You are on page 1of 11

Information Fusion 106 (2024) 102295

Contents lists available at ScienceDirect

Information Fusion
journal homepage: www.elsevier.com/locate/inffus

Representation, optimization and generation of fuzzy measures


Gleb Beliakov a ,∗,1 , Jian-Zhang Wu a ,1 , Weiping Ding b ,1
a
School of Information Technology, Deakin University, Geelong, VIC 3220, Australia
b
School of Information Science and Technology, Nantong University, 226019, China

ARTICLE INFO ABSTRACT

Keywords: We review recent literature on three aspects of fuzzy measures: their representations, learning optimal fuzzy
Fuzzy measure measures and random generation of various types of fuzzy measures. These three aspects are interdependent:
Equivalent representations methods of learning fuzzy measures depend on their representation, and may also include random generation
Fuzzy measure identification
as one of the steps, on the other hand different representations also affect generation methods, while random
Random generation
generation plays an important role in simulation studies for post-hoc analysis of sets of measures learned
Optimization
from data and problem-specific constraints. Explicit modelling of interactions between the decision variables
is a distinctive feature of integrals based on fuzzy measures, but its price is high computational complexity.
To extend their range of applicability efficient representations and computational techniques are required.
All three mentioned aspects provide mathematical and computational tools for novel applications of fuzzy
measures and integrals in decision making and information fusion, allow scaling up significantly the domain
of applicability and reduce their complexity.

1. Introduction This way it accounts not only for the ‘‘sizes’’ of the combined subsets
but also for their ‘‘interaction’’.
In this review we focus on the most recent developments in the In this review we focus on discrete finite sets and functions defined
field of fuzzy measures and integrals, specifically the ones associated on them. In the context of information fusion these universal sets are
with their representations and construction. There are several detailed decision criteria or other inputs that somehow interact with each other
treatments of fuzzy measures presented in the monographs [1–4], from and are combined into a representative value. So fuzzy measures are
which we borrowed the basic definitions. While information fusion and set functions defined on the power sets of inputs. We also impose two
aggregation is one particular and most relevant to this review area, generic conditions to model reasonable application specific require-
where fuzzy measures gained prominence in recent years, these set ments, a rather technical boundary condition (the range of values is
functions are extensively used in mathematical economics under the the unit interval), and the monotonicity condition, which states that the
name of cooperative games, in decision theory and pure mathematics. measure of a set is bigger than the measure of any of its subsets. This
In economics several prominent researchers have been awarded Nobel condition relaxes additivity (for which monotonicity obviously holds),
prizes for their advances in cooperative games, including Harsanyi, and gives rise to a large number of interesting types of behaviour when
Nash, Aumann and Shapley. aggregating the inputs by the fuzzy integrals defined with respect to
As their name indicates, fuzzy measures generalize classical mea- monotone nonadditive set functions.
sures — mathematical objects to quantify properties of sets. Indeed, In this review we focus on some particular aspects of fuzzy measures
fuzzy measures are set functions, that is, functions defined on collec- which are valuable in many operational tasks, such as performing
tions of subsets of some universal set. A classical measure is an additive efficient calculations and simplifying fuzzy measures, simulations in-
function — the measure of a union of two non-intersecting subsets is the volving fuzzy integrals, as well as quantifying and interpreting certain
sum of the measures of each subset. In contrast, a fuzzy, or nonadditive properties of fuzzy measures such as degrees of interaction in a subset
measure results in a value which can be smaller or greater than the sum. of inputs.

∗ Corresponding author.
E-mail address: gleb@deakin.edu.au (G. Beliakov).
1
All autors share equal credit for the development of the methods and the manuscript.

https://doi.org/10.1016/j.inffus.2024.102295
Received 5 November 2023; Received in revised form 1 February 2024; Accepted 4 February 2024
Available online 7 February 2024
1566-2535/© 2024 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
G. Beliakov et al. Information Fusion 106 (2024) 102295

Our survey provides the following contributions. (1) An analysis of Nonmodularity is a broader concept compared to nonadditivity.
recent literature on representations of fuzzy measures and transforms Typically, an additive or modular fuzzy measure indicates a degree
between them, suitability of representations to particular families; (2) of independence among all elements. Similarly, a superadditive/super-
a review of recent methods of learning fuzzy measures from data modular (subadditive/submodular) measure represents complementar-
and problem-specific requirements; (3) a review of recent methods for ity (substitutability) among the inputs, respectively. Supermodularity
random sampling the set of fuzzy measures and its special families. is often associated with the notion of convexity of set functions.
These three areas are inter-dependent: different representations are
suitable for developing efficient methods of learning fuzzy measures, Definition 4 ([8,20]). The dual of a fuzzy measure 𝜇 on 𝑁 is defined as
while random sampling is often a particular step in the learning process. 𝜈(𝐴) = 𝜇(𝑁) − 𝜇(𝑁 ⧵ 𝐴), ∀𝐴 ⊆ 𝑁.
In another direction, learning fuzzy measures may identify not one
optimal fuzzy measure but a whole subset, which in turn can be Definition 5 ([1,10,22]). A fuzzy measure 𝜇 on 𝑁 is called symmetric
sampled to perform post-hoc and sensitivity analysis. measure, if for any 𝐴, 𝐵 ⊆ 𝑁, |𝐴| = |𝐵| implies 𝜇(𝐴) = 𝜇(𝐵). It is called
The review has three main sections after the Introduction and balanced if |𝐴| > |𝐵| implies 𝜇(𝐴) ≥ 𝜇(𝐵).
Preliminaries. In Section 3 we discuss various representations of fuzzy
measures and transformations between them, including also the com- The fuzzy integral is often used as an aggregation function with
putational aspects. In the next section we briefly outline recent progress respect to a fuzzy measure, with typical examples being the Choquet
in eliciting fuzzy measures from data by optimization and machine integral and the Sugeno integral.
learning. Then Section 5 outlines recent methods of random generation
of general and particular types of fuzzy measures for simulation studies. Definition 6 ([6,23]). For a given 𝐱 ∈ [0, 1]𝑛 , the discrete Choquet
The last section points to further directions and presents conclusions. integral  of 𝐱 with respect to fuzzy measure 𝜇 on 𝑁 is defined as
follows:

𝑛
2. Preliminaries (𝐱) = (𝑥(𝑖) − 𝑥(𝑖−1) )𝜇({(𝑖), … , (𝑛)}),
𝑖=1

Let 𝑁 = {1, 2, … , 𝑛}, 𝑛 ≥ 2, be the discrete set of input arguments, where 𝑥(.) is a non-decreasing permutation induced by 𝑥𝑖 , 𝑖 = 1, … , 𝑛,
items or criteria, (𝑁) the power set of 𝑁, and |𝑆| or 𝑠 the cardinality i.e., 𝑥(1) ≤ ⋯ ≤ 𝑥(𝑛) , and 𝑥(0) = 0 by convention.
of a subset 𝑆 ⊆ 𝑁.
Definition 7 ([5,24]). For a given 𝐱 ∈ [0, 1]𝑛 , the discrete Sugeno
Definition 1 ([5–8]). A set function 𝜇 ∶ (𝑁) → [0, 1] is called as a integral  of 𝑥 with respect to fuzzy measure 𝜇 on 𝑁 is defined as
fuzzy measure on 𝑁 if it satisfies follows:
(i) Boundary condition: 𝜇(∅) = 0, 𝜇(𝑁) = 1; ⋁
𝑛
(𝐱) = (𝑥(𝑖) ∧ 𝜇({(𝑖), … , (𝑛)})),
(ii) Monotonicity condition: ∀𝐴, 𝐵 ⊆ 𝑁, if 𝐴 ⊆ 𝐵 then 𝜇(𝐴) ≤ 𝜇(𝐵). 𝑖=1

A fuzzy measure is also called a capacity [6], monotone mea- where 𝑥(.) is a non-decreasing permutation induced by 𝑥𝑖 , 𝑖 = 1, … , 𝑛,
sure [9,10], nonadditive measure [7,11,12], sometimes with further i.e., 𝑥(1) ≤ ⋯ ≤ 𝑥(𝑛) .
qualifications such as ‘‘normalized’’. When condition (ii) is not satisfied,
Fuzzy integral-based aggregation functions are averaging, and they
a fuzzy measure is called a nonmonotonic fuzzy measure [13–15]. A
account not only for different contributions (weights) of individual in-
study on the evolution of notions of fuzzy measure can be found in [16]. puts, but also their coalitions. The special case of independent inputs is
The key feature of fuzzy measures is their ability to model explicitly represented by additive measures, and the respective Choquet integral
the interactions among the inputs, by quantifying not only the relative becomes the weighted arithmetic mean. For non-additive measures the
importance of an input but also that of all possible inputs’ coalitions. behaviour of fuzzy integrals is much richer, offering sophisticated ways
An input, such as decision criterion, may not be particularly important to aggregate the inputs accounting for redundancies and synergies.
by itself but gain importance in presence of other criteria.
Based on that feature one can design aggregation functions to com- 3. Representations and special measures
bine the inputs into a representative output value (or the total score),
which account for all inputs’ interactions. Such aggregation functions Fuzzy measures have multiple representations and one-to-one map-
are expressed by means of fuzzy integrals. pings between them, including the Möbius representation [25], prob-
The direction in which the inputs interact, that is, whether they abilistic simultaneous interaction index [26], possibilistic/necessity
reinforce or weaken each other, are quantified by various interaction Möbius transform [8,27], nonadditivity index [18], nonmodularity in-
indices and also the overall behaviour of fuzzy measures related to their dex [19], probabilistic bipartition interaction index [28], and (max- +)
‘‘convexity’’. transform [29]. These representations play significant roles in eliciting
The interaction index of a subset, which extends the Shapley val- practical interpretations of fuzzy measures, and also in constructing
ues, quantifies the simultaneous marginal interaction. The nonadditiv- special measures to manage the exponential growth (with 𝑛) of the
ity and modularity indices serve as quantifiers of the convexity in a number of coefficients associated with them.
fuzzy measure, providing intuitive insights into the diverse range of An important feature is that most transformations are linear, thus
interaction phenomena observed among elements. allowing to perform calculations in the linear algebra formalism [30,
31] (see Fig. 1 and Table 1).
Definition 2 ([11,12,17,18]). A fuzzy measure 𝜇 on 𝑁 is additive (su-
peradditive, subadditive), if 𝜇(𝐴 ∪ 𝐵) = (≥, ≤)𝜇(𝐴) + 𝜇(𝐵) for nonempty 3.1. Möbius representation and k-additive measures
disjoint subsets 𝐴, 𝐵 ⊆ 𝑁.
Definition 8 ([26,32–34]). Let 𝜇 be a fuzzy measure on 𝑁. The Möbius
Definition 3 ([1,2,19–21]). A fuzzy measure 𝜇 on a set 𝑁 is called value of a subset 𝐴 ⊆ 𝑁 w.r.t. 𝜇 is defined as
modular (supermodular, submodular) if 𝜇(𝐴 ∪ 𝐵) + 𝜇(𝐴 ∩ 𝐵) = (≥, ≤ ∑
𝑚𝜇 (𝐴) = (−1)|𝐴∖𝐶| 𝜇(𝐶),
)𝜇(𝐴) + 𝜇(𝐵), for all subsets 𝐴, 𝐵 ⊆ 𝑁. 𝐶⊆𝐴

