International Journal of Approximate Reasoning 86 (2017) 80–91

Contents lists available at ScienceDirect

International Journal of Approximate Reasoning

www.elsevier.com/locate/ijar

Learning preferences from paired opposite-based semantics ✩

Camilo Franco a,∗ , J. Tinguaro Rodríguez b , Javier Montero b,c
a
IFRO, Faculty of Science, University of Copenhagen, Frederiksberg, 1870, Denmark
b
Faculty of Mathematics, Complutense University, Madrid, 28040, Spain
c
Institute IGEO (CISC-UCM), Complutense University, Madrid, 28040, Spain

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

Article history: Preference semantics examine the meaning of the preference predicate, according to the
Received 22 February 2016 way that alternatives can be understood and organized for decision making purposes.
Received in revised form 21 October 2016 Through opposite-based semantics, preference structures can be characterized by their
Accepted 30 April 2017
paired decomposition of preference into opposite poles, and their respective valuation of
Available online 5 May 2017
binary preference relations. Extending paired semantics by fuzzy sets, preference relations
Keywords: can be represented in a gradual functional form, under an enhanced representational
Paired concepts frame for examining the meaning of preference. Following a semantic argument on the
Fuzzy logic character of opposition, the compound meaning of preference emerges from the fuzzy
Preference structures reinforcement of paired opposite concepts, searching for significant evidence for affirming
Semantic opposition dominance among the decision objects. Here we propose a general model for the paired
Fuzzy reinforcement decomposition of preference, examining its characteristic semantics under a binary and
Significance
fuzzy logical frame, and identifying solutions with different values of significance for
preference learning.
© 2017 Elsevier Inc. All rights reserved.

1. Introduction

The meaning of the preference concept refers to the way decision objects or alternatives can be ordered for decision
making purposes. This concept can be examined through the preference predicate, regarding the paired opposite decom-
position of the positive and negative perceptions of preference. In psychology (see e.g. [3,11,17]), the meaning of concepts
has been studied in relation to its valuation as being either positive or negative, eliciting a subjective measurement from
the individual, but at the same time, requiring that the individual somehow solves the natural ambivalence involved in
understanding opposite perceptions. For doing this, a semantic scale has been commonly used (initially proposed in [17]),
measuring the meaning of concepts according to a given pair of opposite poles.
In decision theory (see e.g. [4,10,22]), such a bi-polarity has been studied from two perspectives. The first one can be
referred as the univariate model, introducing a one-dimensional scale with opposite references as endpoints, in such a way
that one of these references is taken to be (or is understood) as positive, in opposition to the other which is considered
as its negative counterpart. In this setting, it is possible to further introduce a reciprocity assumption, so that the verifica-
tion status of one pole entails by complementation a particular status of the other. For example, if we assume reciprocal
preferences and we have absolute preference for watching a fiction movie over a documentary, then it is already assumed

✩
This paper is part of the virtual special issue on Advances in Weighted Logics for Artificial Intelligence, edited by Marcelo Finger, Lluis Godo, Henri
Prade and Guilin Qi.
*
Corresponding author.
E-mail addresses: cf@ifro.ku.dk (C. Franco), jtrodrig@mat.ucm.es (J.T. Rodríguez), monty@mat.ucm.ess (J. Montero).

http://dx.doi.org/10.1016/j.ijar.2017.04.010
0888-613X/© 2017 Elsevier Inc. All rights reserved.
 C. Franco et al. / International Journal of Approximate Reasoning 86 (2017) 80–91 81

that we have absolutely no preference for watching a documentary. In this (restricted) univariate reciprocal model, where
opposite preference is directly associated with the inverse preference, it is then possible to prefer one or the other, and if
no reciprocity is imposed, we may even prefer nothing at all, such that the value for preference is either positive, negative,
or neither positive nor negative.
On the other hand, the second perspective on bipolarity can be referred as the unipolar bivariate model (see e.g. [10,
11]), where a concept can be positive, negative, neither positive nor negative, or both positive and negative, thus allowing
preference (and aversion) for watching both a fiction and a documentary. Then the relation among poles is not simple, but
rather complex and depends on the particular semantic relation holding among the opposite concepts. Even more, taking
into consideration multi-dimensional concepts, like multiple viewpoints describing the properties of objects, the semantic
relation holding among opposites may require a more complex analysis.
Focusing on the different neutral states holding in between the opposite poles, and stressing the determinant role that
those states have for representing the meaning of concepts, logical paired structures [13] provide an adequate framework
to address such complexity. Particularly, in previous works we have studied how the different neutral states in between
opposite poles that are postulated in the context of paired structures may be useful tools to represent and understand the
complexity of preference concepts (see [8]), allowing to configure a pertinent valuation preference structure for ordering
the decision objects or alternatives.
Furthermore, paired structures [8,13] explicitly represent the different and non-reciprocal sources of information building
up the complex meaning of preference. In this respect, and on a neurological level, it is observed that the meaning of
concepts emerges from the multiple positive and aversive stimuli composing perceptions and emotions [3,9]. In this sense,
different pleasant and unpleasant affective components of the same sensory stimulus, processed separately in different
physical areas of the brain (see e.g. [16,23]), may provide the inputs of human behavior and decision making [1,9]. Therefore,
as neurological observation suggests, the positive and negative counterparts are formed and evaluated separately, in an
independent manner, configuring a significant decision space with respect to the available (positive and negative) evidence.
In this way, opposite sets of evidence can be simultaneously evaluated as separate entities, coming together under the
fuzzy reinforcement of their intensities. That is, given the separate nature of positive and negative aspects, they can be jointly
examined as they reinforce each other under an appropriate aggregation process based on opposition operators [13,18]. As it
will be examined in detail in the later sections of this paper (Sections 5 and 6), opposite pieces of evidence, coming from
different sources, can be used to reinforce each other in order to find greater significance to their inferred meaning.
Here we propose a general setting where the performance of the different preference models can be formally assessed
according to their significance. For this purpose, a measure of relative significance is introduced for evaluating the amount
of evidence for affirming preference. As a result, the aggregation process unravels while searching for significant evidence
on pairwise dominance for preference learning and intelligent (automatic) decision support, where the reliance on the
emotional meaning associated to the alternatives allows explaining and identifying satisfactory (descriptive) viewpoints for
decision-making (see e.g. [12,22]).
In order to examine opposite-based preference semantics for preference learning, Section 2 introduces standard prefer-
ence structures, where the inverse preference relation represents the negative perception on preference. Section 3 extends
the analysis to more complex preference structures, where positive and negative preferences are independently represented
and measured. Then, in Section 4, fuzzy preference structures are introduced, examining a general frame for fuzzy paired
preference semantics. In Section 5, a proposal for the significance of preference orders is given, and in Section 6, the general
methodology for learning preferences is proposed, based on fuzzy reinforcement and maximal significance. Finally, Section 7
offers a numerical example illustrating the proposed methodology, ending with some open challenges for future research.

