Metrizable ordinal proximity measures and their aggregation

Metrizable ordinal proximity measures and their aggregation Jos´e Luis GARC´ IA-LAPRESTA PRESAD Research Group, BORDA Research Unit, IMUVA, Departamento de Econom´ıa Aplicada, Universidad de Valladolid, Spain Raquel GONZ´ ALEZ DEL POZO PRESAD Research Group, IMUVA, Departamento de Econom´ıa Aplicada, Universidad de Valladolid, Spain David P´ EREZ-ROM´ AN PRESAD Research Group, BORDA Research Unit, Departamento de Organizaci´on de Empresas y Comercializaci´on e Investigaci´on de Mercados, Universidad de Valladolid, Spain Abstract Ordered qualitative scales formed by linguistic terms are frequently used for evaluating sets of alternatives in different decision-making problems. These scales are usually implicitly considered as uniform in the sense that the psychological proximity between consecutive terms is perceived as identical. However, sometimes agents can perceive different proximities between the linguistic terms of the scale, and an appropriate method is required for aggregating these perceptions. In this paper we introduce the notion of metrizable ordinal proximity measure, discuss some aggregation procedures and propose a method based on metrics for aggregating experts’ opinions on proximities between linguistic terms on ordered qualitative scales. Keywords: group decision-making; qualitative scales; ordinal proximity measures; judgment aggregation. Email addresses: [email protected] (Jos´e Luis GARC´ IA-LAPRESTA), [email protected] (Raquel GONZ´ ALEZ DEL POZO), [email protected] (David P´ EREZ-ROM ´ AN) Preprint submitted to Information Sciences March 12, 2018 1. Introduction In many real problems in various disciplines (Economics, Management, Marketing, Psychology, Sociology, and Tourism, among others) the use of Ordered Qualitative Scales (OQSs) formed by linguistic terms is quite widespread as a way of collecting the opinions given by a group of agents concerning a number of alternatives. It is usually implicitly assumed that these OQSs are uniform, in the sense that the psychological proximity between all the pairs of consecutive terms of the OQS is perceived as identical. Some OQSs with an odd number of linguistic terms are devised as uniform by fixing a neutral central linguistic term and arranging the other terms around it symmetrically (e.g. the 5-term OQS {‘very bad’, ‘bad’, ‘regular’, ‘good’, ‘very good’}). However, not all OQSs with an odd number of linguistic terms are necessarily uniform1. OQSs with an even number of linguistic terms do not have a neutral central linguistic term. Consequently, it is easier to find non-uniform OQSs in even cases than in odd ones. For instance, the 4-term OQS {‘reject’, ‘major revision’, ‘minor revision’, ‘accept’}used by some scientific journals in evaluating papers can be understood as non-uniform (see Garc´ıa-Lapresta and P´erez-Rom´an [11, 2.4] for empirical evidence). Another example of an OQS with an even number of linguistic terms is provided by Balinski and Laraki [1] when they introduce the Majority Judgment voting system. These authors consider the 6-term OQS {‘to reject’, ‘poor’, ‘acceptable’, ‘good’, ‘very good’, ‘excellent’}for evaluating candidates in political elections. Again, it is not clear that this scale is uniform. There are numerous contributions in the literature that deal with nonuniform OQSs that follow fuzzy techniques involving cardinal approaches (see Herrera-Viedma and L´opez-Herrera [17], Herrera et al. [16], among others). Other authors tackle non-uniform OQSs by means of ordinal ranges avoiding either cardinal or ordinal measurements of the proximities between linguistic terms (see Franceschini et al. [9]). To deal with non-uniform OQSs in a purely ordinal way via psychological proximities between linguistic terms of OQSs, Garc´ıa-Lapresta and P´erez- Rom´an [11] introduce the notion of ordinal proximity measure. This new ap- 1Question HS140 in the European Union Statistics on Income and Living Conditions (EUSILC) survey conducted by Eurostat uses the 3-term OQS {‘a heavy burden’, ‘somewhat a burden’, ‘not burden at all’}for asking individuals about the financial burden of total housing cost. In this OQS the central term is not neutral, so it is not clear that the scale is uniform. 