) aggregation and weight vector
𝐰𝟑 = [0.50, 1.00, 0.25, 0.25].

comparison for four countries: Chile, Uruguay, Suriname, and Venezuela.

Observe that under a particular aggregation (𝖱𝐰 (𝐯)) and weight vector
(𝐰𝟑 = [0.50, 1.00, 0.25, 0.25]) the shape of the WMSD-space and the
isolines of the aggregation function are fixed. Thus, the comparison
of the data from various years only requires superimposing that data
on the WMSD-space. The countries’ positions in the year 2019 are
depicted by solid circles, whereas empty circles mark the countries’
positions in 2023. This dynamic perspective shows that Venezuela’s
and Uruguay’s positions improved while Suriname’s and Chile’s got
worse. As a result, in 2023, Venezuela outranks Suriname and Uruguay
outranks Chile, which was not the case in 2019. All of the considered
ranking transitions resulted from the changes on both 𝑚𝑒𝑎𝑛01 𝐰 (𝐯) and
𝑠𝑡𝑑𝐰01 (𝐯). A closer inspection reveals that out of the four considered
countries, Chile is characterized by the biggest decrease in 𝑚𝑒𝑎𝑛01 𝐰 (𝐯),
which directly caused Chile’s drop in the ranking. On the other hand, Fig. 10. Parallel coordinates plots comparing the countries’ criteria values in 2019
the description of Venezuela in terms of the analyzed criteria became (dashed line) and 2023 (solid line).
notably diversified, which was reflected by the biggest change in
𝑠𝑡𝑑𝐰01 (𝐯). Since Venezuela is situated in the left-hand side of the WMSD-
𝑚𝑒𝑎𝑛(𝐰𝟑 )
space (𝑚𝑒𝑎𝑛01 = 0.25) such an increase of variety does not take into account criteria weights and does not clearly show
𝐰 (𝐕𝐄𝐍) < 2
had a positive effect on the country’s ranking position. The above the ranking. For example, it is practically impossible to notice that
considerations show that WMSD-space is a useful tool not only for Venezuela is ranked higher than Suriname in 2023. Similarly, the
visualizing the impact that weight and aggregations have on the final change in ranking for Chile and Uruguay is hard to spot in the parallel
rankings but also for analyzing changes in rankings over time. coordinates plot, whereas it is clearly visible on the WMSD-space
As mentioned in Section 1, currently available methods that visu- visualization (Fig. 9). Finally, the gradient background of the WMSD
alize dependencies between alternatives and their rankings are very visualization shows the properties of the used aggregation function,
limited. The most common approach to visualizing alternatives is the whereas the parallel coordinates are completely decoupled from the
use of parallel coordinates plots [41]. Fig. 10 presents such a plot de- used aggregation.
picting the criteria values of Chile, Uruguay, Suriname, and Venezuela Summing up, the main goal of the economic freedom case study
in the 2019 and 2023 economic freedom rankings. The parallel co- was to show that WMSD-space visualizations can be applied to real-
ordinates plot is capable of showing all the criteria values for each world TOPSIS ranking problems. Since most publications on TOPSIS are
alternative. With only four criteria and eight alternatives, this makes applications of the method rather than contributions to the methodol-
Fig. 10 fairly readable. However, with more criteria and alternatives, ogy, it was important for us to provide practitioners with a problem
this type of visualization would very quickly become cluttered, whereas setting they are familiar with. That being said, the economic free-
WMSD-space can visualize any number of criteria and tens of alter- dom case study highlights several general decision-making implications
natives and remain readable. Moreover, the parallel coordinates plot connected with TOPSIS:

13
R. Susmaga et al. Applied Soft Computing 153 (2024) 111279

• Criteria weights affect the possible distances to the ideal and Finally, an issue worth discussing is the potential for using WMSD-
anti-ideal solution and, hence, the possible aggregation function space to design improvement actions. As discussed in Section 4 (Fig. 9),
values and the shape of WMSD-space. WMSD-space can show what actions (in terms of changing the mean
• Criteria weights affect aggregation values of the alternatives and, or standard deviation of an alternative) can be taken to change the
hence, their ranking positions. alternative’s ranking position. Of course, such mean and standard
• A permutation of weights (the same set of weights applied to deviation actions can be then translated into actions on criteria values.
different criteria) does not change the space of possible ideal/anti- In this context, the IA-WMSD property (Table 2) sums up the relations
ideal distances and, hence, does not change WMSD-space (al- between mean and standard deviation for different aggregations. It
though it can change ranking positions of individual alternatives). is, therefore, easy to foresee in WMSD-space whether the increase or
• Criteria weights do not affect the properties of aggregation func- decrease of mean or standard deviation will lead to the increase of the
tions and, hence, their gradient in WMSD-space. value of the aggregation function. As such, WMSD-space visualizations
could also prove useful in the potential sensitivity analyses of TOPSIS,
In the next section, we will discuss the limitations of WMSD visual- if such were attempted. Even more importantly, TOPSIS is often ap-
izations and the potential for their further modifications and applica- plied in high-stake domains that impact societies. As an example, in a
tions. recent sustainable energy planning project [58] environment, economy,
technical, and social experts each provided their own preference in
5. Discussion the form of weights on criteria concerning the costs and impact of
using renewable energy sources on a university campus. Such use
WMSD-space offers a visualization method for the standard TOPSIS cases could benefit from WMSD-space visualizations, which could be
methodology. We have shown that multicriteria alternative ranking used to compare rankings obtained from weights provided by different
problems can be visualized in 2D in terms of means and standard experts. Importantly, WMSD-space can help interpret how the mean
deviations, even when the number of criteria exceeds two and even and standard deviation of an alternative affect its position in a given
when the criteria are weighted. Since means and standard deviations ranking. Therefore, as future work, our goal is to make WMSD-space
used in the WMSD-space are easily interpretable, the space offers an visualizations easily available through an open-source tool and to apply
explainable view of TOPSIS rankings, showing whether the means or them to real-world decision-making problems.
standard deviations of particular alternatives impacted a given alterna-
tive’s ranking position. Moreover, by showing the aggregation values 6. Conclusions
of all possible alternatives as a colored gradient, WMSD-space shows
the properties of TOPSIS aggregation functions. In this paper, we have put forward a visual-based method for
Apart from restating the main characteristic of the proposed vi- explaining TOPSIS rankings in practical decision support applications
sualization method, it is worth enumerating the potential limitations with expert-defined criteria weights. To this end, weight-scaled means
of WMSD-space. First, we note that since WMSD-space is a visual- and standard deviations of alternatives were defined as generalizations
ization tool customized for TOPSIS, it inherits the limitations of the of means and standard deviations. Formalizing their relationship with
overall TOPSIS methodology. This implies requiring precise input data distances of an alternative to predefined ideal/anti-ideal points (IA-
(without any ambiguity), batch-like processing mode (without any WMSD property), we have proposed a generalization of MSD-space
user interaction/parametrization), and a linear pre-order output result called WMSD-space. Unlike MSD-space, which is based on regular
(without the relation of incomparability). Moreover, a distance measure means and standard deviation, WMSD-space is based on weight-scaled
for alternative representations must be defined, which often requires means and standard deviations of alternatives. Both, however, are
the criteria defined using finite ranges. Finally, the considered weights equally capable of representing alternatives and aggregation functions
must be non-negative, which is an accepted practice—in this common in a plane regardless of the number of considered criteria. What differ-
interpretation, weights express the magnitude of criterion importance entiates WMSD from MSD is taking into account the criteria weights.
and, as such, cannot be negative. In addition to limitations inherited As such, WMSD-space is a tool for visual-based comparisons of different
from TOPSIS, there occurs a limitation of the visualization method aggregation functions and the impact that weights defined by experts
itself, namely the need to use the Euclidean distance measure. The have on the final rankings. Therefore, the answer to the question
dimensionality reduction in WMSD-space is made possible because of asked in the Introduction, namely: Is it possible to propose a visualization
the application of the Pythagorean theorem, which is derived from the method that generalizes MSD-space to weighted criteria?, is thus definitely
axioms of Euclidean geometry. Thus, for the visualization to be possible positive.
regardless of the number of criteria, the Euclidean distance is needed. We stress that our paper proposes a visualization technique for the
Moreover, as far as weighting is concerned, we assume the standard standard TOPSIS method, not a new version of TOPSIS. The novelty
TOPSIS methodology. As a result, weights are applied multiplicatively, comes from considering all possible alternative representations (instead
which thus constitutes a linear operation. Potential attempts to make of particular datasets), which makes it possible to visualize gradients of
TOPSIS use weights that are in a non-linear relationship with criteria aggregation functions, and the fact that our visualization method is al-
values would require an explicit modification of the method itself. ways two-dimensional, even when the number of criteria is greater than
Additionally, if our WMSD-space based visualization were to handle two. To highlight the practical usefulness of the proposed visualization,
this modified version of TOPSIS, it would require new, potentially two case studies were conducted on a dataset of students described
non-linear versions of projection and rejection. This is because a non- in terms of school grades and on a dataset of countries described in
linear transformation from 𝑈 𝑆 to 𝑉 𝑆 would, in particular, result in terms of factors constituting the Index of Economic Freedom. Using
a non-linear diagonal 𝐷𝟎𝐰 . Should the results of projections onto and WMSD-space visualizations, we discuss how weights affect rankings of
rejections from such a non-linear diagonal be possible and unique, the alternatives under various TOPSIS aggregations and compare the effects
visualization methodology presented in this paper would still hold. of weights provided by multiple experts.
It is also worth noting that the proposed WMSD-space visualization As future research, the WMSD-space methodology could be ex-
does not introduce any significant computational overhead compared tended to take into account uncertainty connected with the input data
to the standard TOPSIS procedure. More precisely, to calculate the or criteria weights. This could involve representing data in the form
coordinates of alternatives in WMSD-space, one needs to calculate the of fuzzy sets, as they are commonly used to model the imprecision
weight-scaled means and standard deviations of each alternative. For and uncertainty of data in many application domains [59,12,7,60].
𝑚 alternatives described by 𝑛 criteria, this requires 𝑂(𝑚𝑛) time. We expect that such fuzzy input will result in alternatives represented