2. Standard preference semantics

Given a set of alternatives A, a crisp preference structure can be defined ∀(a, b) ∈ A2 , by the decomposition of the weak
preference predicate R (a, b) = a is at least as desired as b, into three basic binary relations P , I , J , such that (see e.g. [5]),

P: The pair of alternatives (a, b) ∈ A2 belongs to the strict preference relation P , if and only if a is more desired than b,
expressed by P (a, b). The inverse strict preference predicate, b is more desired than a, is expressed by P (b, a) = P −1 (a, b).
I: The pair of alternatives (a, b) ∈ A2 belongs to the indifference relation I , if and only if a is as much as desired as b,
expressed by I (a, b).
J: The pair of alternatives (a, b) ∈ A2 belongs to the incomparability relation J , if and only if a cannot be compared with
b, expressed by J (a, b).

Under standard preference modeling (see again [5] but also [19] and [14]), it is assumed that ∀a, b ∈ A, the relations
I and J are symmetrical, such that I (a, b) = I (b, a) and J (a, b) = J (b, a) hold, I is reflexive, such that I (a, a) holds, J is
irreflexive, such that J (a, a) does not hold, and P is asymmetrical, such that P (a, b) and P (b, a) cannot hold simultaneously.
Then, the preference structure ( P , I , J ), is such that only one situation holds, as in,

P ∩ I = ∅, (1)
P ∩ J = ∅, (2)
82 C. Franco et al. / International Journal of Approximate Reasoning 86 (2017) 80–91

J ∩ I = ∅, (3)
fulfilling properties,

P ∪ I = R, (4)
−1 −1
P∪I∪P =R∪R , (5)
d
P∪ J=R , (6)

where R d = N ( R −1 ) = ( N ( R ))−1 , for the logical complement N, and being complete in the sense of,

P ∪ P −1 ∪ I ∪ J = A2 . (7)
Exploring preference semantics under the standard structure, the meaning of R is captured between the two predicates
of R and R −1 , which stand as opposite (inverse) poles evaluating the meaning of preference.
Opposite poles offer a set of valuation references for positive and negative judgments, hereby denoted by Q and V ,
respectively. Hence, poles can be assigned any pair of relations having a specific type of opposite semantics between them,
being here the case that Q = R and V = R −1 , such that,

P = Q ∩ N ( V ) = R ∩ N ( R −1 ),
I = Q ∩ V = R ∩ R −1 ,
J = N ( Q ) ∩ N ( V ) = N ( R ) ∩ N ( R −1 ),

but notice here that it could also be, e.g., that Q = P and V = P −1 . In consequence, the meaning assigned to the poles Q
and V may respectively correspond with R and R −1 , or with P and P −1 .
In this way, in the standard structure, poles Q and V are allowed to overlap, under the specific semantic condition that
Q = R and V = R −1 . Otherwise, if poles are assigned the strict preference values, such that Q = P and V = P −1 , then the
neutral situation of indifference (I ) can only be defined by exclusion, as the state holding if no pole is verified. Thus, it
is a general attitude towards preference, following a semantic argument, which allows distinguishing between indifference,
either as a situation defined by exclusion among poles (case of the univariate model), or as a situation defined by the
overlap among poles (case of the bivariate model).
In order to grasp the relevance of this semantic argument, consider the following result, taking the valuation set {0, 1},
such that ∀(a, b) ∈ A2 , R (a, b) = 1 if R (a, b) holds, otherwise R (a, b) = 0, an involutive negation operator N, such that
∀x ∈ {0, 1}, N ( N (x)) = x, and a reciprocal preference relation R, such that Q = R and V = R −1 = N ( R ), defined for all pairs
{(a, b) ∈ A2 |a = b}.1

Proposition 1. If R = P ∪ I is a reciprocal preference relation, i.e., such that its semantic poles are complementary, then R = P .

Proof. Under the hypothesis that N ( V ) = Q , where it holds that Q = R and V = R −1 = N ( R ), it is true that N ( V ) =
N ( N ( R )) = R = Q . According to (4), it holds that R = P ∪ I , and hence that N ( R ) = P −1 . Therefore, it follows that R −1 =
N ( R ) ⇐⇒ R −1 = P −1 , being true that ( R −1 )−1 = ( P −1 )−1 = R = P . 2

As a result, based on Proposition 1, there is a single concept Q and its complementation V = N ( Q ), being there no place
for indifference as a neutral category [13] if the poles of preference are complementary. In this sense, reciprocity entails a
binary unipolar model where either P or P −1 holds (see [8]).

Example 1. Consider the decision problem of choosing a place to live, and a recommendation system offering feedback on
suitable alternatives. Let us have two alternatives a, b ∈ A, assessing their desirability according to the price. After eliciting
the preferences for a over b, a given user states that R (a, b) = 1, and at the same time, that R (b, a) = 1. Then, if the
system takes both pieces of evidence, the result is that I (a, b) = 1 holds, but if the system assumes (additive) reciprocity,
after verifying that R (a, b) = 1, it then obtains that R (b, a) = 0. In this way, under the reciprocal condition, it results that
P (a, b) = 1, and consequently, the recommendation system would identify a as the most desired alternative for the user.
Otherwise, allowing the separate elicitation of the inverse preference, the system would identify both alternatives as being
just as desirable. In consequence, the reciprocity condition entails some important loss of information, as the elicitation of
the inverse preference reveals that the user is indifferent between both alternatives.

Considering the process for the formation of preferences, candidate alternatives are gradually understood and brought
forward regarding their specific attributes, which are perceived under a paired decomposition that allows representing the

1
Notice that being R a reciprocal relation, indifference (I ) is defined by exclusion, such that pairs (a, a) ∈ A2 are always indifferent, i.e., ∀a ∈ A, I (a, a) = 1.
 C. Franco et al. / International Journal of Approximate Reasoning 86 (2017) 80–91 83

emotional character of decisions (see e.g. [1,3,12,23]). Hence, following on Example 1, the meaning of price can refer not
only to the negative condition of establishing a restriction on the user’s budget, but also, to the positive perception on the
social condition that a high price entails for the individual. Therefore, an extended frame for preference representation is
needed to properly take into account both the positive and the negative dimensions of decisions, aiming at capturing the
relevant aspects for identifying and explaining different viewpoints for decision-making.