2
G. Beliakov et al. Information Fusion 106 (2024) 102295

Table 1
Representations and special fuzzy measures.
Representations Special measures References Pros and Cons
Möbius 𝑘-additive measure [26,32–34] Reduces the number of parameters by
representation overlooking higher order interactions.
Easy expression of the Choquet integral.
Needs large number of monotonicity
constraints
Possibilistic and k-maxitive and [27,35] Reduces the number of parameters,
necessity Mobius minitive measure needs no extra constraints. Operations
transform nonlinear, Choquet integral not easily
expressed, but suitable for Sugeno
integrals
Nonadditivity/ k-nonadditive/- [1,18,19,36] Reduces the number of parameters but
nonmodularity nonmodular not monotonicity constraints. The
index measure measure sum based interpretation
approach of interaction in a subset.
Marginal buoyant measure [26,37–40] A subset of derivatives parameterizes the
contribution measure, but requires a number of other
constraints. Useful for modelling
convexity/concavity.
Fuzzy measure 𝑘-interactive, [41–43] Reduces the number of parameters and
itself 𝑘-(in)tolerant monotonicity constraints, easy
measure, 𝑝-symmetric expression for the Choquet integral. A
measure different interpretation of the parameters
as overall interaction in a set.

Definition 9 ([33]). Let 𝑘 ∈ {1, … , 𝑛}. A fuzzy measure 𝜇 on 𝑁 is said


to be 𝑘-additive if its Möbius representation satisfies 𝑚𝜇 (𝐴) = 0 for all
𝐴 ⊆ 𝑁 such that |𝐴| > 𝑘 and there exists at least one subset 𝐴 of 𝑘
elements such that 𝑚𝜇 (𝐴) ≠ 0.

1-additive measure is just the additive measure. To identify a 𝑘-


additive measure 𝜇 on 𝑁, we only need to define the Möbius values
∑ ()
of the 𝑘𝑖=1 𝑛𝑖 subsets whose cardinalities are not greater than 𝑘 [34]
(see Fig. 2).
The dual of any 𝑘-additive measure remains a 𝑘-additive measure.
Furthermore, an additive measure is self-dual.

3.2. Possibilistic and necessity Mobius transform and k-maxitive and mini-
tive measures

Definition 10 ([27]). The possibilistic Möbius transform of a fuzzy


measure 𝜇 on 𝑁 is a mapping 𝑚𝑝𝜇 ∶ (𝑁) → [0, 1] defined by


⎪𝜇(𝐴) if 𝜇(𝐴) > max
𝐵⊂𝐴
𝜇(𝐵),
Fig. 1. A graphical representation of a fuzzy measure for 𝑛 = 8. The sizes of the 𝑚𝑝 (𝐴) = ⎨ (1)
circles reflect the values of 𝜇(𝐴), and the edges reflect monotonicity constraints. The ⎪0 otherwise.
complexity of fuzzy measures is the price for their flexibility. ⎩
The possibilistic Zeta transform of 𝑚𝑝 is the mapping 𝑍𝑚𝑝 ∶ (𝑁) →
[0, 1] defined by:
The linear inverse of the Möbius representation is given as
∑ 𝑚𝑝 (𝐴) = max 𝑚𝑝 (𝐵). (2)
𝐵⊆𝐴
𝜇(𝐴) = 𝑚𝜇 (𝐵), ∀𝐴 ⊆ 𝑁.
𝐵⊆𝐴
Definition 11 ([27,35]). Let 𝑘 ∈ {1, … , 𝑛}. A fuzzy measure 𝜇 is called
Using this inverse transformation called zeta-transform, we can effort-
𝑘-maxitive if its possibilistic Möbius transform satisfies 𝑚𝑝 (𝐴) = 0 for
lessly derive the expressions for boundary and monotonicity conditions,
any 𝐴 such that |𝐴| > 𝑘 and there exists at least one subset 𝐴0 of 𝑁 of
as well as for fuzzy integrals in terms of the Möbius representation [13,
exactly 𝑘 elements such that 𝑚𝑝 (𝐴0 ) ≠ 0.
26,32,34,44].
The Choquet integral can be expressed in terms of Möbius represen- A 𝑘-maxitive fuzzy measure is totally determined by the fuzzy
tation as ∑ ()
measure values at 𝑘𝑖=1 𝑛𝑖 subsets whose cardinalities are equal to or

𝑚 (𝐱) = 𝑚𝜇 (𝐴) min(𝑥𝑖 ), less than 𝑘, since the 𝑘 + 1 and higher order subsets’ values can be
𝑖∈𝐴 obtained by:
𝐴⊆𝑁

Alternatively, it can be represented in terms of fuzzy measure as [44]: 𝜇(𝐴) = 𝜇(𝐵), |𝐴| ≥ 𝑘 + 1. (3)
∑ 𝐵⊂𝐴,|𝐵|=|𝐴|−1
𝜇 (𝐱) = 𝜇(𝐴) max(0, min 𝑥𝑖 − max 𝑥𝑖 ).
𝑖∈𝐴 𝑖∈𝑁∖𝐴
𝐴⊆𝑁 The dual of 𝑘-maxitive measure is called 𝑘-minitive fuzzy mea-
Some generalization of the Choquet integral expressed in Möbius sure [47]. 1-maxitive measure is just the possibility fuzzy measure, and
transform are presented in [45,46]. 1-minitive measure is the necessity fuzzy measure [47,48].

3
G. Beliakov et al. Information Fusion 106 (2024) 102295

Fig. 2. Interaction diagrams for (a) a full (unconstrained) fuzzy measure and (b) a 2-additive fuzzy measure where edges and nodes for 0-valued subsets (in Möbius representation)
are removed.

3.3. (Max, +) transform and 𝑘-maxitive measures A fuzzy measure 𝜇 on 𝑁 can be represented through nonadditivity
index 𝑛𝜇 as
Definition 12 ([29,49,50]). Let 𝜇 be a fuzzy measure on 𝑁. Then, we
define the (Max,+)-transform as the set function 𝜏𝜇 ∶ (𝑁) → [0, 1] ∑ |𝐴|−1
∑ |𝐴|−1 ∏ |𝐴| − 𝑗
𝜇(𝐴) = 𝑛𝜇 (𝐴) + 𝑛 (𝐶), ∀𝐴 ⊆ 𝑁. (7)
such that: 𝐶⊂𝐴 𝑖=|𝐶| 𝑗=𝑖
2𝑗 − 1 𝜇

𝜏𝜇 (𝐵) = 𝜇(𝐵) − max 𝜇(𝐴) (4)