14
R. Susmaga et al. Applied Soft Computing 153 (2024) 111279

𝑣−0
not as points but as polygons—subregions of WMSD-space. On the • 1 (𝑣) = 100−0
,
other hand, the uncertainty of the decision-makers, represented as • 2 (𝑣) = 𝑣−0
.
20−0
fuzzy criteria weights, would require a 3D visualization that would
allow for interpolation between WMSD spaces of potentially different Consider an alternative represented as 𝐸 = [75, 10] ∈ 𝐶𝑆. Its represen-
shapes. Future research could also include the development of TOPSIS tation in 𝑈 𝑆 is found by an application of 1 (𝑣) and 2 (𝑣) to 𝐸 and
modifications that would control the impact that alternative means turns out to be
and standard deviations have on the final rankings. Such modifications 75 − 0 10 − 0
𝐮 = [1 (75), 2 (10)] = [ , ]=
would involve a parameter letting the user define whether the rankings 100 − 0 20 − 0
should be more influenced by the weight-scaled means or standard = [0.7500, 0.5000].
deviations. Such an analysis would generalize the results presented here
Applying the weights vector 𝐰 to 𝐮 creates the 𝑉 𝑆 representation of 𝐮
to other distance-based MCDA ranking methods such as UTA [61] or
SAW [62]. Similarly, WMSD visualization may be potentially applied to as
TOPSIS generalizations based on three-way decision [63,64]. Finally, to 𝐯 = 𝐰◦𝐮 = [1.0000, 0.5000]◦[0.7500, 0.5000] =
meet the needs of practitioners, we plan to develop a publicly available
= [0.7500, 0.2500].
open-source interactive dashboard for WMSD visualizations of user-
provided datasets. Combined with a new spectrum of aggregations Now, projecting 𝐯 = [0.7500, 0.2500] onto 𝐰 = [1.0000, 0.5000]
and possible improvement actions, it would make a valuable tool for produces:
hands-on multi-criteria decision analysis.
𝐯 ⋅ 𝐰𝑇
𝐯↘𝐰 = 𝐰=
CRediT authorship contribution statement ‖𝐰‖22
[0.7500, 0.2500] ⋅ [1.0000, 0.5000]
= ⋅ [1.0000, 0.5000]
Robert Susmaga: Writing – review & editing, Writing – origi- ‖[1.0000, 0.5000]‖2
‖ ‖2
nal draft, Visualization, Software, Methodology, Investigation, Formal 0.7500 ⋅ 1.0000 + 0.2500 ⋅ 0.5000
analysis, Conceptualization. Izabela Szczęch: Writing – review & edit- = √ ⋅ [1.0000, 0.5000] =
2
ing, Writing – original draft, Supervision, Methodology, Investigation, 1.0000 ⋅ 1.0000 + 0.5000 ⋅ 0.5000
Formal analysis, Data curation, Conceptualization. Dariusz Brzezinski: 0.7500 ⋅ 1.0000 + 0.2500 ⋅ 0.5000
= ⋅ [1.0000, 0.5000]
Writing – review & editing, Writing – original draft, Visualization, 1.0000 ⋅ 1.0000 + 0.5000 ⋅ 0.5000
0.875
Validation, Software, Methodology, Investigation, Funding acquisition. = ⋅ [1.0000, 0.5000] = 0.7000 ⋅ [1.0000, 0.5000] =
1.25
= [0.7000, 0.3500].
Declaration of competing interest
Simultaneously, rejecting 𝐯 = [0.7500, 0.2500] from 𝐰 = [1.0000, 0.5000]
The authors declare the following financial interests/personal rela- produces:
tionships which may be considered as potential competing interests:
Izabela Szczech reports financial support was provided by National 𝐯↗𝐰 = 𝐯 − 𝐯↘𝐰 =
Science Centre Poland. Robert Susmaga reports financial support was = [0.7500, 0.2500] − [0.7500, 0.2500] ↘ [1.0000, 0.5000]
provided by National Science Centre Poland. Dariusz Brzezinski reports
= [0.7500, 0.2500] − [0.7000, 0.3500] =
financial support was provided by National Science Centre Poland.
= [0.0500, −0.1000].
Data availability For the given 𝐰 = [1.0000, 0.5000] one gets
‖𝐰‖2 ‖[1.0000, 0.5000]‖
We have shared the data as on-line supplementary materials. ‖ ‖2 1.1180
𝑠= = = = 1.4907.
𝑚𝑒𝑎𝑛(𝐰) 𝑚𝑒𝑎𝑛([1.0000, 0.5000]) 0.7500
Acknowledgments Applying this value in the definitions of 𝑚𝑒𝑎𝑛01 01
𝐰 (𝐯) and 𝑠𝑡𝑑𝐰 (𝐯) results
in:
We would like to thank Adam Ciesiółka, Dariusz Grynia, Bogna Kali-
nowska, and Maciej Woś for their work on the Python implementation ‖𝐯 ↘ 𝐰‖2 ‖ ‖
‖[0.7000, 0.3500]‖2 0.7826
𝑚𝑒𝑎𝑛01
𝐰 (𝐯) = = = = 0.5250,
of the WMSD procedure. This research was partly funded by the Na- 𝑠 1.4907 1.4907
tional Science Centre, Poland, grant number: 2022/47/D/ST6/01770. ‖𝐯 ↗ 𝐰‖2 ‖ ‖
‖[0.0500, −0.1000]‖2 0.1118
𝑠𝑡𝑑𝐰01 (𝐯) = = = = 0.0750.
𝑠 𝑠 1.4907
Appendix A. Detailed example: Space transformations, distance
Because 𝐯 = 𝐯 ↘ 𝐰 = [0.7000, 0.3500] and 𝐯 ↗ 𝐰 = [0.0500, −0.1000],
computations, IA-WMSD property verifications and aggregation
one gets:
values
‖ ‖
‖𝐯 − 𝟎‖ ‖ ‖
All fractional values in the following computations are presented 01 ‖ ‖2 ‖[0.7000, 0.3500] − [0.0000, 0.0000]‖2
𝛿𝐰 (𝐯, 𝟎) = =
𝑠 𝑠
with four significant digits. For a summary of the notation used, see ‖[0.7000, 0.3500]‖
Table 1. ‖ ‖2 0.7826
= = = 0.5250
𝑠 1.4907
01 (𝐯, 𝟎) = 𝑚𝑒𝑎𝑛01 (𝐯) = 0.5250),
(clearly, 𝛿𝐰
A.1. A 2-dimensional example 𝐰
‖ ‖
‖𝐯 − 𝐰‖ ‖ ‖
Let 𝑉1 = [0, 100] and 𝑉2 = [0, 20] be the domains of two type gain 01 ‖ ‖2 ‖[0.7000, 0.3500] − [1.0000, 0.5000]‖2
𝛿𝐰 (𝐯, 𝐰) = =
criteria: 1 and 2 , respectively. This means that 𝑛 = 2, so 𝐶𝑆 and 𝑠 𝑠
𝑈 𝑆 will be 2-dimensional. Additionally, let 𝐰 = [1.0000, 0.5000], which ‖[−0.3000, −0.1500]‖
‖ ‖2 0.3354
= = = 0.2250
implies 𝑛𝑝 = 2 = 𝑛 (no zero weights) and means that also 𝑉 𝑆 will be 𝑠 1.4907
2-dimensional. 01 (𝐯, 𝐰) = 𝑚𝑒𝑎𝑛(𝐰) − 𝑚𝑒𝑎𝑛01 (𝐯) = 0.7500 − 0.5250 = 0.2250),
(clearly, 𝛿𝐰 𝐰
Because both criteria are of type gain, the re-scaling functions 1 (𝑣)
‖ ‖
and 2 (𝑣) corresponding to criteria 1 and 2 , respectively, will be ‖𝐯 − 𝐯‖ ‖ ‖
01 ‖ ‖2 ‖[0.7000, 0.3500] − [0.7500, 0.2500]‖2
defined as follows: 𝛿𝐰 (𝐯, 𝐯) = =
𝑠 𝑠

15
R. Susmaga et al. Applied Soft Computing 153 (2024) 111279

Because all criteria are of type gain, the re-scaling functions 𝑖 (𝑣)
𝑣−0
corresponding to criteria 𝑖 will be defined as follows: 𝑖 (𝑣) = 100−0 .
Consider a vector [62.43, 82.43, 75.37, 81.27] ∈ 𝐶𝑆, which is the repre-
sentation of Chile. Its image in 𝑈 𝑆 is found by an application of 𝑖 (𝑣)
to [62.43, 82.43, 75.37, 81.27] and turns out to be

𝐮 = [1 (62.43), 2 (82.43), 3 (75.37), 4 (81.27)]

62.43 − 0 82.43 − 0 75.37 − 0 81.27 − 0
=[ , , , ]
100 − 0 100 − 0 100 − 0 100 − 0
= [0.6243, 0.8243, 0.7537, 0.8127].