3. Extended preference semantics

The standard preference structure can be extended according to a deeper conceptual decomposition of the preference
predicate, based on four different sources of information (see e.g. [6,8]). On the one hand, for every a, b ∈ A, take the positive
aspects of alternatives to evaluate the preference predicate R + (a, b) = R (a, b) = a is at least as desired as b and its inverse
R + (b, a), and on the other, take their negative aspects for evaluating the negative preference predicate R − (a, b) = a is at
least as non-desired/rejected as b and its inverse R − (b, a). These paired references are associated to logically independent poles
( Q , V ), such that R + (a, b) = Q (a, b), R + (b, a) = Q (b, a) = Q −1 (a, b), R − (a, b) = V (a, b) and R − (b, a) = V (b, a) = V −1 (a, b).
After some neurological evidence on how the brain deals with the multiple positive and aversive stimuli composing
perceptions and emotions (processing separately the pleasant and the unpleasant affective components in different physical
areas of the brain [9,16,23]), it can be examined how opposite-based semantics may provide the inputs for preference-based
decision making. Therefore, preference can be understood under a more complex evaluation space [3], generalizing the one
of the standard model (as shown in [8]), representing the activation channels for positive and negative affections by a pair
of opposite conceptual poles Q = “desire” and V = “rejection”.
Thus, each pole is firstly, and separately, decomposed into their respective standard structures for Preference and Aversion
(P–A) [6], where positive aspects are measured with respect to ( Q , Q −1 ), obtaining the standard structure ( P , I , J ) through
expressions (1)–(7). Analogously, the negative aspects are measured with respect to ( V , V −1 ), obtaining the aversion structure
( Z , G , H ).
In consequence, the aversion structure is defined ∀(a, b) ∈ A2 , by the decomposition of the weak aversion predicate
R − = ( V , V −1 ), into three basic binary relations Z , G , H , such that,

Z: The pair of alternatives (a, b) ∈ A2 belongs to the strict aversion relation Z , if and only if a is more rejected than b,
denoted by Z (a, b).
G: The pair of alternatives (a, b) ∈ A2 belongs to the aversion indifference relation I , if and only if a is as much as rejected
as b, denoted by G (a, b).
H: The pair of alternatives (a, b) ∈ A2 belongs to the aversion incomparability relation J , if and only if a cannot be compared
with b on their negative aspects, denoted by H (a, b).

Hence, ∀a, b ∈ A, the one-dimensional relations G and H are symmetrical, G is reflexive, H is irreflexive, and Z is
asymmetrical, and the three relations are linked together as in,

Z ∩ G = ∅, (8)
Z ∩ H = ∅, (9)
H ∩ G = ∅, (10)
such that,

Z ∪G = V, (11)
−1 −1
Z ∪G∪ Z =V ∪V , (12)
d
Z∪H=V , (13)
−1 2
Z∪Z ∪G∪H =A . (14)

Based on the aversion structure, the negative meaning of R − is captured between V and V −1 , such that,

Z = V ∩ N ( V −1 ),
G = V ∩ V −1 ,
H = N ( V ) ∩ N ( V −1 ).
As a result, every pair a, b ∈ A can be simultaneously valued as belonging to one relation in ( P , I , J ), and one in
( Z , G , H ), building the meaning of preference on the separate and simultaneous verification of the positive and negative
aspects of decisions. Extending this binary model to a continuous one, fuzzy logic allows a gradual valuation of preference,
exploring how are preferences weighed according to the intensity in which they are perceived (see e.g. [5,6,14]). Next, fuzzy
preference structures are examined together with paired-opposite semantics.
84 C. Franco et al. / International Journal of Approximate Reasoning 86 (2017) 80–91

4. Fuzzy paired semantics

Fuzzy preference structures [5,14,24] represent preference relations as gradual predicates which may be verified up to a
certain degree of intensity, fulfilling as much as possible conditions (4)–(7) and (11)–(14). In this way, the decision outcome
is valued by different preference states and their degrees of verification, instead of only one preference/aversion situation as
in the classical-binary setting of (1)–(3) and (8)–(10).
From this standpoint, the characterization of a fuzzy preference relation is given by

R (a, b) = {a, b, μ R (a, b)|a, b ∈ A},

such that

μ R (a, b) ∈ [0, 1]
is the membership function of R, measuring the degree up to which any pair of decision objects verify the preference
predicate. In this way, there are a pair of functions representing the predicates R + and R − , that under the hypothesis that
Q = R + and V = R − , are given by,

μ Q , μ V : A2 → [0, 1].
Taking the paired decomposition of preference according to the P–A framework,

R P − A = ( Q , Q −1 ), ( V , V −1 ) = ( P , I , J ), ( Z , G , H ), (15)
fuzzy preference relations

p , i , j , z, g , h : [0, 1]2 → [0, 1]

can be defined for Q = R + = ( P , I , J ) and V = R − = ( Z , G , H ), such that (following [5,6])

P = p (μ Q , μ Q −1 ) = T (μ Q , N (μ Q −1 )),
I = i (μ Q , μ Q −1 ) = T (μ Q , μ Q −1 ),
J = j (μ Q , μ Q −1 ) = T ( N (μ Q ), N (μ Q −1 )),
Z = z(μ V , μ V −1 ) = T (μ V , N (μ V −1 )),
G = g (μ V , μ V −1 ) = T (μ V , μ V −1 ),
H = h(μ V , μ V −1 ) = T ( N (μ V ), N (μ V −1 )),
where T is a continuous t-norm (a standard fuzzy conjunctive operator), N is a strict negation (in this fuzzy setting, given
by a strictly decreasing, continuous function), and i , j , g and h are symmetrical functions.
Therefore, the standard properties (4)–(7) and (11)–(14), can be respectively formulated in fuzzy logic by means of a
continuous t-conorm S,2 such that (see again [5,6]),

S ( p, i) = μ Q , (16)
−1
S ( p, i, p ) = S (μ Q , μ Q −1 ), (17)
S ( p , j ) = N (μ Q ), (18)
S ( p , p −1 , i , j ) = 1, (19)
and

S ( z, g ) = μ V , (20)
−1
S ( z, g , z ) = S (μ V , μ V −1 ), (21)
S ( z, h) = N (μ V ), (22)
−1
S ( z, z , g , h ) = 1. (23)
As a result, the only solution for ( p , i , j ) and ( z, g , h), characterized by t-norm and t-conorm aggregation operators T , S
and negation N, respectively fulfilling (16)–(19) and (20)–(23), is given by (see e.g. [5,6]),

2
Notice that although T and S are defined here as standard (commutative, associative and monotonic) t-norm and t-conorm operators (see e.g. [5]),
they could also be defined differently, e.g. by means of overlapping and grouping operators [2], respectively defining T and S as not necessarily associative
operators.
 C. Franco et al. / International Journal of Approximate Reasoning 86 (2017) 80–91 85

Table 1
Fuzzy preference/aversion structures.