2 proach has some similarities with difference measurement within classical measurement theory (see Krantz et al. [18, chapter 4] and Roberts [26, section 3.3]), and with non-metric multidimensional scaling, where only the ranks of the psychological distances or proximities are known (see Bennett and Hays [2], Shepard [27], Coombs [4], Kruskal and Wish [19], Cox and Cox [5] and Borg and Groenen [3, chapter 9], among others). Given an OQS, the behavior of some ordinal proximity measures is better than others. This is why in this paper we introduce metrizable ordinal proximity measures. They behave as if the ordinal comparisons between the terms of an OQS were managed through a linear metric on the OQS2. For instance, taking into account the above mentioned OQS {‘reject’, ‘major revision’, ‘minor revision’, ‘accept’}, saying that ‘minor revision’ is closer to ‘accept’ than to ‘major revision’ is equivalent to saying that the numerical distance between ‘minor revision’ and ‘accept’ is shorter than the numerical distance between ‘minor revision’ and ‘major revision’, whatever the corresponding linear metric that generates the ordinal proximity measure3. Deciding whether a given OQS is uniform or not, and in the latter case determining what the ordinal proximities are between the terms of the scale is an important issue for constructing metrizable ordinal proximity measures. This problem can be solved by a well-informed expert. However, sometimes the corresponding metrizable ordinal proximity measure is not easy to construct. In this paper we present an algorithm which generates a metrizable ordinal proximity measure by means of appropriate sequences of questions about the proximities between the terms of the scale. If a group of experts is asked about these proximities, they may have different opinions and different metrizable ordinal proximity measures may therefore emerge. In these situations, it is advisable to find a collective metrizable ordinal proximity measure that represents individual opinions as faithfully as possible. Therefore, an appropriate aggregation procedure is needed. This aggregation problem is not trivial, and different procedures can generate different outcomes 2A linear metric on an OQS is a metric satisfying the requirement that the distance between two terms must be the sum of the distance between the first term and any intermediate term between the two given terms and the distance between that intermediate term and the second given term. 3Notice that this approach is similar to classical preference modeling of weak orders, where one alternative is preferred to another if and only if the utility of the first alternative is greater than the utility of the second, whatever the corresponding numerical utility function that represents the weak order may be. 3 and even inconsistencies. To avoid these problems, we propose some solutions in this paper in the setting of judgment aggregation theory (see Dietrich and List [6], List [22], Mongin [24] and Grossi and Pigozzi [15], among others). In particular, we have devised a weighted-metric-based procedure that provides the metrizable ordinal proximity measure that minimizes the sum of distances (square distances) between itself and the metrizable ordinal proximity measures of the experts. This procedure solves the aforementioned problems. It is illustrated with two real case studies. The rest of the paper is organized as follows. Section 2 introduces, analyzes and generates metrizable ordinal proximity measures. Section 3 studies the problem of how to aggregate experts’ opinions through certain voting systems in order to generate metrizable ordinal proximity measures, and the inconsistencies that can result. Section 4 presents some proposals for avoiding inconsistencies. Section 5 contains some concluding remarks. 2. Ordinal proximity measures Consider an OQS L={l1, . . . , lg}whose terms are arranged from worst to best, with granularity at least 3, i.e., g≥3. In order to recall the notion of ordinal proximity measure on L, introduced by Garc´ıa-Lapresta and P´erez- Rom´an [11], we shall use a linear order ∆ = {δ1, . . . , δh}, with δ1 · · ·  δh, for representing different degrees of proximity (with no meaning) among the terms of L, being δ1and δhthe maximum and minimum degrees, respectively. As usual in the setting of linear orders, δrδsmeans δrδsor δr=δs; δr≺δsmeans δsδr; and δrδsmeans δr≺δsor δr=δs. Definition 1. ([11]) An ordinal proximity measure (OPM) on Lwith values in ∆ is a mapping π:L2−→ ∆, where π(lr, ls) = πrs means the degree of proximity between lrand ls, satisfying the following conditions: 1. Exhaustiveness: For every δ∈∆, there exist lr, ls∈ L such that δ=πrs . 