Applying the weights vector 𝐰 to 𝐮 creates the 𝑉 𝑆 representation of 𝐮

as

Fig. A.1. An illustration of how the weight-scaled mean (WM) and standard deviation 𝐯 = 𝐰◦𝐮
(WSD) define the WMSD-space by the IA-WMSD property.
= [0.2500, 1.0000, 0.2500, 0.5000]◦[0.6243, 0.8243, 0.7537, 0.8127] =
= [0.1561, 0.8243, 0.1884, 0.4063].
‖[−0.0500, −0.1000]‖
‖ ‖2 0.1118 Now, projecting 𝐯 = [0.1561, 0.8243, 0.1884, 0.4063] onto 𝐰 = [0.2500,
= = = 0.0750
𝑠 1.4907
1.0000, 0.2500, 0.5000] produces:
(clearly, 𝛿𝐰 01 (𝐯, 𝐯) = 𝑠𝑡𝑑 01 (𝐯) = 0.0750).
𝐰
Finalizing the example, we get: 𝐯 ⋅ 𝐰𝑇
𝐯↘𝐰 = 𝐰=
‖𝐯 − 𝟎‖ ‖ ‖ ‖𝐰‖22
01 ‖ ‖2 ‖[0.7500, 0.2500] − [0.0000, 0.0000]‖2 [0.1561, 0.8243, 0.1884, 0.4063] ⋅ [0.2500, 1.0000, 0.2500, 0.5000]
𝛿𝐰 (𝐯, 𝟎) = =
𝑠 1.4907 =
‖[0.7500, 0.2500]‖ ‖[0.2500, 1.0000, 0.2500, 0.5000]‖2
‖ ‖2 0.7906 ‖ ‖2
= = = 0.5304 ⋅ [0.2500, 1.0000, 0.2500, 0.5000] =
1.4907 1.4907
0.1561 ⋅ 0.25 + 0.8243 ⋅ 1.00 + 0.1884 ⋅ 0.25 + 0.4063 ⋅ 0.50
‖𝐯 − 𝐰‖2 ‖ ‖ = √
01 ‖[0.7500, 0.2500] − [1.0000, 0.5000]‖2 2
𝛿𝐰 (𝐯, 𝐰) = = 0.25 ⋅ 0.25 + 1.00 ⋅ 1.00 + 0.25 ⋅ 0.25 + 0.50 ⋅ 0.50
𝑠 1.4907
‖[−0.2500, −0.2500]‖ ⋅ [0.2500, 1.0000, 0.2500, 0.5000] =
‖ ‖2 0.3536
= = = 0.2372 0.1561 ⋅ 0.25 + 0.8243 ⋅ 1.00 + 0.1884 ⋅ 0.25 + 0.4063 ⋅ 0.50
1.4907 1.4907 =
0.25 ⋅ 0.25 + 1.00 ⋅ 1.00 + 0.25 ⋅ 0.25 + 0.50 ⋅ 0.50
which allows to verify that:
√ ⋅ [0.2500, 1.0000, 0.2500, 0.5000] =
√ √
• 𝑚𝑒𝑎𝑛01 2 01
𝐰 (𝐯) + 𝑠𝑡𝑑𝐰 (𝐯)
2 = 0.52502 + 0.07502 = 0.2813 = =
1.1136
⋅ [0.2500, 1.0000, 0.2500, 0.5000] = 0.8099
0.5304 = 𝛿𝐰01 (𝐯, 𝟎), 1.3750
√
⋅ [0.2500, 1.0000, 0.2500, 0.5000] =
• 𝑚𝑒𝑎𝑛(𝐰) − (𝑚𝑒𝑎𝑛01 2 01
𝐰 (𝐯)) + 𝑠𝑡𝑑𝐰 (𝐯)
2 =
√ √ √ = [0.2025, 0.8099, 0.2025, 0.4049].
(0.7500 − 0.5250)2 + 0.07502 = 0.22502 + 0.07502 = 0.0563 =
01 (𝐯, 𝐰).
0.2372 = 𝛿𝐰 Simultaneously, rejecting 𝐯 = [0.1561, 0.8243, 0.1884, 0.4063] from 𝐰 =
All these calculated values are shown in Fig. A.1. [0.2500, 1.0000, 0.2500, 0.5000] produces:
What remains to be computed with 𝛿𝐰 01 (𝐯, 𝟎) = 0.5304 and 𝛿 01 (𝐯, 𝐰) =
𝐰 𝐯↗𝐰 = 𝐯 − 𝐯↘𝐰 =
0.2327 is:
01 (𝐯,𝐰)
= [0.1561, 0.8243, 0.1884, 0.4063] − [0.1561, 0.8243, 0.1884, 0.4063]
𝛿𝐰
• 𝖨𝐰 (𝐯) = 1 − = 1 − 0.2327 = 1 − 0.3103 = 0.6897,
𝑚𝑒𝑎𝑛(𝐰) 0.7500 ↘ [0.2500, 1.0000, 0.2500, 0.5000] =
01
𝛿𝐰 (𝐯,𝟎) 0.5304
• 𝖠𝐰 (𝐯) = 𝑚𝑒𝑎𝑛(𝐰)
= 0.7500 = 0.7072, = [0.1561, 0.8243, 0.1884, 0.4063]
01 (𝐯,𝟎)
𝛿𝐰 0.5304
• 𝖱𝐰 (𝐯) = 01 (𝐯,𝐰)+𝛿 01 (𝐯,𝟎) = 0.2327+0.5304 = 0.6951. − [0.2025, 0.8099, 0.2025, 0.4049] =
𝛿𝐰 𝐰

All these aggregation values can also be observed color-coded in Fig. 6. = [−0.0464, 0.0144, −0.0141, 0.0014].