P–A p = z = TM p = z= TL p=z=Tp
p 0.1 0.0 0.05
p −1 0.5 0.4 0.45
i 0.4 0.5 0.45
j 0.0 0.1 0.05
z 0.3 0.0 0.18
z−1 0.4 0.1 0.28
g 0.3 0.6 0.42
h 0.0 0.3 0.12

p (μ Q , μ Q −1 ) = T M (μ Q , N (μ Q −1 )),

i (μ Q , μ Q −1 ) = T L (μ Q , μ Q −1 ),

j (μ Q , μ Q −1 ) = T L ( N (μ Q ), N (μ Q −1 )),

and

z(μ V , μ V −1 ) = T M (μ V , N (μ V −1 )),

g (μ V , μ V −1 ) = T L (μ V , μ V −1 ),

h(μ V , μ V −1 ) = T L ( N (μ V ), N (μ V −1 )),

where N is a strong (involutive) negation, such that ∀x ∈ [0, 1], N ( N (x)) = x, and ∀x, y ∈ [0, 1], T L (x, y ) = max(x + y −
1, 0), T M = min(x, y ), and S = S L = min(x + y , 1). That is, ( T L , S L , N ) is the Lukasiewicz De Morgan triple along with its
respective residual T M . Under this solution, i and j are mutually exclusive in ( p , i , j ), as well as g and h are mutually
exclusive in ( z, g , h), expressing the classical logical impossibility of finding any pair of objects that are equally desired
(rejected) but at the same time incomparable on their desirable (rejectable) attributes.
Another solution (not fully complying with the standard properties) can be identified for fuzzy preference structures [5],
if p (z) is defined as a strongly asymmetrical relation, such that ∀x, y ∈ [0, 1], p (x, y ) > 0 ⇒ p −1 (x, y ) = 0. Then, defining
p , z by means of the t-norm T L , and i , j , g , h by means of T M , both the preference (16), (18)–(19), and the aversion
conditions (20), (22)–(23), are satisfied, but not (17) nor (21).
Lastly (see [24]), there is another limit solution that allows the simultaneous verification of all basic relations in ( p , i , j )
and ( z, g , h), but that does not fulfill (16)–(18) nor (20)–(22). This solution (which is complete in the sense of (19) and
(23)), is given by the multiplicative or probabilistic De Morgan triple ( T p , S p , N ), such that ∀x, y ∈ [0, 1], T p (x, y ) = x · y
and S = S p = x + y − x · y.

Example 2. Following Example 1, the initial two alternatives a, b ∈ A can be assessed regarding their desirability as well
as their rejectability regarding relevant attributes such as price, location and area. Consider here the case where the user
reveals the overall weak preference and aversion values, given by μ Q (a, b) = 0.5, μ Q −1 (a, b) = 0.9, μ V (a, b) = 0.6 and
μ V −1 (a, b) = 0.7. All the different solutions for the fuzzy preference structures can be seen in Table 1, where p = z = T M
stands for the solution complying with all conditions (16)–(23), p = z = T L agrees with the asymmetrical modeling of p
and z, and p = z = T p takes the multiplicative De Morgan triple. As a result, the first solution obtains that b is preferred
to a, although both alternatives seem to be similar both on the positive and negative aspects. The other solutions confirm
their P–A indifference, with a slightly greater intensity for preference on b over a, suggesting a weak recommendation for
choosing b.

These solutions for the multi-valued characterization of R P − A (15), through fuzzy preference structures R + , and R − ,
illustrate the fuzzy-paired semantics of P–A. This approach proposes an analytical neurological foundation for preference
learning (as it will be examined in the following sections), stressing the separate conception of both positive and negative
components, being accordingly treated for understanding the meaning of human perceptions and stimuli. Then, in order to
address the P–A meaning, both opposite dimensions have to be aggregated.
The paired networks of P–A may interact in different ways, according to the semantic requirements on their aggregation.
Thus, they can be either conjunctively or disjunctively aggregated, or under a fair aggregation, they can be allowed to
interact by compensating the positive and the negative aspects, through the reinforcement of desirable and non-rejectable
intensities.
For this purpose, different types of aggregation operators can be examined for representing the P–A meaning, learning
preference orders with different degrees of (evidence-supported) significance.
86 C. Franco et al. / International Journal of Approximate Reasoning 86 (2017) 80–91

5. Significance of preference orders

The meaning of preference responds to the perception of opposite stimuli, which, absorbed through the proposed struc-
ture R P − A (15), offer a pair of complex relational networks that may interact together by reinforcing the intensity for
desire through the corroboration of non-aversion.3 In this sense, the significance of verifying both classes of desirable and
non-rejectable evidence, is greater than verifying only one, or either one of them.4
Consider a preference (weak) order  on A, such that θ a stands for the position of a ∈ A in the order . Then, ∀a, b ∈ A,
it holds one of the following: θ a > θ b , θ b > θ a , or θ b = θ a . In case that θ b > θ a (θ a > θ b ), it holds that a dominates b
(b dominates a), otherwise a and b are equivalent in order . In this way, it is proposed that by learning a significant order
, a position is assigned to every alternative a ∈ A, such that the significance of  increases with the amount of evidence
supporting it.
For every a, b ∈ A, let Q = { Q , Q −1 }∀a,b∈A stand for the set of evidence on the positive aspects of preference, and let
V = { V , V −1 }∀a,b∈A stand for the set of evidence on the negative aspects of aversion. Then significance can be defined as
follows.

Definition 1. Given two separate sets of evidence Q = { Q , Q −1 }∀a,b∈A and V = { V , V −1 }∀a,b∈A , the significance for affirming
preference for a is measured by an evidence counter ea , initially set up to ea = 0. In this way, whenever it holds that
Q (a, b) > Q (b, a), then ea = ea + 1 (i.e., there is one more unit of evidence for affirming that a is a desirable alternative),
and in the same way, if it holds that V (b, a) > V (a, b), then ea = ea + 1 (i.e., there is one more unit of evidence for affirming
that a is a non-rejectable alternative).

Thus, the significance of a preference order can be understood as a monotone, increasing function of the amount of
evidence supporting each pairwise ordering of alternatives. In this sense, if both Q (a, b) > Q (b, a) and V (b, a) > V (a, b)
hold, then ea = ea + 2, and if it holds ∀b = a ∈ A, that Q (a, b) > Q (b, a) and V (b, a) > V (a, b), then ea = 2(n − 1).