2. Symmetry:πsr =πrs , for all r, s ∈ {1, . . . , g}. 3. Maximum proximity:πrs =δ1⇔r=s, for all r, s ∈ {1, . . . , g}. 4. Monotonicity:πrs πrt and πst πrt , for all r, s, t ∈ {1, . . . , g}such that r < s < t. We note that the previous conditions are independent (see Garc´ıa-Lapresta and P´erez-Rom´an [11, Prop. 1]). Every OPM π:L2−→ ∆ can be represented by a g×gsymmetric matrix with coefficients in ∆, where the elements in the main diagonal are πrr =δ1, 4 r= 1, . . . , g:         π11 · · · π1s· · · π1g · · · · · · · · · · · · · · · πr1· · · πrs · · · πrg · · · · · · · · · · · · · · · πg1· · · πgs · · · πgg         . This matrix is called the proximity matrix associated with π. Taking into account the conditions appearing in Definition 1, it is only necessary to show the upper half proximity matrix           δ1π12 π13 · · · π1(g−1) π1g δ1π23 · · · π2(g−1) π2g · · · · · · · · · δ1π(g−1)g δ1           . We note that the minimum proximity between linguistic terms is only reached when comparing the extreme linguistic terms: πrs =δh⇔(r, s)∈ {(1, g),(g, 1)} (see Garc´ıa-Lapresta and P´erez-Rom´an [11, Prop. 2]). We also note that the cardinality of ∆ is located between the granularity of Land a polynomial of degree 2 of that granularity (see Garc´ıa-Lapresta and P´erez-Rom´an [11, Prop. 4]): g≤h≤g·(g−1) 2+ 1. 2.1. Linear metrics The notion of metrizable OPM is based on linear metrics on OQSs. So, we first introduce the notion of linear metric in the setting of OQSs. Definition 2. Alinear metric on an OQS Lis a mapping d:L2−→ R satisfying the following conditions for all r, s, t ∈ {1, . . . , g}: 1. Positiveness:d(lr, ls)≥0. 2. Identity of indiscernibles:d(lr, ls)=0 ⇔r=s. 3. Symmetry:d(ls, lr) = d(lr, ls). 4. Linearity:r < s < t ⇒d(lr, lt) = d(lr, ls) + d(ls, lt). As shown in the following remark, it is possible to generate a linear metric from the distances between consecutive terms of the OQS. 5 Remark 1. Given d(lr, lr+1) = ρr>0 for r= 1, . . . , g −1, there exists a unique linear metric on L,d:L2−→ R, satisfying the given conditions: d(lr, lr+2) = d(lr, lr+1) + d(lr+1, lr+2) = ρr+ρr+1 for every r∈ {1, . . . , g −2}. Iterating this process, we have d(lr, lr+t) = ρr+ρr+1 +· · · +ρr+t−1for all r, t ∈ {1, . . . , g −1}such that r+t≤g. It suffices to define d(ls, lr) = d(lr, ls) and d(lr, lr) = 0 for all r, s ∈ {1, . . . , g}. We now justify that the family of linear metrics is a proper subset of the family of metrics. Proposition 1. Every linear metric d:L2−→ Rsatisfies the triangle inequality, i.e., d(lr, lt)≤d(lr, ls) + d(ls, lt)for all r, s, t ∈ {1, . . . , g}. Proof. There exist 6 cases. 1. r≤s≤t:d(lr, lt) = d(lr, ls) + d(ls, lt). 2. r≤t≤s:d(lr, lt)≤d(lr, lt) + d(lt, ls) = d(lr, ls)≤d(lr, ls) + d(ls, lt). 3. s≤r≤t:d(lr, lt)≤d(lr, lt) + d(ls, lr) = d(ls, lt)≤d(lr, ls) + d(ls, lt). 4. s≤t≤r:d(lr, lt)≤d(lr, lt) + d(ls, lt) = d(ls, lr)≤d(lr, ls) + d(ls, lt). 5. t≤r≤s:d(lr, lt)≤d(lr, lt) + d(lr, ls) = d(lt, ls)≤d(lr, ls) + d(ls, lt). 6. t≤s≤r:d(lr, lt) = d(lt, lr) = d(lt, ls) + d(ls, lr) = d(lr, ls) + d(ls, lt).  Consequently, every linear metric on Lis a metric. The reciprocal is not true, as shown in the following remark. Remark 2. There exist metrics on Lthat are not linear metrics. For instance, consider L={l1, l2, l3}and d:L2−→ Rthe mapping defined as d(l1, l1) = d(l2, l2) = d(l3, l3) = 0, d(l1, l2) = d(l2, l1) = 2, d(l2, l3) = d(l3, l2) = 3 and d(l1, l3) = d(l3, l1) = 4. It is easy to check that dis a metric on L. However, it is not linear: 4 = d(l1, l3)< d(l1, l2) + d(l2, l3) = 5. 2.2. Metrizable OPMs Before introducing the notion of metrizable OPM, we justify that every linear metric on an OQS defines in a natural way an OPM. Proposition 2. Let d:L2−→ Rbe a linear metric. If π:L2−→ ∆is an exhaustive mapping defined as πrs πtu ⇔d(lr, ls)< d(lt, lu), then πis an OPM on L. Proof. 1. Exhaustiveness: By hypothesis. 