𝐴⊂𝐵 Definition 14 ([19]). The nonmodularity index of subset 𝐴 with respect
It can be seen from this definition that 𝜏𝜇 is always positive and to a fuzzy measure 𝜇 on 𝑁 is given as
that for an arbitrary normalized measure 𝜇 and any set 𝐵 ⊂ 𝑁, the 1 ∑
(max, +)-transform 𝜏𝜇 is at most one. When 𝜇 is an additive measure, 𝑑𝜇 (𝐴) = 𝜇(𝐴) − [𝜇({𝑖}) + 𝜇(𝐴∖{𝑖})]. (8)
|𝐴| {𝑖}⊂𝐴
𝜏𝜇 (𝐵) = min𝑖∈𝐵 𝜇({𝑖}).
The inverse transform is obtained by recursion A fuzzy measure 𝜇 on 𝑁 can be represented through its nonmodu-
larity index 𝑑𝜇 as
𝜇(𝐵) = 𝜏𝜇 (𝐵) + max 𝜇(𝐴) (5)
𝐴⊂𝐵 ( )−1
∑ |𝐴| ∑
One can see here that the monotonicity condition is automatically 𝜇(𝐴) = 𝑑𝜇 (𝐴) + 𝑑𝜇 (𝐵) + 𝑑𝜇 ({𝑖}), ∀𝐴 ⊆ 𝑁. (9)
𝐵⊂𝐴,|𝐵|≥2
|𝐵| {𝑖}⊂𝐴
satisfied, but in difference to the possibilistic transform, the values
𝜏𝜇 (𝐴) express the marginal contribution to the total value of the set Nonadditivity and nonmodularity indices possess several desirable
𝐴. Some of those contributions can be explicitly set to 0, similar to properties, including additivity, uniform range, maximality, and mini-
the Möbius values. The transformations between Möbius, possibilistic mality [1,18,19]. The axiomatic characterizations of the nonadditivity
Möbius transforms and (max, +)-transform are specified in [29]. index and nonmodularity index are presented in [55,56].
In difference to the 𝑘-additive fuzzy measures, which despite of the
reduction in the number of parameters still require the same number Definition 15. Let 𝑘 ∈ {1, … , 𝑛}. A fuzzy measure 𝜇 on 𝑁 is said to
of linear constraints to satisfy monotonicity, the values 𝜏𝜇 require be 𝑘-nonadditive (𝑘-nonmodular) if its nonadditivity (nonmodularity)
no additional monotonicity constraints, which has advantages when index satisfies 𝑛𝜇 (𝐴) = 0 (𝑑𝜇 (𝐴) = 0) for all 𝐴 ⊆ 𝑁, |𝐴| > 𝑘 and
learning fuzzy measures in that particular representation. there exists at least one subset 𝐴 of 𝑘 elements such that 𝑛𝜇 (𝐴) ≠ 0
By replacing the operation + with any co-norm operation ⊕ (while (𝑑𝜇 (𝐴) ≠ 0).
keeping max unchanged), and − with the corresponding norm operation
⊖, we obtain the (max, ⊕) transform, which exhibits similar properties The 1-nonadditive/nonmodular fuzzy measure is the additive fuzzy
to the (max, +) transform [29]. measure. A 𝑘-nonadditive/nonmodular fuzzy measure 𝜇 on 𝑁 can be
∑ ()
defined by 𝑘𝑖=1 𝑛𝑖 nonadditivity/nonmodularity indices.
3.4. Nonadditivity/nonmodularity index and k-nonadditive/-nonmodular
measures 3.5. Marginal contribution representation and buoyant measures

To explicitly capture the nature and intensity of interactions among The marginal contribution of any criterion 𝑖 to any subset 𝐵 is
the inputs [51–54], nonadditivity and nonmodularity indices have been defined as [26,40]
sequentially introduced [1,18,19]. These indices provide a framework
for quantifying and characterizing different types of interactions. 𝛥𝑖 𝜇(𝐵) = 𝜇(𝐵 ∪ {𝑖}) − 𝜇(𝐵).

These marginal contributions can be visualized using the latter [57,58]


Definition 13 ([18]). The nonadditivity index of subset 𝐴 with respect
and weighted matrix forms [59]. Then the monotonicity condition can
to a fuzzy measure 𝜇 on 𝑁 is given as
be rewritten as
1 ∑
𝑛𝜇 (𝐴) = 𝜇(𝐴) − 𝜇(𝐶). (6)
2|𝐴|−1 − 1 𝐶⊂𝐴 𝛥𝑖 𝜇(𝐵) ≥ 0, ∀𝑖 ∈ 𝑁, 𝐵 ⊆ 𝑁 ⧵ {𝑖}. (10)

4
G. Beliakov et al. Information Fusion 106 (2024) 102295

Fig. 3. A fuzzy measure for 𝑛 = 4 in marginal contribution representation. The new


variables are the edges marked with thick lines (only some are labelled).

Fig. 4. A 3-interactive fuzzy measure for 𝑛 = 6 and the implied partial order. The
sizes of the circles reflect the values of the respective values. Since all 𝜇(𝐴) are the
same for every fixed cardinality |𝐴| = 4, 5, 6, they are treated as equivalent and hence
A fuzzy measure 𝜇 on 𝑁 is submodular (supermodular) if [2]
represented by single nodes.
𝛥𝑖 𝜇(𝐴) ≤ (≥) 𝛥𝑖 𝜇(𝐵), ∀𝐵 ⊂ 𝐴, 𝐴, 𝐵 ⊆ 𝑁. (11)
A fuzzy measure 𝜇 is buoyant (antibuoyant) [37–39] if

𝛥𝑖 (𝐴 ∪ {𝑖, 𝑗}) ≤ (≥) 𝛥𝑗 (𝐴 ∪ {𝑗}), ∀𝑖, 𝑗 ∉ 𝐴. (12) Definition 18 ([43]). Let 𝜇 be a fuzzy measure on 𝑁, a subset 𝐴 ⊆ 𝑁
A buoyant measure is a distinct subclass of submodular measures is a subset of indifference with respect to 𝜇 if ∀𝐵1 , 𝐵2 ⊆ 𝐴, |𝐵1 | = |𝐵2 |,
[39]. That is, any fuzzy measure that is buoyant must necessarily be then 𝜇(𝐶 ∪ 𝐵1 ) = 𝜇(𝐶 ∪ 𝐵2 ), ∀𝐶 ⊆ 𝑁 ⧵ 𝐴.
submodular. However, the converse is not true, i.e., submodularity does
not imply buoyancy. By applying duality, antibuoyant fuzzy measures Definition 19 ([43]). A fuzzy measure 𝜇 on 𝑁 is said to be 𝑝-symmetric
can be characterized as supermodular but not vice versa. The discrete if the coarsest partition of 𝑁 into subsets of indifference contains
derivatives form a decreasing (increasing) sequence along any maximal exactly 𝑝 subsets 𝐴1 , … , 𝐴𝑝 , where 𝐴𝑖 is a subset of indifference, 𝐴𝑖 ∩
chain. Furthermore a fuzzy measure is anti-buoyant if and only if 𝐴𝑗 = ∅, ∪𝑝𝑖=1 𝐴𝑖 = 𝑁, 𝑖, 𝑗 = 1, … , 𝑝, and a partition 𝜋 is coarser than
the respective Choquet integral satisfies the Pigou–Dalton principle of another partition 𝜋 ′ if all subsets of 𝜋 are union of some subsets of 𝜋 ′ .
progressive transfers [38]. The partition {𝐴1 , … , 𝐴𝑝 } is called the basis of 𝜇.

3.6. Special families defined directly on fuzzy measure values 1-symmetric fuzzy measure is just the symmetric fuzzy measure.
Let {𝐴1 , … , 𝐴𝑝 } be the basis of a 𝑝-symmetric fuzzy measure 𝜇 on 𝑁,
Another approach involves directly fixing the values of the fuzzy then any subset 𝑆 ⊆ 𝑁 can be identified with a 𝑝-dimensional vector
measure for all subsets with a cardinality greater than 𝑘, as seen
𝐛𝑆 = (𝑏1 , … , 𝑏𝑝 ), where 𝑏𝑖 = |𝑆 ∩ 𝐴𝑖 |, 𝑖 = 1, … , 𝑝. That is, a 𝑝-symmetric
in the examples of the 𝑘-interactive measure [41] and the 𝑘-tolerant ∏
fuzzy measure needs to define 𝑝𝑖=1 (|𝐴𝑖 | + 1) coefficients [43].
measure [42].
We also briefly mention ways to encode fuzzy measures and their
Definition 16 ([41]). A fuzzy measure is called 𝑘-interactive if for some transformations for computation. The values of a set function can be
chosen 𝐾 ∈ [0, 1] stored in an array of length 2𝑛 on a computer. The most direct scheme
𝑎−𝑘−1 is to use cardinality based ordering, that is, to store first the empty
𝜇(𝐴) = 𝐾 + (1 − 𝐾), for all 𝐴, 𝑎 > 𝑘. set, singletons, pairs, triples, followed by all 𝑘-tuples. This scheme is
𝑛−𝑘−1
In particular, the 𝑘-interactive measure maximizes the average con- usually used for k-order fuzzy measures, such as k-additive, because one
tribution of the 𝑛 − 𝑘 − 1 smallest inputs as well as the partial entropy of does not need to store many zeros (i.e. the Möbius values for subsets of
the calculated over subsets of cardinality greater than 𝑘. The 𝑘-tolerant higher cardinalities). However identifying the subset composition (and
fuzzy measures arise as the special cases (see Fig. 4). hence performing set operations) from the index of a given element in
The 𝑘-interactive fuzzy measures exhibit a notable property: the the array is complicated.
Möbius values of their duals are all zero for subsets of cardinalities Another way to order the fuzzy measure values in an array is to use
ranging from 2 to 𝑛 − 𝑘, provided that 𝑛 − 𝑘 ≥ 2. Therefore they can binary ordering, where the index of an element in binary expresses the
also be represented compactly in Möbius representation through their
set composition (e.g., 01101𝑏 represents {1, 3, 4}). However set functions
duals.
with many zero values (call them sparse fuzzy measures) are not stored
efficiently, although the set operations are very efficient in binary
Definition 17 ([42]). Let 𝑘 ∈ {1, … , 𝑛}. A fuzzy measure 𝜇 on 𝑁 is said
to be 𝑘-tolerant if 𝜇(𝐴) = 1 for all 𝐴 ⊆ 𝑁 such that |𝐴| ≥ 𝑘 and there arithmetic.
exists a subset 𝐵 ⊆ 𝑁, with |𝐵| = 𝑘 − 1, such that 𝜇(𝐵) ≠ 1; Another related scheme is used to store fuzzy measures in marginal
contributions representation [30,40], see Fig. 3. General sparse fuzzy
The dual of 𝑘-tolerant measure is called the 𝑘-intolerant measure,
measures can also be stored in sparse matrix representation [60], where
i.e., if 𝜇(𝐴) = 0 for all 𝐴 ⊆ 𝑁 such that |𝐴| ≤ 𝑛 − 𝑘 and there exists a
only the nonzero values together with their set composition are stored,
subset 𝐵 ⊆ 𝑁, with |𝐵| = 𝑛 − 𝑘 + 1, such that 𝜇(𝐵) ≠ 0.
Indeed, most of the special fuzzy measures mentioned earlier can be which provides efficient subset operations and storage, but at the cost
uniformly termed as lower/upper 𝑘-order representative measures [36], of extra data to keep.
∑ ()
each defined by 𝑘𝑖=1 𝑛𝑖 coefficients. Some operations on fuzzy measures, such as transformations and
Another special type of special fuzzy measure is proposed from differentiation, can be expressed in sparse linear formalism as presented
indifference sets. in [1,31,61].