For the given 𝐰 = [0.2500, 1.0000, 0.2500, 0.5000] one gets

A.2. A 4-dimensional example
‖𝐰‖2 ‖[0.2500, 1.0000, 0.2500, 0.5000]‖
‖ ‖2 1.1726
𝑠= = = = 2.3452.
Let us now consider a fully analogous but slightly more-dimensional 𝑚𝑒𝑎𝑛(𝐰) 𝑚𝑒𝑎𝑛([0.2500, 1.0000, 0.2500, 0.5000]) 0.5000
(precisely: 4-dimensional) example to appreciate the actual dimension-
Applying this value in the definitions of 𝑚𝑒𝑎𝑛01 01
𝐰 (𝐯) and 𝑠𝑡𝑑𝐰 (𝐯) results
ality reduction that takes place in such a situation. In the previous
in:
example, a vector from 2-dimensional 𝐶𝑆 was transformed to equally
2-dimensional WMSD-space. Now, a vector from 4-dimensional 𝐶𝑆 will ‖𝐯 ↘ 𝐰‖2 ‖ ‖
‖[0.2025, 0.8099, 0.2025, 0.4049]‖2
𝑚𝑒𝑎𝑛01
𝐰 (𝐯) = =
be transformed to 2-dimensional WMSD-space, although no information 𝑠 2.3452
relevant to the TOPSIS method will be lost. This ‘lossless’ character of the 0.9497
= = 0.4049,
reduction concerns of course all 𝑛 > 2. 2.3452
Let 𝑖 ∈ {1, 2, 3, 4}. Additionally, let 𝑉𝑖 = [0, 100] be the domains of
criteria 𝑖 , all of type gain. This means that 𝑛 = 4, so 𝐶𝑆 and 𝑈 𝑆 will ‖𝐯 ↗ 𝐰‖2 ‖ ‖
‖[−0.0464, 0.0144, −0.0141, 0.0014]‖2
𝑠𝑡𝑑𝐰01 (𝐯) = =
be 4-dimensional. Now, let 𝐰 = [0.2500, 1.0000, 0.2500, 0.5000], which 𝑠 𝑠
corresponds to the second sets of weights in the analysis of Index of 0.0506
= = 0.0216.
Economic Freedom, as described in Section 4.2 of Section 4. The vector 2.3452
of weights implies 𝑛𝑝 = 4 = 𝑛 (no zero weights) and means that also The above coordinates of Chile (CHL) in WMSD-space can noticed
𝑉 𝑆 will be 4-dimensional. on Fig. 8B.