2. Symmetry: By symmetry of d. 6 3. Maximum proximity: Since πis exhaustive, δ1=πrs for some r, s ∈ {1, . . . , g}. If r6=s, then d(lr, lr)=0< d(lr, ls). Then, πrr πrs =δ1, that is a contradiction. Consequently, r=s, i.e., δ1=πrr . In order to prove that πtt =δ1for every t∈ {1, . . . , g}, suppose that πtt 6=πrr for some t∈ {1, . . . , g}. If πtt πrr , then 0 = d(lt, lt)< d(lr, lr), that is a contradiction. Analogously, from πrr πtt we obtain a contradiction. 4. Monotonicity: Consider r, s, t ∈ {1, . . . , g}such that r < s < t. Since d(lr, lt) = d(lr, ls) + d(ls, lt), d(lr, ls)>0 and d(ls, lt)>0, we have d(lr, lt)> d(lr, ls), i.e., πrs πrt , and d(lr, lt)> d(ls, lt), i.e., πst πrt .  Definition 3. An OPM π:L2−→ ∆ is metrizable if there exists a linear metric d:L2−→ Rsuch that πrs πtu ⇔d(lr, ls)< d(lt, lu), for all r, s, t, u ∈ {1, . . . , g}. We say that πis generated by d. Remark 3. If π:L2−→ ∆ is a metrizable OPM generated by a linear metric d:L2−→ Rand d0:L2−→ Ris defined as d0(lr, ls) = λ·d(lr, ls) for some λ > 0, then d0is a linear metric and πis also generated by d0. However, a metrizable OPM can be generated by non proportional linear metrics. Consider L={l1, l2, l3}and π:L2−→ ∆ the OPM associated with the matrix A32 (see Subsection 2.3). Let d:L2−→ Rbe the mapping defined as d(l1, l1) = d(l2, l2) = d(l3, l3) = 0, d(l1, l2) = d(l2, l1) = 2, d(l2, l3) = d(l3, l2) = 1 and d(l1, l3) = d(l3, l1) = 3. Let d0:L2−→ Rbe the mapping defined as d0(l1, l1) = d0(l2, l2) = d0(l3, l3) = 0, d0(l1, l2) = d0(l2, l1) = 3, d0(l2, l3) = d0(l3, l2) = 1 and d0(l1, l3) = d0(l3, l1) = 4. It is easy to see that d and d0are linear metrics and that πis simultaneously generated by dand d0. Nevertheless, d06=λ·dfor every λ > 0. 2.3. Constructing metrizable OPMs We now show the proximity matrices associated with all the metrizable OPMs for g= 3,4. The subindices of the matrices A’s correspond to the subindices of the δ’s appearing in the coefficients just over the main diagonal, π12, π23, . . . , π(g−1) g. These ordinal proximities correspond to the comparisons between all the pairs of consecutive linguistic terms. For g= 3 there are three OPMs and all of them are metrizable. If an expert declares π12 =π23 ,π12 π23 or π12 ≺π23 , then the matrix associated with the corresponding metrizable OPM will be A22 (see Figure 1), A23 (see Figure 7 2) or A32 (see Figure 3), respectively: A22 =   δ1δ2δ3 δ1δ2 δ1   , A23 =   δ1δ2δ4 δ1δ3 δ1   , A32 =   δ1δ3δ4 δ1δ2 δ1   . Figure 1: Ordinal proximity measure with associated matrix A22. l1l2 δ2l3 δ2 Figure 2: Ordinal proximity measure with associated matrix A23. l1l2 δ2l3 δ3 Figure 3: Ordinal proximity measure with associated matrix A32. l1l2 δ3l3 δ2 For g > 3, experts may find some difficulties in directly constructing the OPM (or the associated proximity matrix) that reflects their own opinions about the proximities between the terms of the OQS. This can be done through appropriate sequences of questions, the answers to which lead to a metrizable OPM. For g= 4 there are 51 OPMs, but only 25 of them are metrizable. Figure 4 contains an algorithm for g= 4 that guides the sequence of questions, depending on the answers provided by an expert, in order to obtain one of the 25 metrizable OPMs. This algorithm starts by asking the expert about ordinal proximities π12 and π23 . The next question differs depending on whether one of these ordinal proximities is greater than the other or they are the same. The procedure continues with similar questions comparing the ordinal proximities between the remaining pairs of terms of the OQS until the OPM is obtained. It is interesting to note that 4, 12 and 9 matrices are achieved after answering 2, 3 and 4 questions, respectively (3.2 questions on average). Unfortunately, for g > 4 the complexity of the algorithm dramatically increases. The associated proximity matrices of these 25 metrizable OPMs are: 8 π12 versus π23 π12 π23 π12 π34 —— π12 =δ2 π23 π34 —— π23 =δ3 π13 π34 —— π13 =δ4 π34 =δ5 A235 π13 =π34 —— π13 =π34 =δ4 A0 234 π13 ≺π34 —— π13 =δ5 π34 =δ4 A234 π23 =π34 —— π23 =π34 =δ3 A233 π23 ≺π34 —— π23 =δ4 π34 =δ3 A243 π12 =π34 —— π12 =π34 =δ2 π23 =δ3 A232 π12 ≺π34 —— π12 =δ3 π23 =δ4 π34 =δ2 A342 π12 =π23 π12 π34 —— π12 =π23 =δ2 π13 π34 —— π13 =δ3 π34 =δ4 A224 π13 =π34 —— π13 =π34 =δ3 A0 223 π13 ≺π34 —— π13 =δ4 π34 =δ3 A223 π12 =π34 —— π12 =π23 = =π34 =δ2 A222 π12 ≺π34 —— π12 =π23 =δ3 π34 =δ2 A332 π12 ≺π23 π23 π34 —— π23 =δ2 π12 π34 —— π12 =δ3 π13 π34 —— π13 =δ4 π34 =δ5 A325 π13 =π34 —— π13 =π34 =δ4 A0 324 π13 ≺π34 —— π13 =δ5 π34 =δ4 A324 π12 =π34 —— π12 =π34 =δ3 A323 π12 ≺π34 —— π34 =δ3 π12 π24 —— π12 =δ4 π24 =δ5 A423 π12 =π24 —— π12 =π24 =δ4 A0 423 π12 ≺π24 —— π12 =δ5 π24 =δ4 A523 π23 =π34 —— π23 =π34 =δ2 π12 π24 —— π12 =δ3 π24 =δ4 A322 π12 =π24 —— π12 =π24 =δ3 A0 322 π12 ≺π24 —— π12 =δ4 π24 =δ3 A422 π23 ≺π34 —— π23 =δ3 π34 =δ2 π12 π24 —— π12 =δ4 π24 =δ5 A432 π12 =π24 —— π12 =π24 =δ4 A0 432 π12 ≺π24 —— π12 =δ5 π24 =δ4 A532 Figure 4: Algorithm for g= 4. 