5
G. Beliakov et al. Information Fusion 106 (2024) 102295

4. Optimization models and machine learning for fuzzy measures classification [111], sparse modelling [112], image quality thresh-
olding [113], tuberculosis detection from chest X-ray images, gene
4.1. Fuzzy measure fitting with nonlinear optimization models selection, and cancer prediction [107,110], edge detection [114], clus-
ter analysis [115], facial emotion recognition [108], brain–computer
Fuzzy measure identification approaches are typically built upon interface, social network analysis [116], as well as cancer histology
optimization models featuring nonlinear objective functions alongside classification [109].
explicit and implicit preference constraints [34,62]. These models
encompass various methodologies, including least-squares-based ap- 4.5. Non-monotonic measure learning
proaches [24,34], maximum split approaches [63], TOMASO (Tool
for Ordinal Multi-Attribute Sorting and Ordering) approaches [64, Non-monotonic measures are normalized set functions without
65], the least absolute deviation criterion-based linear programming monotonicity constraints, which offer greater flexibility than regu-
model [66], nonadditive robust ordinal regression models [67–73], lar fuzzy measures in representing intricate and chaotic situations.
minimum mean absolute error models [74], maximum log-likelihood Refs. [14,117] provide the axiomatic characterization of non-monotonic
principle-based optimization models [75,76], maximum margin measures in conjunction with Choquet integral and the Shapley index.
principle-based optimization models [77], the maximum entropy prin- This framework is extended to regression tasks, showcasing improved
ciple [78], the compromise principle [79], least square and absolute de- performance, swift computation, and strong interpretability [118].
viation principles based on MCCPI (Multiple Criteria Correlation Prefer- Additionally, it proves effective for autoregressive models, facilitating
ence Information) [80,81], interaction index-oriented models [62,73], efficient time series analysis that incorporates nonlinear features [119].
OWA vector and probability vector based crescent method [82–84], Furthermore, the non-monotonic measure-based Choquet integral has
basic uncertain information based generation methods [85,86], param- been extended to include interval values [120,121], set values [122,
eterized construction and transformation methods [87,88], weighting 123], and fuzzy valued [124–126] forms by using the Möbius repre-
models [89], preference levelled evaluation functions based construc- sentation [45].
tion methods [90], and tolerance perspective determining method [91].
4.6. Other special optimization tasks

4.2. Linear representation of optimization models In the context of portfolio selection, the knapsack problem inher-
ently addresses interdependencies, often synergies, among items. This
Linear programming offers advantages over nonlinear program- can be effectively formulated through the utilization of the Choquet
ming due to its computational efficiency in finding global optima, integral with a supermodular measure, ultimately resulting in the for-
well-established methods, algorithmic maturity, interpretability, and mulation of linear mixed-integer programs. For large-scale knapsack
straightforward sensitivity analysis [37]. Some nonlinear objectives can problems, the scalability and interaction representation capabilities of
be replaced by linear ones. For instance, the least square form can be a 2-additive measure make it a suitable choice [127].
alternatively represented by the least absolute deviation and further The Difference of Convex Functions Algorithm (DCA) is an op-
into multiple goal linear programming [66,80]. Nonlinear constraints timization technique designed for problems where the objective is
on special fuzzy measures can also be converted into linear constraints. expressed as the difference between two convex or concave functions.
For example, the 𝑘-maxitive measure defined in (3) has equivalent Through iterative steps, it minimizes these components, progressively
representations in the following 0-1 linear constraints [47,92]: converging to solutions for non-convex problems by tackling simpler
convex subproblems [94]. This approach can be employed effectively
⎧𝜇(𝐴) − 𝜇(𝐴∖{𝑖}) ≤ 𝑦 , |𝐴| ≥ 𝑘 + 1, 𝑖 ∈ 𝐴,
⎪ { 𝐴,𝑖 in the optimization of Sugeno and Choquet integrals, leveraging the
⎪ 0 if 𝜇(𝐴) − 𝜇(𝐴∖{𝑖}) = 0 insight that any fuzzy measure can be represented as a difference
⎨𝑦𝐴,𝑖 = between two supermodular fuzzy measures [2,17]. By appropriately de-
⎪ 1 if 𝜇(𝐴) − 𝜇(𝐴∖{𝑖}) ≠ 0
⎪∑ 𝑦 ≤ |𝐴| − 1. composing the objective function into convex or concave components,
⎩ 𝐴 𝐴,𝑖
the optimization of these integrals can be formulated [57,93,94,128,
Choquet integral and Sugeno integral based objective function both 129]. Moreover, the 𝑘-order measure and Möbius representation offer
can be rewritten in linear goal constraints and multiple goal liner suitable simplifications for managing the complexity and sparsity of
objectives [30,38,93,94]. large-scale problems (see Table 2).
(max,+)-transforms do not require any monotonicity constraints
4.3. Inconsistency checking and adjustment and are used successfully in genetic algorithms for fuzzy measure
identification [29,49].
The inconsistency check and adjustment approaches of the correla-
tive preference information can be classified into two major branches. 5. Random generation methods for fuzzy measures
One branch uses the 0-1 mixed linear programming to find the con-
tradictory constraints and then directly remove them from the set 5.1. Challenges for generating fuzzy measures
without further programmable process of adjustment [68,69,71]. An-
other branch uses the multiple goals linear programming (MGLP) to Random generation of fuzzy measures plays a pivotal role in solving
recognize the contradictory as well as the redundant constraints and the previously mentioned optimization models by introducing stochas-
then provide some adjustment suggestions [95–98]. tic elements into the problem-solving process. It facilitates exploration
of the solution space, aiding the algorithm in escaping local optima and
4.4. Fuzzy measure based machine and deep learning diversifying its search for better solutions and empowers optimization
algorithms to address uncertainty, complex constraints, and nonlinear-
Fuzzy measure-based Choquet and Sugeno integral models find ity effectively [134]. Methods based on evolutionary heuristics depend
applications in machine and deep learning models, serving as aggre- heavily on generation random solutions in a feasible domain.
gation functions for combining and ensembling classifiers [99–103]. It turns out that random generation of fuzzy measures with uniform
These models are also used to fuse and regularize outputs from mod- distribution is a very challenging problem. The difficulties are related
ular artificial neural networks (ANN) or convolutional neural net- to the high dimensionality of the space containing fuzzy measures and
works (CNN) [104–110], with applications spanning highly imbalanced a complicated structure of the set of fuzzy measures. It is known [135]

6
G. Beliakov et al. Information Fusion 106 (2024) 102295

Table 2
Optimization and machine learning for fuzzy measures.
Topic Content References
Fuzzy measure fitting Implicit and explicit information based fitting and [1,19,34,80]
learning models
Linear represented Linear expressions of special families of fuzzy [38,47,92–94]
optimization models measures, Choquet integral and Sugeno integral
objective function oriented optimization models
Inconsistency 0-1 mixed linear programming and multiple goals [68,69,71,95–98].
checking and linear programming (MGLP)
adjustment
Least squares Quadratic expressions for special families of fuzzy [34,76,99,130–133]
optimization models measures, Choquet integral objective function
oriented optimization models
Fuzzy measure based Combining outputs and classifiers ensembles, ANNs [105–110]
machine and deep and CNNs
learning
Non-monotonic Offer much flexibility without monotonicity [14,117–119]
measure learning constraints, easier to fit, but lack of any
monotonicity may be questionable in applications

that the set of all fuzzy measures is a polytope with an extremely large 5.3. Vertex–convex combination generation method
number of vertices related to the Dedekind numbers 𝑀(𝑛) (the OEIS
sequence A000372: 2, 3, 6,20, 168, 7581, 7828354, 2414682040998, The feasible range determined by a set of linear constraints is a
…), and hence it is infeasible to use convex combinations of its vertices, convex set, and hence convex combinations of feasible solutions will
for example, for random sampling. General methods of uniformly sam- also be feasible. The convex combination of fuzzy measures 𝜇1 , … , 𝜇𝑡 ,
pling polytopes [136,137] are also infeasible in such a high dimension. 𝑡 > 1, is defined as [138]:
On the other hand, the structure of the polytopes of fuzzy measures