Proposition 2. Given two separate sets of evidence Q = { Q , Q −1 }∀a,b∈A and V = { V , V −1 }∀a,b∈A , and two preference orders on A,
1 and 2 , such that 1 is supported by both sets of evidence Q and V, and 2 is supported by only one set of evidence from Q or V.
If there exists some a, b ∈ A where both Q (a, b) > Q (b, a) and V (b, a) > V (a, b) hold, then the significance of 1 is greater than the
significance of 2 .

Proof. For every a ∈ A, count the evidence for affirming preference for a under index ea , such that if Q (a, b) > Q (b, a)
holds, then ea = ea + 1, and in the same way, if V (b, a) > V (a, b) holds, then ea = ea + 1. It then follows that ∀a ∈ A,

the value of ea under 2 , ea 2 , counts the evidence on either Q (a, b) > Q (b, a) or V (b, a) > V (a, b), such that if both
  
Q (a, b) > Q (b, a) and V (b, a) > V (a, b) hold, then ea 2 = ea 2 + 1. But under 1 , ea 1 counts the evidence on both Q (a, b) >
 
Q (b, a) and V (b, a) > V (a, b), such that if both Q (a, b) > Q (b, a) and V (b, a) > V (a, b) hold, then ea 1 = ea 1 + 2, such that
1 2
ea > ea . Thus, if there exists some a, b ∈ A where both Q (a, b) > Q (b, a) and V (b, a) > V (a, b) hold, then the significance
of 1 is greater than the significance of 2 . 2

Based on Proposition 2, a measure of relative significance can be proposed, given two separate sets of evidence, one for
the desired attributes Q, and another for the rejected attributes V, and given a reference order 1 , built from both bodies
of evidence Q and V. In this way, the degree of significance of another order 2 , with respect to the reference order 1 , is
measured by

1 
σ (1 , 2 ) = 1 − dist (θ1a , θ2a ), (24)
|A|2
∀a∈A

where dist can be taken e.g. as a 1-norm distance, such that

dist (θ1a , θ2a ) = |θ1a − θ2a |.

The relative significance measure (24) computes the significance of the order 2 with respect to a reference, signif-
icant order 1 , coming from the joint verification of desire and non-aversion. Thus, it holds that σ (1 , 2 ) = 1, only if
1 = 2 , being equivalently maximal significant orders. Otherwise, 2 is a significant order up to a degree, given by
0 < σ (1 , 2 ) < 1. Notice that on a strictly formal level, there is a certain similarity between (24) and Spearman’s rank
correlation coefficient. Nonetheless, they belong to different contexts, as the degree of significance (24) measures if there

3
Notice that this approach is different from the one that represents the meaning of R P − A by aggregating its opposite conceptual structures as in
( P , I , J ) × ( Z , G , H ), generating 16 different relational situations (see e.g. [6,8]).
4
This observation on the greater significance of verifying the evidence on both the desirable and non-rejectable aspects is grounded on a dialectical
decision principle (see e.g. [7]), stating that needs are more important to satisfy than desires, where needs refer to both desired and not rejected alternatives.
 C. Franco et al. / International Journal of Approximate Reasoning 86 (2017) 80–91 87

is any difference between two rankings (where one of them is fixed as a reference, significant order); and the Spearman
correlation measures the statistical dependence or correlation between the ranking of two random variables.
Now, the procedure for building significant orders from the available evidence requires representing the way that both
opposite paired concepts of P–A are aggregated, as two separate complex perceptions, together with how they are weighed
along the aggregation process. Therefore, the meaning of R P − A is examined next, modeling the weighted compound predi-
cate for Q and not-V , as it organizes the positive and negative stimuli under a neurological frame for preference learning.

6. Learning preferences by fuzzy reinforcement

Focusing on the separate nature of positive and negative perceptions, and how they are weighed and transformed into
meaningful knowledge, it has been observed that opposite pieces of evidence, coming from different sources, can be used
to reinforce each other in order to find greater significance to their inferred meaning. Therefore, the meaning of Q can
be reinforced by not-V , once the predicate V has been transformed by some involutive operation into representing that
which needs to be avoided (see e.g. [6,7]). Here, opposition operators [13,18] allow modeling in a general and flexible way
the meaning of the predicate not-V , which is independent from the predicate Q , such that Q = not-V and not-Q = V .
In this way, ∀x, y ∈ [0, 1] and a strong negation N, an opposition operator A : [0, 1] → [0, 1], is such that [18],

A ( A (x)) = x,
x ≤ y ⇒ A ( y ) ≤ A (x).

Then, A is an antonym [21] if it holds that,

A ≤ N,

and A is an antagonism [18], also called sub-antonym [13], if it holds that,

A ≥ N.

An example for the opposition operator A, previously assuming the standard strong negation N = 1 − x, can be given by
the Sugeno family of strong negations or λ-complement [20], defined by (∀x ∈ [0, 1], λ > −1),

1−x
N λ (x) = , (25)
1 + λx
being an antonym whenever λ ≥ 0, and an antagonist whenever −1 < λ ≤ 0.
Through the opposition operator A, the reinforced interaction can take place among the paired network formed by Q
and A ( V ), denoted by R  =  Q , A ( V ), such that ∀a, b ∈ N, the degree of reinforced preference, or dominance, is given by,

R  (a, b) = C (μ Q (a, b), A (μ V (a, b))), (26)

where C is an aggregation function C : [0, 1]2 → [0, 1], being {0,1}-idempotent (i.e., ∀x ∈ {0, 1}, C (x, x) = x), and monotonic
(i.e., ∀x, y , z, w ∈ [0, 1], C (x, y ) > C ( w , z) whenever x > w and y > z.
If instead of taking the weak predicates Q = R + , V = R − , the strict predicates Q = P , V = Z are taken as opposite
poles, following a strict decision attitude towards preference, the dominance relation R  , can be transformed by a semantic
argument for taking the strict reinforcing predicates of P and A ( Z ). Thus, based on (26), the weak dominance relation is
characterized by the weak assignation of opposites Q = R + and V = R − , while the strict dominance relation is characterized
by the strict assignation of opposites Q = P and V = Z .
Under a strict attitude, fuzzy preference structures hold a set of solutions which allow the joint verification of
( p , p −1 , i , j ) and (z, z−1 , g , h) under the multiplicative solution; or having mutual exclusion among p and p −1 , z and z−1 ,
i and j or g and h, according to any of the Lukasiewicz–De Morgan solutions.
In consequence, considering the decision attitude towards preference (weak or strict), the specific semantic require-
ments for A (antonym or antagonist), and the type of aggregation for C , the meaning for R  (26) can be measured under
three general perspectives. The first perspective develops from a disjunctive meaning, where either degree of desirability
or non-rejection allows affirming a general preference order. The second one is more demanding, requiring both degrees
of desirability and non-rejection to be verified, while the third one allows some fair compensation among them, allowing
further interaction among the paired fuzzy reinforcing concepts of μ Q and A (μ V ).