9 3.3. Consistency For all r, s, t, u, v, w ∈ {1, . . . , g}the following conditions must be satisfied: πrs πtu ∧πtu πvw⇒πrs πvw (1) πrs πtu ∧πtu =πvw⇒πrs πvw (2) πrs =πtu ∧πtu πvw⇒πrs πvw (3) πrs =πtu ∧πtu =πvw⇒πrs =πvw (4) r<s<t ⇒πrs πrt (5) r<s<t ⇒πst πrt (6) Conditions (1), (2) and (3) refer to the linear order on ∆; (4) to the transitivity of = ; and (5) and (6) to the monotonicity of π. 3.4. A case study To show how some voting systems can be applied for generating an OPM from the opinions of a group of experts, we consider the data from the case study reported in Garc´ıa-Lapresta and P´erez-Rom´an [11, 2.4]. We now present the results of a survey conducted on 76 members of the Spanish Society for Fuzzy Logic and Technology (ESTYLF) about the degrees of proximity between the usual decisions of some journal editors (see Table 1). l1l2l3l4 Reject Major revision Minor revision Accept Table 1: Meaning of the linguistic terms. Table 2 contains the data obtained in the survey. We now present the social outcomes generated by simple majority and some qualified majorities when aggregating experts’ opinions about the ordinal proximities between pairs of linguistic terms. •Simple majority: π34 π23 π12 π24 π13 . This information allows us to assign the following degrees of proximity πrr =δ1π34 =δ2π23 =δ3π12 =δ4π24 =δ5π13 =δ6π14 =δ7 16 π12 versus π23 Number % π12 π23 27 35.5 π12 ≺π23 32 42.1 π12 =π23 17 22.4 π23 versus π34 Number % π23 π34 6 7.9 π23 ≺π34 69 90.8 π23 =π34 1 1.3 π12 versus π34 Number % π12 π34 10 13.2 π12 ≺π34 54 71.0 π12 =π34 12 15.8 π12 versus π24 Number % π12 π24 42 55.3 π12 ≺π24 18 23.7 π12 =π24 16 21.0 π13 versus π34 Number % π13 π34 0 0.0 π13 ≺π34 75 98.7 π13 =π34 1 1.3 π13 versus π24 Number % π13 π24 1 1.3 π13 ≺π24 53 69.7 π13 =π24 22 29.0 Table 2: Data of the survey. and the metrizable OPM with associated proximity matrix A432 =      δ1δ4δ6δ7 δ1δ3δ5 δ1δ2 δ1       . •Qualified majorities of thresholds q∈[0.5,0.552): π34 π12 =π23  π24 π13 . This information allows us to assign the following degrees of proximity πrr =δ1π34 =δ2π12 =π23 =δ3π24 =δ4π13 =δ5π14 =δ6 and the metrizable OPM with associated proximity matrix A332 =      δ1δ3δ5δ6 δ1δ3δ4 δ1δ2 δ1       . •Qualified majorities of thresholds q∈[0.697,0.710): π34 π12 =π23 = 17 π13 =π24 . Since π13 =π12 and π24 =π23 ,πviolates the monotonicity conditions (5) and (6) and, consequently, πis not an OPM. Nevertheless, the obtained ordinal proximities can be arranged in the following matrix       δ1δ3δ3δ4 δ1δ3δ3 δ1δ2 δ1       . •Qualified majorities of thresholds q∈[0.710,0.907): π12 =π23 ,π12 = π34 ,π34 π13 ,π34 π23 ,π12 =π24 and π13 =π24 . Since π34 =π12 , π12 =π23 and π34 π23 ,πviolates condition (4), the transitivity of = , and, consequently, it is not an OPM. In this case it is not possible to arrange the ordinal proximities obtained in any matrix. We now present the outcomes generated by some scoring rules when aggregating experts’ opinions on the proximities between pairs of linguistic terms. Consider the normalized scoring vector (1, s, 0), with 0 ≤s≤1. •s∈[0,0.2]: we obtain the metrizable OPM with associated proximity matrix A432. •s∈[0.3,0.5): we obtain the metrizable OPM with associated proximity matrix A332. •s∈(0.5,0.7]: π34 π23 =π12 =π24 . Since π24 =π23 ,πviolates the monotonicity condition (5) and, consequently, πis not an OPM. •s∈[0.8,0.9]: π12 =π23 ,π12 =π34 ,π34 π13 ,π34 π23 ,π12 =π24 and π13 =π24 . Since π34 =π12 ,π12 =π23 and π34 π23 ,πviolates condition (4), the transitivity of = , and, consequently, it is not an OPM. Again, it is not possible to arrange the obtained ordinal proximities in any matrix. All these inconsistencies may be considered as specific problems of judgment aggregation within social choice theory, in the sense that the aggregation of judgments over multiple interconnected issues may produce inconsistent outcomes (see, for instance, Grossi and Pigozzi [15]). Some proposals aimed at avoiding these inconsistencies are made in the next section. 