𝑡
𝑃 , called the order polytopes, is established [2,135,138]. It allows one 𝜆1 𝜇1 + ⋯ + 𝜆𝑡 𝜇𝑡 , 𝜆𝑖 = 1, 𝜆𝑖 ≥ 0. (13)
to perform decomposition of 𝑃 into simplices (of equal volume) [139], 𝑖=1
then randomly pick up one simplex and finally generate a random point This formula allows any number of feasible fuzzy measures to serve
within that simplex (the last step is almost trivial [137,140]). The catch as the parent measures for generating new fuzzy measures between
here is the extremely large cardinality of the simplicial partition, given them. However, one of the ultimate goals of a random generation
by the OEIS sequence A046873, which reads (starting from 𝑛 = 1) 1, method is to cover the entire feasible range as much as possible. In
2, 48, 1680384, 14807804035657359360,. . . . Hence these simplices this context, the use of extreme fuzzy measures [138], which usually
cannot be explicitly enumerated for 𝑛 > 4. correspond to the boundary or vertex points of the polytope of fuzzy
Recent research has introduced efficient techniques for randomly measures, can provide a better coverage of the feasible domain [143].
generating fuzzy measures and their specific types, including random If one restricts herself only to the boundary and monotonicity con-
node generation [135,141], topological sort generation [134], linear ditions, the extreme values of 𝜇(𝐴) can reach 0 and 1 in two opposite
extensions based generation [135,139,142], vertex–convex combina- directions, and one basic fact is that any fuzzy measure is a convex
tion generation [138,143]. combination of 0-1 valued extreme fuzzy measures [2]. The number of
It should be noted that random generation of fuzzy measures can be extreme fuzzy measures that take only 0 or 1 as their values is given
performed in any representation related to 𝜇 linearly, such as Möbius, by 𝑀(𝑛) − 2, where 𝑀(𝑛) is a Dedekind number that grows very quickly
interaction, non-additivity and non-modularity, mentioned in Section 3. with 𝑛.
It is due to the fact that a uniform distribution remains uniform under
a linear transformation, hence some approaches use Möbius represen- 5.4. Linear extension based generation method
tation, for instance. However order polytopes in one representation are
no longer order polytopes in another, and hence specialized methods The set of all fuzzy measures, which is an order polytope 𝑃 , is
are not equally applicable in different representations. determined by the 𝑑 = 2𝑛 − 2 non-negative values 𝜇(𝐴) and the 𝑛2𝑛−1
non-redundant monotonicity constraints. Note that the values 𝜇(𝐴)
5.2. Random node generation method constitute a poset, which corresponds to the Boolean lattice 𝐵𝑛 . It is
known that this polytope has simplicial partition, and each simplex
A straightforward approach to generating a random fuzzy measure corresponds to a linear extension of (𝑁). A linear extension of a poset
on set 𝑁 is to randomly select a subset (node) from 𝑁 first, and  is a linear order (i.e. chain) compatible with the order relations in .
then to identify a random value within its allowable range [138,141]. The number of linear extensions of the Boolean lattice 𝐵𝑛 is given
The most notable advantage of this method lies in its low complexity by the OEIS sequence A046873, which makes it prohibitively expensive
and fast execution speed. However, from a theoretical standpoint, to construct the partition explicitly. Some (approximately) random
the fuzzy measures it generates are not uniform. This non-uniformity algorithms for picking linear extensions have been proposed [145].
stems from two factors: the dependence of a subset’s allowable range The MinimalsPlus approach [139] employs random walks to select
on its cardinality and rank, and the distribution of generated values simplices in a non-enumerative manner, making it suitable for both
deviating significantly from a uniform distribution [144]. Moreover, general fuzzy measures and 𝑝-symmetric fuzzy measures. For 2-additive
efforts have been made to enhance random node generation by lever- fuzzy measures, a method based on the simplicial partition in Möbius
aging distribution information of measure values and probabilities representation has been developed [22]. An alternative technique uti-
of linear extensions, particularly under monotonicity and additional lizes Markov chains to approximate linear extensions, resulting in a
conditions [144]. two-layer approximation that significantly reduces time consumption

7
G. Beliakov et al. Information Fusion 106 (2024) 102295

Table 3
Fuzzy measure generation methods.
Methods Basic principle References Features
Random node Randomly select a subset (node) from N [135,141] Simple to
generation first and identify a random value within implement but does
its allowable range not produce uniform
distribution
Topological sort Considers poset  as a directed acyclic [134,147] Efficient, produces
generation graph (DAG) where an edge between linear extensions
vertices 𝑢 and 𝑣 means 𝑢 ≤ 𝑣 in that distributed
ordering uniformly
Linear extensions View the random generation of fuzzy [22,135,139,145] Produces linear
based generation measures as a process of assigning extensions
ordered random values to the linear distributed
extensions of all subsets in (𝑁). uniformly, needs
many Markov steps
Vertex–convex The convex combination of boundary or [37,138] Better coverage but
combination vertex measures can provide a better for uniformity needs
generation coverage of the feasible domain. additional steps

compared to traditional Markov chain methods [145]. Another way of subsequently huge computational complexity, which involves not only
picking a linear extension of a poset is by topological sorting [146] the parameters but relations among them, such as monotonicity or
which considers poset  as a directed acyclic graph (DAG) where an convexity. Modern practical applications produce large universal sets,
edge between vertices 𝑢 and 𝑣 means 𝑢 ≤ 𝑣 in that ordering. The oldest which imply the need to develop suitable representation, simplification,
topological sort method is due to Kahn [146], which is implemented characterization, eliciting, computation, and other tools.
in most unix-based computer systems, as a useful routine to perform In this survey we reviewed three aspects related to modelling with
objects linking. A different algorithm is based on depth-first search from fuzzy measures: representations, learning and random generation. An
the nodes of the DAG in arbitrary order [147]. important issue is fuzzy measure simplification to reduce their com-
One additional simple technique to improve the uniformity of the plexity, both in terms of computational cost and interpretability [149].
randomly generated measures is symmetrization. Note that we expect The latter is particularly useful for eXplainable Artificial Intelligence
the distribution to be independent on the way the inputs are labelled. (XAI), where high predictive power is delivered side by side with
Whatever is the permutation of the indices of the inputs {1, 2, … , 𝑛}, explainability and interpretability of parameters [150].
the distribution must be the same. The symmetrization technique picks Promising new directions related to fuzzy measure representations
an arbitrary permutation 𝜎 and then relabels the inputs 𝑖 → 𝜎(𝑖) in are sparse fuzzy measures and hierarchical decompositions of the fuzzy
every subset 𝐴 ⊂ 𝑁, thus making the output of any random generation integrals [37,103,149]. Sparse representations disregard some of the
algorithm nonsensitive to the way the inputs are labelled. less important inputs’ interactions and focus on the most significant
ones. Methods of identification and learning sparse fuzzy measures are
5.5. Generation of special measures needed. Hierarchical decompositions also lead to sparse representations
but in a structured and top-down driven way, mimicking the approach
A method for generating supermodular measures has been proposed taken by people to decompose complex problems into simpler blocks.
in [148], which used MinimalsPlus, randomized topological sort and Scaling up fuzzy measure-based models to larger universes is an-
Markov chain random walk, combined with a smaller set of linear con- other direction which aims at broadening their applicability. Big data
straints and amenable to acceptance/rejection. This approach translates leads to big fusion problems that at the moment are beyond the reach of
into an efficient algorithm for generating supermodular capacities, and fuzzy integrals. Suitable simplifications and decompositions, as well as
by duality, also the submodular ones. The implementation has been sparse modelling must be combined with highly efficient computational
released as part of fmtools package (in C/C++ with interfaces to R and tools and high performance computing.
Python) [60]. Another promising direction is fuzzy measure-based optimization
Relying on the fact that convex combinations of (anti)buoyant models. One aspect of it is aggregation in multiobjective optimization
fuzzy measures remain within the same class, Ref. [37] constructed (by reducing multiobjective problems to single-objective). It involves
various subsets of vertices for (anti)buoyant fuzzy measures based on dynamic data-driven approximations to the Pareto fronts, which ac-
their desired cardinalities and linear extensions. These subsets were counts for mutual dependencies of the objectives. Such problems in-
further utilized to create possible representations of (anti)buoyant fuzzy volve repeated learning and adjustments of fuzzy measures within an
measures by employing their convex combinations. optimization algorithm, as well as their random generation.
In [147], the MinimalsPlus and topological sort methods were em- Another aspect is formulating and executing optimization models
ployed to generate 𝑘-interactive capacities, and a random walk (Markov with the objectives and constraints based on fuzzy integrals. Linear
chain) was used on a modified Boolean lattice corresponding to 𝑘- objectives which underpin the area of linear programming assume
interactive capacities to ensure more uniform distribution of outputs independence of the variables, which is often not justified. Choquet
(see Table 3). integral-based objectives generalize linear programming, and for spe-
cific types of interactions (e.g., variables synergy) can be transformed
6. Further directions and conclusion into larger linear programming problems. When the variables interac-
tions are mixed, alternative formulations are being developed [128].
Aggregation methods in which only the individual inputs are weigh- In the same vein, mixed linear integer problems, such as knapsack, can
ted implicitly assume inputs independence. In reality a large number of be generalized by using fuzzy integrals, and sometimes converted into
practical problems involve inputs synergies and redundancies. Fuzzy larger mixed linear integer problems [127].
measures quantify possible inputs interaction by assigning values to Overall, aggregation based on fuzzy measures facilitates develop-
all of inputs’ subsets. The main challenge of this modelling approach ment of sophisticated models which account for inputs interactions,
is the high number of inputs coalitions (for 𝑛 inputs it is 2𝑛 ), and but at the cost of high computational complexity. This challenge can