6.1. Disjunctive meaning

Examining the meaning of R  from a disjunctive perspective, requires using a disjunctive aggregation operator. For this,
take an aggregation operator C = C s , such that C s (0, 1) = C s (1, 0) = 1. Hence, this is the most tolerant type of aggregation,
as the verification of either degree of desirability or non-rejection is enough for having some degree of dominance. Take e.g.
the (smallest) t-conorm C s = S M (being also a grouping operator [2]), such that ∀x, y ∈ [0, 1], S M = max(x, y ).
88 C. Franco et al. / International Journal of Approximate Reasoning 86 (2017) 80–91

As a result, the disjunctive dominance fuzzy relation is defined ∀a, b ∈ A, such that,
s
R (a, b) = C s (μ Q (a, b), A (μ V (a, b))). (27)
Hence, ∀a, b ∈ A, a simple voting procedure can be applied for preference learning based on R  , such that a vote for the s
s s
dominance of a over b is conceeded when R  (a, b) > R  (b, a). Thus, alternatives can be ranked (weakly ordered) according
to their total number of votes. This order is referred to as the disjunctive order, based on (27) and the votes in favor of, or
not against dominance, due to the intensity of preference on desire or non-rejection. Such a weak order is denoted by s .
The order s is the result of the most tolerant ordering procedure, as the verification of just one of the intensities being
aggregated is enough to affirm dominance. Hence (as pointed out in Proposition 2), this is a solution that fails to comply
with the more significant one of the conjunctive procedure, which is examined next.

6.2. Conjunctive meaning

From a conjunctive perspective, the meaning of R  requires a conjunctive aggregation operator. Thus, take an aggrega-
tion operator C = C t , such that C t (0, 1) = C t (1, 0) = 0. Therefore, this is the most demanding type of aggregation, as the
verification of both a degree of desirability and of non-rejection is needed for having some degree of dominance. Take e.g.
the t-norm C t = T M (being the largest t-norm and also an overlap operator [2]).
In consequence, the conjunctive dominance fuzzy relation is defined ∀a, b ∈ A, such that,

R t (a, b) = C t (μ Q (a, b), A (μ V (a, b))). (28)

Applying the same voting procedure as before, this time based on R  , ∀a, b ∈ A, a vote for the dominance of a over b
t

is conceeded when R t (a, b) > R t (b, a). In this way, alternatives are ranked according to their total number of votes under
the conjunctive order, based on (28) and the votes in favor of, and not against dominance, due to the intensity of preference
on desire and non-rejection. This order is denoted by t .
The order t is the result of the most demanding ordering procedure, as the verification of both the positive and
non-negative intensities is necessary to affirm dominance. Hence, this is the most significant solution, which nonetheless
can be difficult to obtain due to its strong requirements. In usual decision problems, in particular multi-criteria problems
(see e.g. [10,19]), some trade-off or compensation among the decision attributes is often required to arrive at satisfactory
solutions. For this reason, a compensative meaning is explored next, with the purpose of finding a compromise over the
positive and non-negative intensities of preference.

6.3. Compensative meaning

Focusing on a compromise solution, the meaning of R  requires a compensating or averaging aggregation operator. For
this purpose, take an aggregation operator C = C v , such that 0 < C v (0, 1) = C v (1, 0) < 1. In a general sense, this type of
aggregation allows some interaction among the reinforcing concepts, being a compromise solution in between the limit
solutions of (27) and (28).
Such a compensating operator can be modeled by means of the Choquet integral C v , widely used in decision literature,
generalizing the weighted arithmetic mean (see e.g. [10,15,22]). Given the set X = {x1 , x2 }, ∀x1 , x2 ∈ [0, 1], the discrete
Choquet integral can be defined with respect to a non-additive set function or fuzzy measure v : 2 X → [0, 1], such that,


2
C v (x1 , x2 ) = v ( X (i ) )(x(i ) − x(i −1) ),
i =1

where (·) indicates a permutation of X , such that x(0) = 0 ≤ x(1) ≤ x(2) , and X (1) = {(1), (2)} and X (2) = {(2)}. For example,
|X|
the uniform fuzzy measure v ( X ) = 2 , can be used to compute the standard arithmetic mean.
It follows that the compensative dominance fuzzy relation is defined ∀a, b ∈ A2 , such that,
v
R (a, b) = C v (μ Q (a, b), A (μ V (a, b))). (29)
Applying the same voting procedure as with R  and R  , ∀a, b ∈ A, a vote for the dominance of a over b is here
s t
v v
conceeded whenever R  (a, b) > R  (b, a). Then, alternatives are ranked according to their total number of votes under the
compensative order, based on (29) and a compromise on the votes in favor of, and not against dominance, due to the intensity
of preference on desire and/or non-rejection. This weak order is denoted by  v .
Therefore, given an attitude towards preference (weak or strict), together with the type of semantic opposition (antonym
or antagonist), ∀a, b ∈ A, the dominance fuzzy relation (26) can be computed, estimating the number of votes supporting
the preference of a ∈ A over all other alternatives b = a ∈ A. Then, different orders are obtained under a disjunctive or
compensating meaning, respectively given by s (27) and  v (29), whose significance (24) can be measured with respect
to t (28), learning the order ∗ = t with maximal significance.
As a result, the conjunctive order t stands as the most significant decision, at the expense of being too strong in
its aggregation procedure. Hence, focusing on a compensating solution, decision support can be offered in the form of a
 C. Franco et al. / International Journal of Approximate Reasoning 86 (2017) 80–91 89

Table 2
Weak and strict preference/aversion intensities.

μQ μ Q −1 μV μ V −1 p p −1 z z−1
(a, b) 0.5 0.9 0.6 0.7 0.1 0.5 0.3 0.4
(a, c ) 0.9 0.4 0.3 0.7 0.6 0.1 0.3 0.7
(a, d) 0.3 0.6 0.4 0.4 0.3 0.6 0.4 0.4
(b, c ) 0.1 0.3 0.1 0.2 0.1 0.3 0.1 0.2
(b, d) 0.8 0.3 0.9 0.9 0.7 0.2 0.1 0.1
(c , d) 0.6 0.9 0.5 0.5 0.1 0.4 0.5 0.5

negotiated decision among the P–A positive and negative perceptions. In this way, it is possible to learn the compensating
solution that maximizes its significance for an optimal value of λ, by computing fuzzy reinforcement (26) over a set of
values associated to the antonym/antagonist λ-complement operator (25). The overall methodology is illustrated in the
following example.