16
R. Susmaga et al. Applied Soft Computing 153 (2024) 111279

01 (𝐯,𝐰)
𝛿𝐰 0.0975
Because 𝐯 = 𝐯 ↘ 𝐰 = [0.2025, 0.8099, 0.2025, 0.4049] and 𝐯 ↗ 𝐰 = • 𝖨𝐰 (𝐯) = 1 − 𝑚𝑒𝑎𝑛(𝐰)
=1− 0.5000
= 1 − 0.1949 = 0.8051,
[−0.0464, 0.0144, −0.0141, 0.0014], one gets: 01 (𝐯,𝟎)
𝛿𝐰 0.5055
• 𝖠𝐰 (𝐯) = 𝑚𝑒𝑎𝑛(𝐰)
= 0.5000
= 0.8110,
‖ ‖
‖𝐯 − 𝟎‖ • 𝖱𝐰 (𝐯) =
01 (𝐯,𝟎)
𝛿𝐰 0.4055
01 ‖ ‖2 01 (𝐯,𝐰)+𝛿 01 (𝐯,𝟎) = = 0.8062.
𝛿𝐰 (𝐯, 𝟎) = = 𝛿𝐰 𝐰 0.0975+0.4055
𝑠
‖[0.2025, 0.8099, 0.2025, 0.4049] − [0.000, 0.000, 0.000, 0.000]‖ This final value (0.8062) can also be seen in column 4 of the first row
‖ ‖2
= of Table 6 (three significant digits).
𝑠
‖[0.2025, 0.8099, 0.2025, 0.4049]‖
‖ ‖2
= = Appendix B. The lower and upper boundary of the WMSD-space
𝑠
0.9497
= = 0.4049 Recall that 𝑚𝑒𝑎𝑛01
2.3452 𝐰 (𝐯) is defined as the re-scaled length of a pro-
01 (𝐯, 𝟎) = 𝑚𝑒𝑎𝑛01 (𝐯) = 0.4049), jection of 𝐯 ∈ 𝑉 𝑆 onto 𝐰, 𝑠𝑡𝑑𝐰01 (𝐯) is defined as the re-scaled length
(clearly, 𝛿𝐰 𝐰 of a rejection of 𝐯 ∈ 𝑉 𝑆 from 𝐰, and that the diagonal 𝐷𝟎𝐰 equals
01
𝛿𝐰 (𝐯, 𝐰) {𝑑 ⋅ 𝐰 ∣ 𝑑 ∈ [0, 1]}.
‖ ‖ Additionally, given 𝐰 such that 𝑛𝑝 ≥ 2, and 𝑡 ∈ [0, 𝑚𝑒𝑎𝑛(𝐰)], consider
‖𝐯 − 𝐰‖
‖ ‖2 the following parametrized (non-linear) programming problems:
= =
𝑠
‖[0.2025, 0.8099, 0.2025, 0.4049] − [0.2500, 1.0000, 0.2500, 0.5000]‖ find 𝑝𝑙 (𝑡) as: min 𝑠𝑡𝑑𝐰01 (𝐯)
‖ ‖2 𝐯∈𝑉 𝑆
= subject to: 𝑚𝑒𝑎𝑛01
𝑠 𝐰 (𝐯) = 𝑡
‖−0.0475, −0.1901, −0.0475, −0.0951‖
‖ ‖2 and
= =
𝑠 find 𝑝𝑢 (𝑡) as: max 𝑠𝑡𝑑𝐰01 (𝐯)
0.2229 𝐯∈𝑉 𝑆
= = 0.0950
2.3452 subject to: 𝑚𝑒𝑎𝑛01
𝐰 (𝐯) = 𝑡
01 (𝐯, 𝐰) = 𝑚𝑒𝑎𝑛(𝐰) − 𝑚𝑒𝑎𝑛01 (𝐯) = 0.5000 − 0.4049 = 0.0950),
(clearly, 𝛿𝐰 where 𝑡 is the parameter.
𝐰
01
For all 𝑡 ∈ [0, 𝑚𝑒𝑎𝑛(𝐰)] the resulting vectors:
𝛿𝐰 (𝐯, 𝐯)
‖ ‖ • [𝑡, 𝑝𝑙 (𝑡)] represent points of the ‘lower perimeter’ of WMSD-space,
‖𝐯 − 𝐯‖
‖ ‖2 • [𝑡, 𝑝𝑢 (𝑡)] represent points of the ‘upper perimeter’ of WMSD-space.
= =
𝑠
‖[0.2025, 0.8099, 0.2025, 0.4049] − [0.1561, 0.8243, 0.1884, 0.4063]‖ As 𝑝𝑙 (𝑡) = 0 for all 𝑡 ∈ [0, 𝑚𝑒𝑎𝑛(𝐰)], the ‘lower perimeter’ of WMSD-space
‖ ‖2
= simply constitutes a horizontal segment. This is because in 𝑉 𝑆 the re-
𝑠
‖[0.0464, −0.0144, 0.0141, −0.0014]‖ scaled length of a rejection of 𝐯 ∈ 𝑉 𝑆 from 𝐰 (i.e. the ‘orthogonally’
‖ ‖2
= = computed distance between 𝐯 and 𝐷𝟎𝐰 ), is minimal when 𝐯 ∈ 𝐷𝟎𝐰 and
𝑠
0.0506 equals 0 in such cases.
= = 0.0216 On the other hand, 𝑝𝑢 (𝑡) = 0 for 𝑡 ∈ {0, 𝑚𝑒𝑎𝑛(𝐰)}, while 𝑝𝑢 (𝑡) > 0
2.3452
01 (𝐯, 𝐯) = 𝑠𝑡𝑑 01 (𝐯) = 0.0216). for 𝑡 ∈ (0, 𝑚𝑒𝑎𝑛(𝐰)) (it also satisfies the following symmetry condition:
(clearly, 𝛿𝐰 𝐰 𝑝𝑢 ( 𝑚𝑒𝑎𝑛(𝐰)
2
− ℎ) = 𝑝𝑢 ( 𝑚𝑒𝑎𝑛(𝐰)
2
+ ℎ) for every ℎ ∈ [0, 𝑚𝑒𝑎𝑛(𝐰)
2
]). This is
Finalizing the example, we get:
because in 𝑉 𝑆 the re-scaled length of a rejection of 𝐯 ∈ 𝑉 𝑆 from 𝐰
01
𝛿𝐰 (𝐯, 𝟎) (i.e. the ‘orthogonally’ computed distance between 𝐯 and 𝐷𝟎𝐰 ) is 0 when
‖𝐯 − 𝟎‖ 𝐯 ∈ {𝟎, 𝐰} and greater than 0 otherwise. In these cases, the maxima
‖ ‖2
= = are attained at the borders of 𝑉 𝑆 and, because of its intricate shape,
𝑠
‖[0.1561, 0.8243, 0.1884, 0.4063] − [0.000, 0.000, 0.000, 0.000]‖ are dependent of 𝑚𝑒𝑎𝑛01 𝐰 (𝐯). Unfortunately, the closed-form formula for
‖ ‖2 these maxima, and thus for the ‘upper perimeter’ of WMSD-space is
=
2.3452
‖[0.1561, 0.8243, 0.1884, 0.4063]‖ unknown in the general case (as opposed to MSD-space [35], which
‖ ‖2 constitutes a special case of WMSD-space for 𝐰 = 𝟏).
= =
2.3452
0.9510
= = 0.4055 Appendix C. Supplementary data
2.3452

01 Supplementary material related to this article can be found online

𝛿𝐰 (𝐯, 𝐰)
at https://doi.org/10.1016/j.asoc.2024.111279.
‖𝐯 − 𝐰‖2
= =
𝑠 References
‖[0.1561, 0.8243, 0.1884, 0.4063] − [0.2500, 1.0000, 0.2500, 0.5000]‖
‖ ‖2
=
2.3452 [1] V. Belton, T.J. Stewart, Multiple Criteria Decision Analysis: An Integrated
‖[−0.0939, −0.1757, −0.0616, −0.0937]‖
‖ ‖2 Approach, Springer, 2002.
= = [2] A. Ishizaka, P. Nemery, Multi-Criteria Decision Analysis: Methods and Software,
2.3452
0.2286 Wiley, 2013.
= = 0.0975 [3] R. Bisdorff, L. Dias, P. Meyer, V. Mousseau, M. Pirlot, Evaluation and Decision
2.3452 Models with Multiple Criteria, Springer, 2015.
which allows us to verify that: [4] S. Greco, M. Ehrgott, J.R. Figueira, Multiple Criteria Decision Analysis: State of
√ √ √ the Art Surveys, Springer, 2016.
• 𝑚𝑒𝑎𝑛01 2 01 2
𝐰 (𝐯) + 𝑠𝑡𝑑𝐰 (𝐯) = 0.40492 + 0.02162 = 0.1644 = 0.4054 = [5] M. Cinelli, M. Kadziński, G. Miebs, M. Gonzalez, R. Słowiński, Recommending
01 (𝐯, 𝟎), multiple criteria decision analysis methods with a new taxonomy-based decision
𝛿𝐰
√ √ support system, European J. Oper. Res. 302 (2022) 633–651.
• 𝑚𝑒𝑎𝑛(𝐰) − (𝑚𝑒𝑎𝑛01 2 01 2
𝐰 (𝐯)) + 𝑠𝑡𝑑𝐰 (𝐯) = (0.5000 − 0.4049)2 + 0.02162 [6] C.L. Hwang, K. Yoon, Multiple Attribute Decision Making: Methods and
√ √ Applications, Springer-Verlag, 1981.
2 2
= 0.0951 + 0.0216 = 0.0095 = 0.0975 = 𝛿𝐰 01 (𝐯, 𝐰).
[7] E. Bottani, A. Rizzi, A fuzzy TOPSIS methodology to support outsourcing of
logistics services, Supply Chain Manag. 11 (2006) 294–308.
01 (𝐯, 𝟎) = 0.4055 and 𝛿 01 (𝐯, 𝐰) =
What remains to be computed with 𝛿𝐰 𝐰 [8] W.-P. Wang, Toward developing agility evaluation of mass customization systems
0.0975 is: using 2-tuple linguistic computing, Expert Syst. Appl. 36 (2009) 3439–3447.