8
G. Beliakov et al. Information Fusion 106 (2024) 102295

be overcome through fuzzy measure simplifications, hierarchical and [21] M. Grabisch, J.-L. Marichal, R. Mesiar, E. Pap, Aggregation Functions, in:
sparse representation, and the development of efficient computational Encyclopedia of Mathematics and Its Foundations, Cambridge University Press,
2009.
tools. The potential gain is in better interpretability of the parameters
[22] P. Miranda, P. García-Segador, Combinatorial structure of the polytope of 2-
of such models and predictability of their behaviour, something that additive measures, IEEE Trans. Fuzzy Syst. http://dx.doi.org/10.1109/TFUZZ.
current black-box based models are missing. 2019.2945243.
[23] M. Grabisch, Fuzzy integral in multicriteria decision making, Fuzzy Sets and
CRediT authorship contribution statement Systems 69 (3) (1995) 279–298.
[24] M. Grabisch, C. Labreuche, A decade of application of the Choquet and Sugeno
integrals in multi-criteria decision aid, Ann. Oper. Res. 175 (1) (2010) 247–286.
Gleb Beliakov: Conceptualization, Formal analysis, Writing – origi- [25] A. Chateauneuf, J.-Y. Jaffray, Some characterizations of lower probabilities and
nal draft, Writing – review & editing. Jian-Zhang Wu: Formal analysis, other monotone capacities through the use of Möbius inversion, Math. Social
Investigation, Methodology, Resources, Writing – original draft, Writ- Sci. 17 (3) (1989) 263–283.
[26] M. Grabisch, J.-L. Marichal, M. Roubens, Equivalent representations of set
ing – review & editing. Weiping Ding: Conceptualization, Methodol-
functions, Math. Oper. Res. 25 (2) (2000) 157–178.
ogy, Writing – original draft, Writing – review & editing. [27] R. Mesiar, Generalizations of k-order additive discrete fuzzy measures, Fuzzy
Sets and Systems 102 (3) (1999) 423–428.
Declaration of competing interest [28] J.-Z. Wu, G. Beliakov, Probabilistic bipartition interaction index of multiple
decision criteria associated with the nonadditivity of fuzzy measures, Int. J.
Intell. Syst. 34 (2) (2019) 247–270.
The authors declare the following financial interests/personal rela-
[29] V. Torra, (Max, ⊕)-transforms and genetic algorithms for fuzzy measure
tionships which may be considered as potential competing interests: identification, Fuzzy Sets and Systems 451 (2022) 253–265.
Beliakov reports financial support was provided by Australian Research [30] G. Beliakov, D. Divakov, On representation of fuzzy measures for learning
Council. If there are other authors, they declare that they have no choquet and sugeno integrals, Knowl.-Based Syst. 189 (2020) 105134.
known competing financial interests or personal relationships that [31] R.-J. Xi, J.-Z. Wu, G. Beliakov, Matrix representation of capacity-based
multicriteria decision analysis, IEEE Access 7 (2019) 185543–185553.
could have appeared to influence the work reported in this paper. [32] K. Fujimoto, I. Kojadinovic, J.-L. Marichal, Axiomatic characterizations of
probabilistic and cardinal-probabilistic interaction indices, Games Econom.
Data availability Behav. 55 (1) (2006) 72–99.
[33] M. Grabisch, K-order additive discrete fuzzy measures and their representation,
Fuzzy Sets and Systems 92 (2) (1997) 167–189.
No data was used for the research described in the article.
[34] M. Grabisch, I. Kojadinovic, P. Meyer, A review of methods for capacity iden-
tification in Choquet integral based multi-attribute utility theory: Applications
Acknowledgement of the Kappalab R package, European J. Oper. Res. 186 (2) (2008) 766–785.
[35] R. Mesiar, A. Kolesárová, k-maxitive aggregation functions, Fuzzy Sets and
The work was supported by the Australian Research Council Discov- Systems 346 (2018) 127–137.
[36] J.-Z. Wu, G. Beliakov, k-order representative capacity, J. Intell. Fuzzy Systems
ery project DP210100227.
38 (3) (2020) 3105–3115.
[37] G. Beliakov, M. Gagolewski, S. James, Reduction of variables and constraints
References in fitting antibuoyant fuzzy measures to data using linear programming, Fuzzy
Sets and Systems 451 (2022) 266–284.
[1] G. Beliakov, S. James, J.-Z. Wu, Discrete Fuzzy Measures: Computational [38] G. Beliakov, S. James, Choquet integral optimisation with constraints and the
Aspects, Springer, Cham, Switzerland, 2020. buoyancy property for fuzzy measures, Inform. Sci. 578 (2021) 22–36.
[2] M. Grabisch, Set Functions, Games and Capacities in Decision Making, Springer, [39] G. Beliakov, S. James, Choquet integral-based measures of economic welfare
Berlin, New York, 2016. and species diversity, Int. J. Intell. Syst. 37 (4) (2022) 2849–2867.
[3] B. Peleg, P. Sudholter, Introduction to the Theory of Cooperative Games, [40] J.-Z. Wu, G. Beliakov, Marginal contribution representation of capacity-based
Springer, Berlin, 2007. multicriteria decision making, Int. J. Intell. Syst. 35 (3) (2020) 373–400.
[4] G. Chalkiadakis, E. Elkind, M. Wooldridge, Computational Aspects of Co- [41] G. Beliakov, J.-Z. Wu, Learning fuzzy measures from data: simplifications and
operative Game Theory, Morgan and Claypool Publishers, San Rafael, USA, optimisation strategies, Inform. Sci. 494 (2019) 100–113.
2012. [42] J.-L. Marichal, K-intolerant capacities and choquet integrals, European J. Oper.
[5] M. Sugeno, Theory of Fuzzy Integrals and Its Applications (Ph.D. thesis), Tokyo Res. 177 (3) (2007) 1453–1468.
Institute of Technology, 1974. [43] P. Miranda, M. Grabisch, P. Gil, p-Symmetric fuzzy measures, Int. J. Uncertain.
[6] G. Choquet, Theory of capacities, Ann. l’inst. Fourier 5 (1954) 131–295. Fuzziness Knowl.-Based Syst. 10 (supp01) (2002) 105–123.
[7] E. Pap, Null-Additive Set Functions, Kluwer Academic Pub, Dordrecht, 1995. [44] G. Beliakov, H. Bustince Sola, T. Calvo, A Practical Guide to Averaging
[8] Z. Wang, G.J. Klir, Generalized Measure Theory, Springer Science & Business Functions, Springer, New York, 2016.
Media, New York, 2010. [45] J. Fernandez, H. Bustince, L. Horanská, R. Mesiar, A. Stupňanová, A general-
[9] Z. Wang, G.J. Klir, Fuzzy Measure Theory, Springer Science & Business Media, ization of the Choquet integral defined in terms of the Möbius transform, IEEE
2013. Trans. Fuzzy Syst. 28 (10) (2019) 2313–2319.
[10] M. Grabisch, T. Murofushi, M. Sugeno, J. Kacprzyk, Fuzzy Measures and [46] G. Pereira Dimuro, J. Fernandez, B. Bedregal, R. Mesiar, J. Sanz, G. Lucca, H.
Integrals: Theory and Applications, Springer, 2000. Bustince, The state-of-art of the generalizations of the Choquet integral: From
[11] D. Denneberg, Non-Additive Measure and Integral, Springer Science & Business aggregation and pre-aggregation to ordered directionally monotone functions,
Media, Dordrecht, 1994. Inf. Fusion 57 (2020) 27–43.
[12] V. Torra, Y. Narukawa, M. Sugeno, Non-Additive Measures: Theory and [47] J.-Z. Wu, G. Beliakov, k-minitive capacities and k-minitive aggregation
Applications, vol. 310, Springer, 2013. functions, J. Intell. Fuzzy Systems 37 (2) (2019) 2797–2808.
[13] M. Grabisch, The representation of importance and interaction of features by [48] L.A. Zadeh, Fuzzy sets as a basis for a theory of possibility, Fuzzy Sets and
fuzzy measures, Pattern Recognit. Lett. 17 (6) (1996) 567–575. Systems 100 (1999) 9–34.
[14] T. Murofushi, M. Sugeno, M. Machida, Non-monotonic fuzzy measures and the [49] V. Torra, Optimal transport and the wasserstein distance for fuzzy measures:
Choquet integral, Fuzzy Sets Syst. 64 (1) (1994) 73–86. An example, in: International Conference on Intelligent and Fuzzy Systems,
[15] Y. Narukawa, T. Murofushi, M. Sugeno, Space of fuzzy measures and Springer, 2023, pp. 39–44.
convergence, Fuzzy Sets and Systems 138 (3) (2003) 497–506. [50] V. Torra, The transport problem for non-additive measures, European J. Oper.
[16] L. Garmendia, The evolution of the concept of fuzzy measure, in: Intelligent Res. (2023).
Data Mining: Techniques and Applications, Springer, 2005, pp. 185–200. [51] I. Kojadinovic, Modeling interaction phenomena using fuzzy measures: on the
[17] I. Gilboa, D. Schmeidler, Additive representations of non-additive measures and notions of interaction and independence, Fuzzy Sets Syst. 135 (3) (2003)
the choquet integral, Ann. Oper. Res. 52 (1994) 43–65. 317–340.
[18] J.-Z. Wu, G. Beliakov, Nonadditivity index and capacity identification method in [52] J.-L. Marichal, Aggregation of interacting criteria by means of the discrete
the context of multicriteria decision making, Inform. Sci. 467 (2018) 398–406. choquet integra, in: T. Calvo, G. Mayor, R. Mesiar (Eds.), Aggregation Operators:
[19] J.-Z. Wu, G. Beliakov, Nonmodularity index for capacity identifying with New Trends and Applications, Springer, 2002, pp. 224–244.
multiple criteria preference information, Inform. Sci. 492 (2019) 164–180. [53] J.-Z. Wu, L.-P. Yu, G. Li, J. Jin, B. Du, The sum interaction indices of some
[20] G. Beliakov, A. Pradera, T. Calvo, Aggregation Functions: A Guide for particular families of monotone measures, J. Intell. Fuzzy Systems 31 (3) (2016)
Practitioners, Springer, Berlin, Heidelberg, 2007. 1447–1457.