7. Learning preferences for recommendation systems

Consider the specific problem of looking for a place to live on rent (following Examples 1 and 2). The options are
restricted to a given city, looking for alternative houses or apartments, studying them while thinking on the pros and cons
of choosing one or the other, and trying to arrive at a satisfactory solution for an adequate and decent home. In a general
sense, the decision is perceived based on its negative and the positive aspects, where the positive refers to the attributes
that can be associated to the desirable aspects, and the negative refers to the rejectable ones. Think e.g. on common
aspects related to the price: while they can be perceived as negative, the cheaper the better, they can also be perceived as
positive, for a fair price indicates good quality and may offer even more confidence than a suspiciously low price (commonly
associated to scams). After studying the market and the available alternatives, people articulate their preferences around
plausible candidates homes, identifying the one that fits better to their desires and needs. This process can be aided by a
recommendation system, which could be implemented as an open-access website.5
Suppose the website needplace2live.com exists, offering support on housing options by replicating the individual’s decision
process and identifying a priority order with maximum significance. As input data from the users, it is required that they
introduce information on their attitudes, whether their demands are weak or strict, and their specific desires and needs
on the maximum price they are willing to pay, pmax, the greatest distance they are willing to live away from the city
center, dmax, and the minimum area that they are looking for, mina. Then, the system offers the service of finding useful
recommendations, presenting them in decreasing order of preference.
Given a database containing all the available information on the housing options (such as the price, distance/location and
area), the system goes through every option verifying if they fulfill the specific requirements introduced by the user. This
first filtering process is intended to identify a small set of candidate alternatives, thus reducing the initial complexity of the
following pairwise comparison process (which requires a number of 2(|A|2 − |A|) pairwise comparisons). In this way, for
every pair (a, b) ∈ A2 , the desirability for a over b increases with a relatively lower (greater) price and distance (size), over
an acceptance fuzzy area around the pmax and dmax (mina) reference values; while the rejectability for a over b increases
with a relatively higher (lower) price and distance (size), over a rejectance fuzzy area around the pmax and dmax (mina)
reference values.6 The process obtains the fuzzy weak P–A intensities (μ Q , μ Q −1 ) and (μ V , μ V −1 ), as shown in Table 2.
Having filtered four candidate alternatives, the intensities of weak preference and aversion are computed, together with
the strict preference and aversion values for the three limit solutions of t-norms T L , T M and T p (see Table 2 for the weak
and strict (p , z = T M ) preference and aversion values. The other limit solutions for p , z are omitted as they all obtain the
same ordering results). Then, as negative perceptions weigh more than positive ones under general conditions of uncer-
tainty (see e.g. [22]), the opposition operator is taken as an antagonist operator, such that N = 1 − x and λ ∈ (−1, 0), and
preferences are computed for every meaning of R  (27)–(29). The system identifies the recommended compromise solution
with greater significance, given for the optimal λ value (see Table 3 and Table 4, respectively showing the estimations for
preference orders under a weak and a strict attitude, including the maximal-significant solution together with some other
λ-based estimations).
In Table 3, the different λ-based results for a weak attitude are shown. Based on the number of votes (#) for every
alternative, a ranking is built allowing ties among different alternatives (the position in the ranking is given by θ ). It can
be seen that, for all values of λ ∈ (−1, 0), the ranking under a compensative meaning (29) obtains a tie among the most
preferred alternatives b and d, while b is ranked first under a disjunctive meaning (27). On the other hand, under the more
demanding conjunctive meaning (28), the ranking depends on the particular value of λ, resulting in different significance
scores (RS) for both the disjunctive and compensative meanings. On the other hand, under the strict attitude of Table 4,

5
We could also imagine a mobile application, capturing information on house-searching behavior through a tracking system, and automatically gathering
the user’s revealed preferences for suggesting recommendations.
6
P–A intensities could even be measured by means of a brain scanner, or some type of virtual tracking device, presenting images of the housing options
to the user and estimating the positive and negative units of pleasure and aversion that each option stimulates in the brain.
90 C. Franco et al. / International Journal of Approximate Reasoning 86 (2017) 80–91

Table 3
Weak attitude preference orders, with the number of votes (#) and ranking position (θ ) of the
alternatives under the conjunctive (t), disjunctive (s) and compensative (v) solutions, and their
relative significance (RS).

λ λ
− 0.8 (#, θ)t (#, θ)s (#, θ) v − 0.6 (#, θ)t (#, θ)s (#, θ) v
a (1, 2) (1, 2) (1, 2) (1, 2) (1, 2) (1, 2)
b (2, 1) (3, 1) (2, 1) (1, 2) (3, 1) (2, 1)
c (1, 2) (0, 3) (1, 2) (1, 2) (0, 3) (1, 2)
d (2, 1) (1, 2) (2, 1) (2, 1) (1, 2) (2, 1)

RS 1 0.88 1 1 0.81 0.94

λ λ
− 0.4 (#, θ)t (#, θ)s (#, θ) v − 0.2 (#, θ)t (#, θ)s (#, θ) v
a (2, 1) (1, 2) (1, 2) (2, 1) (1, 2) (1, 2)
b (0, 3) (3, 1) (2, 1) (0, 3) (3, 1) (2, 1)
c (1, 2) (0, 3) (1, 2) (1, 2) (0, 3) (1, 2)
d (2, 1) (1, 2) (2, 1) (1, 2) (1, 2) (2, 1)

RS 1 0.69 0.81 1 0.75 0.75

Table 4
Strict attitude preference orders.

λ λ
−1 (#, θ)t (#, θ)s (#, θ) v (−1, 0] (#, θ)t (#, θ)s (#, θ) v
a (1, 2) (0, 1) (1, 2) (1, 2) (2, 1) (1, 2)
b (2, 1) (0, 1) (2, 1) (2, 1) (1, 2) (2, 1)
c (1, 2) (0, 1) (1, 2) (1, 2) (0, 3) (1, 2)
d (2, 1) (0, 1) (2, 1) (2, 1) (0, 3) (2, 1)

RS 1 0.88 1 1 0.69 1

the conjunctive and compensative meanings entail the same ranking for all values of λ ∈ (−1, 0), where b and d are ranked
first, while the disjunctive meaning obtains different orders for λ = −1, having absolute indifference, and for λ ∈ (−1, 0],
being a the most preferred one.
As a result, following the compensating solution ((#, θ) v ) with maximum significance, b and d are both assigned the first
position, followed by a and c. This ranking is identified under a weak attitude, for a λ value of −0.8, and it is confirmed
by the results associated to the strict attitude, where the same compensative order holds for any value of an antagonistic
λ ∈ (−1, 0]. In case a total order is required, a solution can be found according to some relevant argument that allows
breaking the tie between b and d. Thus, the associated disjunctive solution can be examined, as it can be understood as a
solution not complying with the significant one. It is then observed that under a weak attitude, b is ranked over d with
significance 0.88, while for a strict attitude, b is ranked over d with a significance of 0.69. Hence, forcing a total order, d is
placed in second place, followed by a and then by c.