17
R. Susmaga et al. Applied Soft Computing 153 (2024) 111279

[9] Z. Zhang, H. Jiang, T. Shao, Q. Shao, Understanding the selection of intelligent [38] R. Guidotti, A. Monreale, S. Ruggieri, F. Turini, F. Giannotti, D. Pedreschi, A
engineering B2B platform in China through the fuzzy DANP and TOPSIS survey of methods for explaining black box models, ACM Comput. Surv. (CSUR)
techniques: A multi-study analysis, Appl. Soft Comput. 141 (2023) 110277. 51 (2018) 1–42.
[10] X. Yu, S. Guo, J. Guo, X. Huang, Rank B2C e-commerce websites in e-alliance [39] B. Pradhan, A. Dikshit, S. Lee, H. Kim, An explainable AI (XAI) model for
based on AHP and fuzzy TOPSIS, Expert Syst. Appl. 38 (2011) 3550–3557. landslide susceptibility modeling, Appl. Soft Comput. 142 (2023) 110324.
[11] M. Piwowarski, D. Miłaszewicz, M.L. atuszyńska, M. Borawski, K. Nermend, [40] S. Itani, F. Lecron, P. Fortemps, A one-class classification decision tree based on
TOPSIS and VIKOR methods in study of sustainable development in the EU kernel density estimation, Appl. Soft Comput. 91 (2020) 106250.
countries, Procedia Comput. Sci. 126 (2018) 1683–1692. [41] P. Ziemba, M. Piwowarski, K. Nermend, Visualization of uncertain data in the
[12] S.-S. Lin, A. Zhou, S.-L. Shen, Safety assessment of excavation system via TOPSIS- NEAT F-PROMETHEE method, MethodsX 10 (2023) 102166.
based MCDM modelling in fuzzy environment, Appl. Soft Comput. 138 (2023) [42] J. Černevičienė, A. Kabašinskas, Review of multi-criteria decision-making meth-
110206. ods in finance using explainable artificial intelligence, Front. Artif. Intell. 5
[13] M. Behzadian, S. Khanmohammadi Otaghsara, M. Yazdani, J. Ignatius, A (2022) 827584.
state-of the-art survey of TOPSIS applications, Expert Syst. Appl. 39 (2012) [43] K.W. De Bock, K. Coussement, A.D. Caigny, R. Słowiński, B. Baesens, R.N. Boute,
13051–13069. T.-M. Choi, D. Delen, M. Kraus, S. Lessmann, S. Maldonado, D. Martens, M.
[14] E. Zavadskas, A. Mardani, Z. Turskis, A. Jusoh, K. Nor, Development of Óskarsdóttir, C. Vairetti, W. Verbeke, R. Weber, Explainable AI for operational
TOPSIS method to solve complicated decision-making problems - an overview research: A defining framework, methods, applications, and a research agenda,
on developments from 2000 to 2015, Int. J. Inf. Technol. Decis. Mak. 15 (2016). European J. Oper. Res. (2023).
[15] S. Zyoud, D. Fuchs-Hanusch, A bibliometric-based survey on AHP and TOPSIS [44] D. Brzezinski, J. Stefanowski, R. Susmaga, I. Szczech, Visual-based analysis
techniques, Expert Syst. Appl. 78 (2017) 158–181. of classification measures and their properties for class imbalanced problems,
[16] S. Opricovic, G.H. Tzeng, Compromise solution by MCDM methods: A com- Inform. Sci. 462 (2018) 242–261.
parative analysis of VIKOR and TOPSIS, European J. Oper. Res. 156 (2004) [45] D. Brzezinski, J. Stefanowski, R. Susmaga, I. Szczech, Tetrahedron: Barycentric
445–455. measure visualizer, in: Machine Learning and Knowledge Discovery in Databases,
[17] E. Zavadskas, A. Zakarevičius, J. Antucheviciene, Evaluation of ranking accuracy Proceedings of ECML PKDD 2017, Part III, in: Lecture Notes In Computer Science,
in multi-criteria decisions, Informatica, Lith. Acad. Sci. 17 (2006) 601–618. vol. 10536, Springer, 2017, pp. 419–422.
[18] S. Chakraborty, C.-H. Yeh, A simulation comparison of normalization procedures [46] R. Susmaga, I. Szczech, Can interestingness measures be usefully visualized? Int.
for TOPSIS, in: 2009 International Conference on Computers & Industrial J. Appl. Math. Comput. Sci. 25 (2015) 323–336.
Engineering, 2009, pp. 1815–1820. [47] R. Susmaga, I. Szczech, Visualization support for the analysis of properties of
[19] D.-F. Li, Relative ratio method for multiple attribute decision making problems, interestingness measures, Bull. Pol. Acad. Sci. Tech. Sci. 63 (2015) 315–327.
Int. J. Inf. Technol. Decis. Mak. 08 (2009) 289–311. [48] J.M. Martel, B. Roy, Analyse de la signifiance de diverses procédures d’agrégation
[20] T. Kuo, A modified TOPSIS with a different ranking index, European J. Oper. multicritere, INFOR Inf. Syst. Oper. Res. 44 (2006) 191–214.
Res. 260 (2017) 152–160. [49] C.D. Meyer, Matrix Analysis and Applied Linear Algebra, SIAM, Philadelphia,
[21] S. Abootalebi, A. Hadi-Vencheh, A. Jamshidi, Ranking the alternatives with a 2000.
modified TOPSIS method in multiple attribute decision making problems, IEEE [50] Index of economic freedom, 2023, URL: https://www.heritage.org/index/.
Trans. Eng. Manage. (2019) 1–6. [51] A.B. Kim, 2023 Index of Economic Freedom, The Heritage Foundation, NY, USA,