9
G. Beliakov et al. Information Fusion 106 (2024) 102295

[54] J.-Z. Wu, L.-P. Yu, G. Li, J. Jin, B. Du, Using the monotone measure sum to [83] B. Llamazares, On the relationship between the Crescent Method and SUOWA
enrich the measurement of the interaction of multiple decision criteria, J. Intell. operators, IEEE Trans. Fuzzy Syst. 28 (10) (2019) 2645–2650.
Fuzzy Systems 30 (5) (2016) 2529–2539. [84] L. Jin, R. Mesiar, R.R. Yager, The properties of crescent preference vectors
[55] G. Beliakov, J.-Z. Wu, The axiomatic characterisations of non-modularity index, and their utility in decision making with risk and preferences, Fuzzy Sets and
Int. J. Gen. Syst. 49 (2020) 675–688. Systems 409 (2021) 114–127.
[56] J.-Z. Wu, G. Beliakov, Axiomatic representations for nonadditivity and non- [85] L. Jin, R. Mesiar, S. Borkotokey, M. Kalina, Certainty aggregation and the
modularity indices: describing interactions of fuzzy measures, J. Intell. Fuzzy certainty fuzzy measures, Int. J. Intell. Syst. 33 (4) (2018) 759–770.
Systems (2023) in press. [86] L. Jin, R. Mesiar, R.R. Yager, Some fuzzy measures constructing paradigm and
[57] G. Beliakov, D. Divakov, Aggregation with dependencies: Capacities and fuzzy methods from given measure spaces and evaluation information, IEEE Trans.
integrals, Fuzzy Sets and Systems 446 (2022) 222–232. Fuzzy Syst. 30 (3) (2020) 878–884.
[58] J.-Z. Wu, G. Beliakov, Some ideas about graphic representations of discrete [87] L. Jin, R. Mesiar, R.R. Yager, Derived fuzzy measures and derived Cho-
fuzzy measures, 2023, arXiv preprint arXiv:2310.06871. quet integrals with some properties, IEEE Trans. Fuzzy Syst. 29 (5) (2020)
[59] A.R. Buck, D.T. Anderson, J.M. Keller, T. Wilkin, M.A. Islam, A weighted matrix 1320–1324.
visualization for fuzzy measures and integrals, in: 2020 IEEE International [88] R. Mesiar, S. Borkotokey, L. Jin, M. Kalina, Aggregation functions and
Conference on Fuzzy Systems, FUZZ-IEEE, IEEE, 2020, pp. 1–8. capacities, Fuzzy Sets and Systems 346 (2018) 138–146.
[60] G. Beliakov, Rfmtool package, version 4, 2018-2022, https://CRAN.R-project.
[89] L. Jin, R. Mesiar, G. Qian, Weighting models to generate weights and capacities
org/package=Rfmtool, http://personal-sites.deakin.edu.au/~gleb/fmtools.html.
in multicriteria group decision making, IEEE Trans. Fuzzy Syst. 26 (4) (2017)
[61] G. Beliakov, On the derivatives of set functions in matrix representation, Inform.
2225–2236.
Sci. 546 (2021) 469–480.
[90] C. Zhu, L. Jin, R. Mesiar, R.R. Yager, Using preference leveled evaluation
[62] J.-Z. Wu, E. Pap, A. Szakal, Two kinds of explicit preference information
functions to construct fuzzy measures in decision making and evaluation, Int.
oriented capacity identification methods in the context of multicriteria decision
J. Gen. Syst. 49 (2) (2020) 161–173.
analysis, Int. Trans. Oper. Res. 25 (2018) 807–830.
[91] J. Li, X. Yao, X. Sun, D. Wu, Determining the fuzzy measures in multiple criteria
[63] J.-L. Marichal, M. Roubens, Determination of weights of interacting criteria
decision aiding from the tolerance perspective, European J. Oper. Res. 264 (2)
from a reference set, European J. Oper. Res. 124 (3) (2000) 641–650.
(2018) 428–439.
[64] P. Meyer, M. Roubens, Choice, ranking and sorting in fuzzy multiple criteria
decision aid, in: J. Figueira, S. Greco, M. Ehrogott (Eds.), Multiple Criteria [92] G. Beliakov, J.-Z. Wu, Learning k-maxitive fuzzy measures from data by mixed
Decision Analysis: State of the Art Surveys, Springer New York, 2005, pp. integer programming, Fuzzy Sets and Systems 412 (2021) 41–52.
471–503. [93] G. Beliakov, M. Gagolewski, S. James, Aggregation on ordinal scales with the
[65] M. Roubens, Ordinal multiattribute sorting and ordering in the presence of sugeno integral for biomedical applications, Inform. Sci. 501 (2019) 377–387.
interacting points of view, in: Aiding Decisions with Multiple Criteria: Essays [94] G. Beliakov, M. Gagolewski, S. James, DC optimization for constructing discrete
in Honor of Bernard Roy, Springer, 2002, pp. 229–246. Sugeno integrals and learning nonadditive measures, Optimization 69 (12)
[66] G. Beliakov, Construction of aggregation functions from data using linear (2020) 2515–2534.
programming, Fuzzy Sets and Systems 160 (1) (2009) 65–75. [95] J.-Z. Wu, L. Huang, R.-J. Xi, Y.-P. Zhou, Multiple goal linear programming-
[67] S. Angilella, M. Bottero, S. Corrente, V. Ferretti, S. Greco, I.M. Lami, Non based decision preference inconsistency recognition and adjustment strategies,
Additive Robust Ordinal Regression for urban and territorial planning: an Information 10 (7) (2019) 223.
application for siting an urban waste landfill, Ann. Oper. Res. 245 (1–2) (2016) [96] J.-Z. Wu, R.-J. Xi, Y. Zhu, Correlative decision preference information consis-
427–456. tency check and comprehensive dominance representation method, J. Intell.
[68] S. Angilella, S. Greco, B. Matarazzo, Non-additive robust ordinal regression: A Fuzzy Systems 38 (2) (2020) 2009–2019.
multiple criteria decision model based on the Choquet integral, European J. [97] G. Beliakov, J.-Z. Wu, D. Divakov, Towards sophisticated decision models:
Oper. Res. 201 (1) (2010) 277–288. Nonadditive robust ordinal regression for preference modeling, Knowl.-Based
[69] S. Corrente, S. Greco, M. Kadziński, R. Słowiński, Robust ordinal regression in Syst. 190 (2020) 105351.
preference learning and ranking, Mach. Learn. 93 (2–3) (2013) 381–422. [98] J.-Z. Wu, F.-F. Chen, Y.-Q. Li, L. Huang, Capacity random forest for correlative
[70] S. Corrente, S. Greco, A. Ishizaka, Combining analytical hierarchy process multiple criteria decision pattern learning, Mathematics 8 (8) (2020) 1372.
and choquet integral within non-additive robust ordinal regression, Omega 61 [99] M. Uriz, D. Paternain, H. Bustince, M. Galar, A supervised fuzzy measure
(2016) 2–18. learning algorithm for combining classifiers, Inform. Sci. 622 (2023) 490–511.
[71] S. Greco, V. Mousseau, R. Slowinski, Ordinal regression revisited: Multiple [100] M. Uriz, D. Paternain, I. Dominguez-Catena, H. Bustince, M. Galar, Unsupervised
criteria ranking using a set of additive value functions, European J. Oper. Res. fuzzy measure learning for classifier ensembles from coalitions performance,
191 (2) (2008) 416–436. IEEE Access 8 (2020) 52288–52305.
[72] S. Greco, V. Mousseau, R. Słowiński, Robust ordinal regression for value [101] S. Abbaszadeh, E. Hüllermeier, Machine learning with the sugeno integral: The
functions handling interacting criteria, European J. Oper. Res. 239 (3) (2014) case of binary classification, IEEE Trans. Fuzzy Syst. 29 (12) (2020) 3723–3733.
711–730.
[102] G.J. Scott, R.A. Marcum, C.H. Davis, T.W. Nivin, Fusion of deep convolutional
[73] J.-Z. Wu, G. Beliakov, Nonadditive robust ordinal regression with nonadditivity
neural networks for land cover classification of high-resolution imagery, IEEE
index and multiple goal linear programming, Int. J. Intell. Syst. 34 (2019)
Geosci. Remote Sens. Lett. 14 (9) (2017) 1638–1642.
1732–1752.
[103] C.-Y. Chen, J.-J. Huang, Integration of genetic algorithms and neural networks
[74] M. Gagolewski, S. James, G. Beliakov, Supervised learning to aggregate data
for the formation of the classifier of the hierarchical Choquet integral, Inform.
with the Sugeno integral, IEEE Trans. Fuzzy Syst. 27 (4) (2019) 810–815.
Sci. 537 (2020) 46–61.
[75] A. Fallah Tehrani, W. Cheng, E. Hüllermeier, Preference learning using the
[104] S.-L. Wu, Y.-T. Liu, T.-Y. Hsieh, Y.-Y. Lin, C.-Y. Chen, C.-H. Chuang, C.-T.
Choquet integral: The case of multipartite ranking, in: Proceedings of the 20th
Lin, Fuzzy integral with particle swarm optimization for a motor-imagery-based
Workshop Computational Intelligence, 2010, pp. 119–130.
brain–computer interface, IEEE Trans. Fuzzy Syst. 25 (1) (2016) 21–28.
[76] E. Hüllermeier, A.F. Tehrani, Efficient learning of classifiers based on the 2-
additive choquet integral, in: C. Moewes, A. Nürnberger (Eds.), Computational [105] G. Martinez, P. Melin, O. Castillo, Optimization of modular neural networks
Intelligence in Intelligent Data Analysis, Springer Berlin Heidelberg, 2013, pp. using hierarchical genetic algorithms applied to speech recognition, in: Pro-
17–29. ceedings. 2005 IEEE International Joint Conference on Neural Networks, 2005,
[77] A. Fallah Tehrani, W. Cheng, K. Dembczynski, E. Hüllermeier, Learning mono- Vol. 3, IEEE, 2005, pp. 1400–1405.
tone nonlinear models using the Choquet integral, Mach. Learn. 89 (1–2) (2012) [106] P. Melin, A. Mancilla, M. Lopez, J. Soria, O. Castillo, Pattern recognition for
183–211. industrial monitoring and security using the fuzzy Sugeno integral and modular
[78] J.-L. Marichal, Entropy of discrete choquet capacities, European J. Oper. Res. neural networks, in: 2007 International Joint Conference on Neural Networks,
137 (3) (2002) 612–624. IEEE, 2007, pp. 2977–2981.
[79] J.-Z. Wu, Q. Zhang, Q. Du, Z. Dong, Compromise principle based methods of [107] S. Dey, R. Roychoudhury, S. Malakar, R. Sarkar, An optimized fuzzy ensemble
identifying capacities in the framework of multicriteria decision analysis, Fuzzy of convolutional neural networks for detecting tuberculosis from Chest X-ray
Sets and Systems 246 (2014) 91–106. images, Appl. Soft Comput. 114 (2022) 108094.
[80] J.-Z. Wu, S. Yang, Q. Zhang, S. Ding, 2-additive capacity identification methods [108] C.-J. Lin, C.-H. Lin, S.-H. Wang, C.-H. Wu, Multiple convolutional neural
from multicriteria correlation preference information, IEEE Trans. Fuzzy Syst. networks fusion using improved fuzzy integral for facial emotion recognition,
23 (6) (2015) 2094–2106. Appl. Sci. 9 (13) (2019) 2593.
[81] J.-Z. Wu, Y.-P. Zhou, L. Huang, J.-J. Dong, Multicriteria correlation pref- [109] P. Bhowal, S. Sen, J.D. Velasquez, R. Sarkar, Fuzzy ensemble of deep learning
erence information (MCCPI) based ordinary capacity identification method, models using Choquet fuzzy integral, coalition game and information theory for
Mathematics 7 (3) (2019) 300. breast cancer histology classification, Expert Syst. Appl. 190 (2022) 116167.
[82] L. Jin, R. Mesiar, R.R. Yager, Melting probability measure with OWA operator [110] J. Wang, Z. He, S. Huang, H. Chen, W. Wang, F. Pourpanah, Fuzzy measure
to generate fuzzy measure: the crescent method, IEEE Trans. Fuzzy Syst. 27 (6) with regularization for gene selection and cancer prediction, Int. J. Mach. Learn.
(2018) 1309–1316. Cybern. 12 (2021) 2389–2405.