18 4. Avoiding inconsistencies In order to avoid the inconsistencies shown in Section 3 when experts’ opinions are aggregated through voting systems, we now introduce some distancebased procedures. Two of them are related to the characterizations of the median and the mean, in the sense that the outcomes are the metrizable OPMs that minimize the sum of distances and squared distances, respectively, to the metrizable OPMs of the agents. First we introduce metrics on OPMs. 4.1. Distances between OPMs With P(L) we denote the set of OPMs on L. With M(L) we denote the set of metrizable OPMs on L. Definition 5. Let π1:L2−→ ∆1and π2:L2−→ ∆2be two OPMs, w: N−→ Ra weighting function such that w(1) = 1 ≥w(2) ≥ · · · ≥ w(g−1) >0, β: (∆1∪∆2)2−→ N∪ {0}the mapping defined as β(δi, δj) = |i−j|and S={(r, s)∈ {1, . . . , g −1}2|r+s≤g}. Then, Dwπ1, π2is defined as Dwπ1, π2=X (r,s)∈S w(s)·βπ1 r(r+s), π2 r(r+s).(7) Remark 7. An equivalent formulation of Eq. (7) is Dwπ1, π2= g−1 X r=1 βπ1 r(r+1), π2 r(r+1)+w(2) · g−2 X r=1 βπ1 r(r+2), π2 r(r+2)+ +w(3) · g−3 X r=1 βπ1 r(r+3), π2 r(r+3)+· · · Remark 8. Some simple examples of weighting functions are: 1. Power: w(s) = 1 sα, with α≥1. 2. Exponential: w(s) = 1 αs−1, with α > 1. 3. Linear: w(s)=1−1−α g−2·(s−1), with 0 < α < 1 (note that α=w(g−1)). Proposition 5. If w:N−→ Ris a weighting function such that w(1) = 1 ≥ w(2) ≥ · · · ≥ w(g−1) >0, then the mapping Dw:P(L)2−→ Rdefined from Eq. (7) is a metric. 19 Proof. Taking into account that w(s)>0 for every s∈ {1, . . . , g −1} and the mapping β:{δi|i∈N} −→ Rdefined as β(δi, δj) = |i−j|is a metric, we have that Dwπ1, π2is a positive linear combination of distances. Consequently, Dwis a metric.  Remark 9. For g= 4, let w:N−→ Rbe the weighting functions with power weights for α= 1, i.e., w(1) = 1, w(2) = 1 2and w(3) = 1 3, and α= 2, i.e., w(1) = 1, w(2) = 1 4and w(3) = 1 9, and exponential weights for α= 2, i.e., w(1) = 1, w(2) = 1 2and w(3) = 1 4, and α= 3, i.e., w(1) = 1, w(2) = 1 3and w(3) = 1 9. The closest metrizable OPM, with respect to Dw, to the non metrizable OPMs with associated proximity matrices A1 ijk and A2 ijk (see the Appendix) are just the ones associated with Aijk or A0 ijk (see Subsection 2.3). However, linear weighting functions do not follow this pattern. For instance, if α= 0.3, i.e., w(1) = 1, w(2) = 0.85 and w(3) = 0.7, then the closest metrizable OPM, with respect to Dw, to the non metrizable OPMs with associated proximity matrices A1 222 and A2 222 (see the Appendix) is not the one associated with A222, but the ones associated with A0 223 and A0 322, respectively. For this reason, on the sequel we do not consider linear weighting functions. In what follows we focus on metrizable OPMs. Consider a profile of metrizable OPMs (π1, . . . , πm)∈M(L)massociated with mexperts6. We are interested in finding a metrizable OPM that represents, as faithfully as possible, the opinions of the experts. As shown in Subsection 3.4, applying a voting system for generating the ordinal proximities between pairs of linguistic terms from the comparisons of the experts may generate inconsistencies. In the following subsections we propose different approaches for solving these problems. 4.2. Voting systems In spite of the inconsistency problems mentioned in Section 3, in some cases it is possible to use a voting system to obtain a matrix (πrs) that represents the ordinal proximities between pairs of linguistic terms7. If (πrs) corresponds to a metrizable OPM π∈M(L), then πis the outcome. Otherwise, once a 6Notice that experts are not required directly to provide metrizable OPMs. They only need to answer some simple questions. In particular, for g= 4, following the algorithm included in Figure 4, the corresponding metrizable OPMs are easily generated once between two and four questions have been answered. 7As shown in the examples provided in Subsection 3.4, such a matrix cannot exist if condition (4) is not satisfied. 20 metric Dwhas been fixed, we find8 {π∈M(L)| ∀π0∈M(L)Dw(π, π)≤Dw(π0, π)}, i.e., the solution of arg π∈M(L) min Dw(π, π). If this set contains more than one metrizable OPM, a tie-breaking procedure needs to be applied. One possibility is to set a sequence of weighting functions and apply them lexicographically. 