[22] S. Greco, R. Słowiński, J.R. Figueira, V. Mousseau, Robust ordinal regression, in: 2023.
M. Ehrgott, J.R. Figueira, S. Greco (Eds.), Trends in Multiple Criteria Decision [52] D.F. de Lima Silva, L. Ferreira, A.T. de Almeida Filho, Preference disaggregation
Analysis, Springer US, Boston, MA, 2010, pp. 241–283. on TOPSIS for sorting applied to an economic freedom assessment, Expert Syst.
[23] P. Zielniewicz, A ranking method based on the aggregate distance measure Appl. 215 (2023) 119341.
function in the value space, Int. J. Inf. Technol. Decis. Mak. 16 (2017) 685–710. [53] A. Puska, A. Stilic, I. Stojanovic, Approach for multi-criteria ranking of balkan
[24] S. Corrente, M. Tasiou, A robust TOPSIS method for decision making problems countries based on the index of economic freedom, J. Decis. Anal. Intell. Comput.
with hierarchical and non-monotonic criteria, Expert Syst. Appl. 214 (2023) 3 (2023) 1–14.
119045. [54] S.C. Dinç, N.A. Erilli, Hybrid ranking proposal based on spatial econometrics:
[25] L. Yu, W. Yang, Y. Duan, X. Long, A study on the application of coordinated An example of European economic freedoms, Ekoist J. Econom. Stat. 36 (2022)
TOPSIS in evaluation of robotics academic journals, Math. Probl. Eng. 2018 205–233.
(2015) 5456064. [55] D.F. de Lima Silva, A.T. de Almeida Filho, Sorting with TOPSIS through boundary
[26] P. Chen, A novel coordinated TOPSIS based on coefficient of variation, and characteristic profiles, Comput. Ind. Eng. 141 (2020) 106328.
Mathematics 7 (2019). [56] I. Brkić, N. Gradojević, S. Ignjatijević, The impact of economic freedom on
[27] G. Tian, H. Zhang, M. Zhou, Z. Li, AHP, gray correlation, and TOPSIS combined economic growth? New European dynamic panel evidence, J. Risk Financial
approach to Green performance evaluation of design alternatives, IEEE Trans. Manag. 13 (2020) 26.
Syst. Man Cybern. A 48 (2018) 1093–1105. [57] T. Miller, A.B. Kim, J.M. Roberts, 2019 Index of Economic Freedom - 25th
[28] K.P. Yoon, W.K. Kim, The behavioral TOPSIS, Expert Syst. Appl. 89 (2017) Anniversary Edition, The Heritage Foundation, NY, USA, 2019.
266–272. [58] C. Becchio, M. Bottero, S. Corgnati, F. Dell’Anna, G. Vergerio, Energy audit and
[29] M.A.A. Cox, T.F. Cox, Multidimensional scaling, in: Handbook of Data multi-criteria decision analysis to identify sustainable strategies in the university
Visualization, Springer Berlin Heidelberg, Berlin, Heidelberg, 2008, pp. 315–347. campuses: Application to politecnico di torino, in: C. Bevilacqua, F. Calabrò,
[30] I. Borg, P.J.F. Groenen, Modern Multidimensional Scaling. Theory and L. Della Spina (Eds.), New Metropolitan Perspectives, Springer International
Applications, Springer Science+Business Media, New York, 2005. Publishing, Cham, 2021, pp. 1187–1197.
[31] M. Walesiak, Visualization of linear ordering results for metric data with the [59] C.T. Chen, Extension of the TOPSIS for group decision-making under fuzzy
application of multidimensional scaling, Econometrics. Ekonometria. Adv. Appl. environment, Fuzzy Sets and Systems 114 (2000) 1–9.
Data Anal. 2 (2016) 9–21. [60] E. Aydoğdu, E. Güner, B. Aldemir, H. Aygün, Complex spherical fuzzy TOPSIS
[32] L. van der Maaten, G. Hinton, Visualizing data using t-SNE, J. Mach. Learn. Res. based on entropy, Expert Syst. Appl. 215 (2023) 119331.
9 (2008) 2579–2605. [61] E. Jacquet-Lagrèze, Y. Siskos, Assessing a set of additive utility functions for
[33] L. van der Maaten, G.E. Hinton, Visualizing non-metric similarities in multiple multicriteria decision-making, the UTA method, European J. Oper. Res. 10 (1982)
maps, Mach. Learn. 87 (2012) 33–55. 151–164.
[34] C.O. Wilke, Fundamentals of Data Visualization: A Primer on Making Informative [62] F. Ciardiello, A. Genovese, A comparison between TOPSIS and SAW methods,
and Compelling Figures, O’Reilly Media, 2019. Ann. Oper. Res. (2023) 1–28.
[35] R. Susmaga, I. Szczech, P. Zielniewicz, D. Brzezinski, MSD-space: Visualizing [63] Y. Wang, P. Liu, Y. Yao, BMW-TOPSIS: A generalized TOPSIS model based on
the inner-workings of TOPSIS aggregations, European J. Oper. Res. 308 (2023) three-way decision, Inform. Sci. 607 (2022) 799–818.
229–242. [64] Q. Zhan, L. Jin, R.R. Yager, A modified TOPSIS approach with three-way
[36] J.-C. Lu, M.-J. Li, An improved TOPSIS within the DEA framework, Asia-Pac. J. decision, Int. J. Uncertain. Fuzziness Knowl.-Based Syst. 31 (2023) 795–823.
Oper. Res. (2023) 2250034.
[37] P. Chen, Effects of the entropy weight on TOPSIS, Expert Syst. Appl. 168 (2021)
114186.

18