10
G. Beliakov et al. Information Fusion 106 (2024) 102295

[111] M.X. Uriz Martín, D. Paternain Dallo, H. Bustince Sola, M. Galar Idoate, [130] I. Kojadinovic, Minimum variance capacity identification, European J. Oper.
An empirical study on supervised and unsupervised fuzzy measure construc- Res. 177 (2) (2007) 498–514.
tion methods in highly imbalanced classification, in: 2020 IEEE International [131] M. Grabisch, I. Kojadinovic, P. Meyer, A review of methods for capacity
Conference on Fuzzy Systems, FUZZ-IEEE, Glasgow, UK, 2020, 1–8, IEEE, 2020. identification in Choquet integral based multi-attribute utility theory, Eur. J.
[112] H.E. de Oliveira, L.T. Duarte, J.M.T. Romano, Identification of the Choquet Oper. Res. 186 (2008) 766–785.
integral parameters in the interaction index domain by means of sparse [132] J. Murillo, S. Guillaume, E. Tapia, P. Bulacio, Revised HLMS: A useful algorithm
modeling, Expert Syst. Appl. 187 (2022) 115874. for fuzzy measure identification, Inf. Fusion 14 (4) (2013) 532–540.
[113] F. Bardozzo, B. De La Osa, L. Horanská, J. Fumanal-Idocin, M. delli Priscoli, L.
[133] J. Murillo, S. Guillaume, P. Bulacio, k-maxitive fuzzy measures: A scalable
Troiano, R. Tagliaferri, J. Fernandez, H. Bustince, Sugeno integral generalization
approach to model integractions, Fuzzy Sets and Systems 324 (2017) 33–48.
applied to improve adaptive image binarization, Inf. Fusion 68 (2021) 37–45.
[134] G. Beliakov, J.-Z. Wu, Random generation of capacities and its application in
[114] G.E. Martínez, C.I. Gonzalez, O. Mendoza, P. Melin, General type-2 fuzzy sugeno
comprehensive decision aiding, Inform. Sci. 577 (2021) 424–435.
integral for edge detection, J. Imag. 5 (8) (2019) 71.
[115] M. Boczek, M. Kaluszka, On the extended Choquet–Sugeno-like operator, [135] E.F. Combarro, I. Díaz, P. Miranda, On random generation of fuzzy measures,
Internat. J. Approx. Reason. 154 (2023) 48–55. Fuzzy Sets Syst. 228 (4) (2013) 64–77.
[116] J. Fumanal-Idocin, Z. Takáč, L. Horanská, T. da Cruz Asmus, G. Dimuro, C. [136] R.L. Smith, Efficient Monte Carlo procedures for generating points uniformly
Vidaurre, J. Fernandez, H. Bustince, A generalization of the Sugeno integral to distributed over bounded regions, Oper. Res. 32 (6) (1984) 1296–1308.
aggregate interval-valued data: An application to brain computer interface and [137] P.A. Rubin, Generating random points in a polytope, Comm. Statist. Simulation
social network analysis, Fuzzy Sets and Systems 451 (2022) 320–341. Comput. 13 (3) (1984) 375–396.
[117] M. Cardin, S. Giove, Aggregation functions with non-monotonic measures, Fuzzy [138] E. Combarro, P. Miranda, Identification of fuzzy measures from sample data
Econ. Rev. 13 (2) (2008) 3. with genetic algorithms, Comput. Oper. Res. 33 (10) (2006) 3046–3066.
[118] T.C. Havens, D.T. Anderson, Machine learning of choquet integral regression [139] E.F. Combarro, J. Hurtado de Saracho, I. Díaz, Minimals Plus: An improved
with respect to a bounded capacity (or non-monotonic fuzzy measure), in: 2019 algorithm for the random generation of linear extensions of partially ordered
IEEE International Conference on Fuzzy Systems, FUZZ-IEEE, IEEE, 2019, pp. sets, Inform. Sci. 501 (2019) 50–67.
1–6. [140] L. Devroye, Non-Uniform Random Variate Generation, Springer Verlag, New
[119] S.H. Kwon, M. Sugeno, Extension of the subset interactive AR model using non- York, 1986.
monotonic fuzzy measures, in: FUZZ-IEEE’99. 1999 IEEE International Fuzzy
[141] T.C. Havens, A.J. Pinar, Generating random fuzzy (capacity) measures for
Systems. Conference Proceedings (Cat. No. 99CH36315), Vol. 2, IEEE, 1999,
data fusion simulations, in: 2017 IEEE Symposium Series on Computational
pp. 1191–1195.
Intelligence, SSCI, IEEE, 2017, pp. 1–8.
[120] Z. Gong, X. Kou, T. Xie, Interval-valued choquet integral for set-valued
[142] J.-Z. Wu, G. Beliakov, Generating fuzzy measures from additive measures, 2023,
mappings: definitions, integral representations and primitive characteristics,
arXiv preprint arXiv:2309.15399.
AIMS Math. 5 (6) (2020) 6277–6297.
[121] B. Mihailovic, E. Pap, Non-monotonic set functions and general fuzzy integrals, [143] J.-Z. Wu, G. Beliakov, Random generation of linearly constrained fuzzy mea-
in: 2008 6th International Symposium on Intelligent Systems and Informatics, sures and domain coverage performance evaluation, Inform. Sci. 659 (2024)
IEEE, 2008, pp. 1–4. 120080.
[122] D. Zhang, C. Guo, D. Liu, Set-valued choquet integrals revisited, Fuzzy Sets and [144] P. Sun, M. Grabisch, C. Labreuche, An improvement of random node generator
Systems 147 (3) (2004) 475–485. for the uniform generation of capacities, in: International Conference on
[123] Y. Huang, C. Wu, Real-valued choquet integrals for set-valued mappings, Scalable Uncertainty Management, Springer, 2022, pp. 202–216.
Internat. J. Approx. Reason. 55 (2) (2014) 683–688. [145] M. Grabisch, C. Labreuche, P. Sun, An approximation algorithm for random
[124] P. Meyer, M. Roubens, On the use of the Choquet integral with fuzzy numbers generation of capacities, Order (2023) 1–26.
in multiple criteria decision support, Fuzzy Sets and Systems 157 (7) (2006) [146] A.B. Kahn, Topological sorting of large networks, Commun. ACM 5 (1962)
927–938. 558–562.
[125] P. Fortemps, M. Roubens, Ranking and defuzzification methods based on area [147] G. Beliakov, F.J. Cabrerizo, E. Herrera-Viedma, J.-Z. Wu, Random generation
compensation, Fuzzy Sets and Systems 82 (3) (1996) 319–330. of k-interactive capacities, Fuzzy Sets and Systems 430 (2022) 48–55.
[126] R. Yang, Z. Wang, P.-A. Heng, K.-S. Leung, Fuzzy numbers and fuzzification of
[148] G. Beliakov, On random generation of supermodular capacities, IEEE Trans.
the choquet integral, Fuzzy Sets and Systems 153 (1) (2005) 95–113.
Fuzzy Syst. 30 (1) (2022) 293–296.
[127] G. Beliakov, Knapsack problems with dependencies through non-additive
[149] S.G. Arcidiacono, S. Corrente, S. Greco, As simple as possible but not simpler in
measures and choquet integral, European J. Oper. Res. 301 (1) (2022) 277–286.
multiple criteria decision aiding: the robust-stochastic level dependent Choquet
[128] G. Beliakov, Optimising non-convex choquet integrals using DC (difference of
integral approach, European J. Oper. Res. 280 (2020) 988–1007.
convex) algorithm, in: 2022 IEEE International Conference on Fuzzy Systems,
FUZZ-IEEE, IEEE, 2022, pp. 1–6. [150] B. Murray, M.A. Islam, A.J. Pinar, T.C. Havens, D.T. Anderson, G. Scott,
[129] G. Beliakov, J.-Z. Wu, Discrete choquet integral optimisation with positive Explainable ai for understanding decisions and data-driven optimization of the
and negative interactions, IEEE Trans. Fuzzy Syst. (2023) http://dx.doi.org/ Choquet integral, in: 2018 IEEE International Conference on Fuzzy Systems,
10.1109/TFUZZ.2023.3318631. FUZZ-IEEE, IEEE, 2018, pp. 1–8.

11

You might also like