4.3. Minimizing aggregated distances We now introduce a proposal for obtaining collective metrizable OPMs from the individual ones9. It is based on the characterizations of the statistical notions of median and mean. 4.3.1. The median It is well known that the medians of a list of numbers is the set of real numbers that minimize the sum of distances to the numbers of the list. In this way, the metrizable OPMs that minimize the sum of distances (for a fixed metric Dw) to the metrizable OPMs of the agents, π1, . . . , πm, can be said to be their medians. Thus, the medians of π1, . . . , πmare the elements of the following set medwπ1, . . . , πm= (π∈M(L)| ∀π0∈M(L) m X i=1 Dwπ, πi≤ m X i=1 Dwπ0, πi), i.e., the solution of medwπ1, . . . , πm= arg π∈M(L) min m X i=1 Dwπ, πi.(8) 8We extend Eq. (7) to matrices that do not necessarily correspond to OPMs. 9This proposal is related to the one provided by Grossi and Pigozzi [15, 4.3.3] in a specific problem of judgment aggregation. See also Eckert and Klamler [7] and Lang et al. [20]. 21 4.3.2. The mean It is also well known that the mean of a list of numbers is the real number that minimizes the sum of the squared distances to the numbers of the list. The metrizable OPM that minimizes the sum of squared distances (for a fixed metric Dw) to the metrizable OPMs of the agents, π1, . . . , πm, can be said to be its mean. Thus, the mean of π1, . . . , πmis the following set meanwπ1, . . . , πm= (π∈M(L)| ∀π0∈M(L) m X i=1 Dwπ, πi2≤ m X i=1 Dwπ0, πi2), i.e., the solution of meanwπ1, . . . , πm= arg π∈M(L) min m X i=1 Dwπ, πi2.(9) Notice that the sets of medians and means may have more than one metrizable OPM. In such cases, a tie-breaking procedure needs to be applied. As in Subsection 4.2, one possibility is to set a sequence of weighting functions and apply them lexicographically. Example 1. Table 3 shows the matrices associated with the OPMs of the 76 participants in the survey shown in Subsection 3.4 after the algorithm included in Figure 4 for g= 4 has been applied. If we consider Dwfor the weighting functions appearing in Remark 9, power weights for α= 1, i.e., w(1) = 1, w(2) = 1 2and w(3) = 1 3, and α= 2, i.e., w(1) = 1, w(2) = 1 4and w(3) = 1 9, and exponential weights for α= 2, i.e., w(1) = 1, w(2) = 1 2and w(3) = 1 4, and α= 3, i.e., w(1) = 1, w(2) = 1 3and w(3) = 1 9, then the median and the mean of π1, . . . , π76is the metrizable OPM with associated matrix A332, just the same as the one obtained when absolute majority and closed qualified majorities are applied in Subsection 3.4. Remark 10. The sums appearing in 4.3.1 and 4.3.2 can be changed for an aggregation function F(for instance an OWA operator, such as the median or a trimmed mean, a quasiarithmetic mean, etc.). Thus, (8) and (9) become arg π∈M(L) min FDwπ, π1, . . . , Dw(π, πm) and arg π∈M(L) min FDwπ, π12,...,(Dw(π, πm))2, 22 Matrix Frequency % A222 4 5.3 A223 1 1.3 A232 5 6.6 A243 6 7.9 A323 1 1.3 A332 12 15.8 A342 16 21.1 A432 11 14.5 A0 432 4 5.3 A532 16 21.1 76 100 Table 3: Matrices of the survey in Example 1. respectively. For instance, in Example 1 if the power weighting function for α= 1, i.e., w(1) = 1, w(2) = 1 2and w(3) = 1 3is considered, and the distances are aggregated through the trimmed mean that removes the three highest and the three lowest values, then the metrizable OPM with associated matrix A432 is obtained, which is just the same as the one obtained when simple majority is applied in Subsection 3.4. Notice that the metrizable OPMs with the highest frequency (16), those with associated proximity matrices A342 and A532, are not selected as the social outcome for any of the procedures considered. This is not surprising at all, since plurality rule10 could not faithfully represent individual opinions when there are more than two alternatives (see, for instance, Morales [25] –English translation in McLean and Urken [23]– and Laslier [21]). Example 2. In this example we apply the proposed procedure to a 4-term OQS used in the Trends in International Mathematics and Science Study (TIMSS). This study evaluates the home, community, school, and student factors associated with student achievement in mathematics and science at fourth and eighth grade levels in more than 60 countries. To that end, data is collected through questionnaires completed by students, parents, teachers, and school principals. The TIMSS has been conducted every four years since 1995. The latest data 10In this voting system each agent votes for only one alternative and the winners are the alternatives with most number of votes. 