8. Final comments

Preference semantics have been applied to the problem of preference learning, building weak orders from the fuzzy re-
inforcement of opposite paired preferences. Based on dominance fuzzy relations and different attitudes towards preference,
positive and negative perceptions are weighed according to their opposite character, and different pieces of evidence are
aggregated for preference learning under different levels of significance.
For future research, the formulation of different solutions for fuzzy (paired) preference structures remains to be explored
under other types of aggregation operators, which may not satisfy the standard properties of associativity, commutativity
or monotonicity. Besides, the construction of total orders should be explored in more detail, further refining weak orders,
and interval valued preferences should be considered, extending the paired fuzzy frame for the representation of imprecise
fuzzy preferences, learning from interval-valued or type-II fuzzy reinforcement. Finally, on an applied level, the proposed
preference learning technique by fuzzy reinforcement should be tested on real data, measuring the performance of the
results together with their respective degrees of significance.

Acknowledgements

This research has been partially supported by the Government of Spain (grant TIN2015-66741-P), the Government of
Madrid (grant S2013/ICE-2845, CASI-CAM-CM), and by the Danish Industry Foundation.
 C. Franco et al. / International Journal of Approximate Reasoning 86 (2017) 80–91 91

References

[1] A. Bechara, The role of emotion in decision-making: evidence from neurological patients with orbitofrontal damage, Brain Cogn. 55 (2004) 30–40.
[2] H. Bustince, M. Pagola, R. Mesiar, E. Hüllermeier, F. Herrera, Grouping, overlap and generalized bi-entropic functions for fuzzy modeling of pairwise
comparisons, IEEE Trans. Fuzzy Syst. 20 (2012) 405–415.
[3] J.T. Cacioppo, G.G. Berntson, Relationship between attitudes and evaluative space: a critical review, with emphasis on the separability of positive and
negative substrates, Psychol. Bull. 115 (1994) 401–423.
[4] D. Dubois, H. Prade, An introduction to bipolar representations of information and preference, Int. J. Intell. Syst. 23 (2008) 866–877.
[5] J. Fodor, M. Roubens, Fuzzy Preference Modelling and Multicriteria Decision Support, Kluwer Academic Publishers, Dordrecht, 1994.
[6] C. Franco, J. Montero, J.T. Rodríguez, A fuzzy and bipolar approach to preference modeling with application to need and desire, Fuzzy Sets Syst. 214
(2013) 20–34.
[7] C. Franco, J.T. Rodríguez, J. Montero, An ordinal approach to computing with words and the preference-aversion model, Inf. Sci. 258 (2014) 239–248.
[8] C. Franco, J.T. Rodríguez, J. Montero, Building the meaning of preference from logical paired structures, Knowl.-Based Syst. 83 (2015) 32–41.
[9] F. Grabenhorst, E.T. Rolls, Ch. Margot, M.A. da Silva, M.I. Velazco, How pleasant and unpleasant stimuli combine in different brain regions: odor
mixtures, J. Neurosci. 27 (2007) 13532–13540.
[10] M. Grabisch, Ch. Labreuche, A decade of application of the Choquet and Sugeno integrals in multi-criteria decision aid, Ann. Oper. Res. 175 (2010)
247–290.
[11] K. Kaplan, On the ambivalence-indifference problem in attitude theory and measurement: a suggested modification of the semantic differential tech-
nique, Psychol. Bull. 77 (1972) 361–372.
[12] L. Lee, O. Amir, D. Ariely, In search of the Homo Economicus: cognitive noise and the role of emotion in preference consistency, J. Consum. Res. 36
(2009) 173–187.
[13] J. Montero, H. Bustince, C. Franco, J.T. Rodríguez, D. Gómez, M. Pagola, J. Fernandez, E. Barrenechea, Paired structures in knowledge representation,
Knowl.-Based Syst. 100 (2016) 50–58.
[14] J. Montero, J. Tejada, V. Cutello, A general model for deriving preference structures from data, Eur. J. Oper. Res. 98 (1997) 98–110.
[15] Y. Narukawa, V. Torra, Fuzzy measures and Choquet integral on discrete spaces, in: B. Reusch (Ed.), Computational Intelligence, Theory and Applications,
Springer, Berlin, 2004, pp. 573–581.
[16] J. O’Doherty, M. Kringelback, E. Rolls, J. Hornak, C. Andrews, Abstract reward and punishment representations in the human orbitofrontal cortex, Nat.
Neurosci. 4 (2001) 95–102.
[17] Ch. Osgood, G. Suci, P. Tannenbaum, The Measurement of Meaning, University of Illinois Press, Urbana, 1958.
[18] J.T. Rodríguez, C. Franco, D. Gómez, J. Montero, Paired structures, imprecision types and two-level knowledge representation by means of opposites,
Adv. Intell. Syst. Comput. 401 (2016) 3–15.
[19] B. Roy, Partial preference analysis and decision-aid: the fuzzy outranking relation concept, in: D. Bell, R. Keeney, H. Raiffa (Eds.), Conflicting Objectives
in Decisions, Wiley and Sons, New York, 1977, pp. 40–75.
[20] M. Sugeno, Fuzzy measures and fuzzy integrals: a survey, in: M.M. Gupta (Ed.), Fuzzy Automata and Decision Processes, North-Holland, Amsterdam,
1977, pp. 89–102.
[21] E. Trillas, C. Moraga, S. Guadarrama, S. Cubillo, E. Castineira, Computing with antonyms, Stud. Fuzziness Soft Comput. 217 (2007) 133–153.
[22] A. Tversky, D. Kahneman, Advances in prospect theory: cumulative representation of uncertainty, J. Risk Uncertain. 5 (1992) 297–323.
[23] M. Viinikainen, P. Jääskeläinen, Y. Alexandrov, M.H. Balk, T. Autti, M. Sams, Nonlinear relationship between emotional valence and brain activity:
evidence of separate negative and positive valence dimensions, Hum. Brain Mapp. 31 (2010) 1030–1040.
[24] B. van der Walle, B. de Baets, E. Kerre, Characterizable fuzzy preference structures, Ann. Oper. Res. 80 (1998) 105–136.