23 available is from TIMSS 2015 (see [28]). In this example, a 4-term OQS used in TIMSS 2015 to assess some school problems (arriving late at school, vandalism, cheating, classroom disturbances, etc.) is considered. The linguistic terms of the scale appear in Table 4. l1l2l3l4 Serious problem Moderate problem Minor problem Not a problem Table 4: Meaning of the linguistic terms. As in Example 1, we carried out an on-line survey as to the proximities between the terms of the scale. A total of 19 Spanish school principals who participated in TIMSS 2015 contributed to this survey. The algorithm included in Figure 4 for g= 4 is applied and the matrices associated with the OPMs of the 19 participants are then shown in Table 5. Matrix Frequency % A222 6 31.59 A232 1 5.26 A234 1 5.26 A322 2 10.53 A332 3 15.79 A342 1 5.26 A432 1 5.26 A532 4 21.05 19 100 Table 5: Matrices of the survey in Example 2. Considering Dwfor the same weighting functions as in Example 1, the metrizable OPM obtained using the median and the mean is the matrix A332. 5. Concluding remarks Given an OQS, determining whether the scale is uniform or not (and if not what the ordinal proximities between the linguistic terms of the scale are) is an important issue. Our proposal is to ask some experts their opinions about these ordinal proximities and aggregate these opinions to obtain a metrizable OPM on the OQS. This is not a trivial problem, since inconsistencies could appear (see Section 3). We propose various procedures for generating a metrizable OPM from the opinions of the experts, paying special attention to OQSs with three or four 24 terms. Once this problem is solved, the OPM obtained can be used in different decision-making and classification problems in which agents show their opinions about a set of alternatives through an OQS equipped with a metrizable OPM: Measuring consensus in a group of agents on a subset of alternatives, and consensus-based clustering procedures, as in Garc´ıa-Lapresta and P´erez-Rom´an [11]; consensus-reaching processes, as in Garc´ıa-Lapresta and P´erez-Rom´an [13]; implementing an appropriate voting system, such as the one introduced and analyzed in Garc´ıa-Lapresta and P´erez-Rom´an [14]; etc. As shown in Subsection 2.3, for g= 3, answering a single question is sufficient to assign the corresponding metrizable OPM; for g= 4, the number of questions should be between 2 and 4 (see Figure 4). Following this pattern, it is possible to determine all the metrizable OPMs of OQSs with granularity g∈ {5,6,7}(OQSs with more than seven linguistic terms are neither usual nor appropriate). This tedious task needs to be carried out if the proposal included in Subsections 4.2 and 4.3 is applied. Taking into account Proposition 4, totally uniform OPMs are metrizable only through πrs πtu ⇔ρ· |s−r|< ρ · |u−t|, with ρ > 0. Thus, it is appropriate to associate a real number with each term of the OQS (for instance, lrcan be identified with r). However, this identification is not possible for nonuniform OQSs (equipped with the corresponding OPMs), even when they are metrizable, since different non proportional linear metrics may generate the same OPM (see Remark 3). On the meaningless of assigning numerical values to the linguistic terms of non-uniform OQSs, see Roberts [26], Franceschini et al. [9] and Fattore et al. [8], among others. As further research, the present analyses could be extended to the framework of intervals of linguistic terms, when agents are allowed to assign several consecutive terms of the OQS, if they hesitate (see Garc´ıa-Lapresta and P´erez-Rom´an [12] and Garc´ıa-Lapresta and Gonz´alez del Pozo [10]). Acknowledgments. The authors are grateful to Gabriella Pigozzi for her comments and suggestions and her kind hospitality during the stay of the first author at LAMSADE in June 2015, and also to four anonymous referees for their comments and suggestions. The authors are also grateful to Direcci´on General de Innovaci´on y Equidad Educativa of Junta de Castilla y Le´on for allowing us to conduct the surveys on school principals who participated in TIMSS 2015. Financial support from Spanish Ministerio de Econom´ıa y Competitividad (project ECO2016- 25