Efficiency assessment using fuzzy production possibility set and enhanced Russell Graph measure

This paper studies the efficiency assessment of Decision Making Units (DMUs) when their inputs and outputs are fuzzy sets. An axiomatic derivation of the fuzzy production possibility set is presented and a fuzzy enhanced Russell graph measure is formulated using a fuzzy arithmetic approach. The proposed approach uses polygonal fuzzy sets and LU-fuzzy partial orders, and provides crisp efficiency measures (and associated efficiency ranking) as well as fuzzy efficient targets. The proposed approach has been compared with other fuzzy DEA approaches on different datasets from the literature, and the results show that it has more discriminant power and more flexibility in modelling the input and output data.


Introduction
need to make an efficient use of the limited resources available. And not only the benefits involve reduced costs and increased revenue but also reduced natural resources depletion, reduced pollution, and increased sustainability. Hence, the need to develop efficiency assessment models and tools that can be used in practical settings where uncertainty may be present in the data.
Data envelopment analysis (DEA) is the methodology most commonly used for assessing the relative efficiency of a set of Decision Making Units (DMUs). The first step in DEA is deriving a Production Possibility Set (PPS) from the input and output data of the observed DMUs and using certain axioms (such as envelopment, free disposability, and convexity). The non-dominated subset of the PPS corresponds to the efficient frontier. If a DMU lies on the efficient frontier it is said to be efficient. Otherwise, it is inefficient and the distance to the efficient frontier, which is a function of the potential input and output improvements (i.e., the input and output slacks), is used to compute an efficiency score. There are different ways to define the PPS (e.g., Constant Returns to Scale, CRS, Variable Returns to Scale, VRS) as well as different ways to compute the efficiency score. In this paper, we will use the enhanced Russell graph measure (ERM) proposed in Pastor et al. (1999), which, in the case of crisp data, is equivalent to the slacks-based measure of efficiency (SBM) of Tone (2001).
Although still a very active research field, the DEA methodology is very well developed in the case of crisp data. There are however situations in which there is uncertainty in the data. Dyson and Shale (2010) discusses possible sources of uncertainty in DEA and reviews different approaches for handling this uncertainty. A classic approach for handling data uncertainty is stochastic DEA methods, such as Chance Constrained DEA (Cooper et al. 2002(Cooper et al. , 2004 or Monte Carlo simulation (e.g., Kao andLiu 2009, 2019). These approaches assume that the input and output data are random variables whose joint distribution function is known (generally a multivariate normal) or can be fitted from historical data (e.g., a beta distribution). The resulting efficient frontier is random and so are the efficiency scores. Olesen and Petersen (2016) present a thorough review of stochastic DEA approaches, including Stochastic Frontier Analysis and Chance-constrained DEA. The reader is also referred to Salahi et al. (2019) and Izadikhah (2021) for examples of Robust optimization and Chanceconstrained DEA approaches, respectively, that also use the ERM efficiency considered in this paper.
An alternative to considering random data is modelling data uncertainty using fuzzy sets. Thus, often, the data variability is not due to randomness. There can be multiple factors such as inherent data variability, effectiveness of measurement instruments, multiplicity of data sources, etc. data can lead to uncertainty in the input and output data. Take, for example, a typical input variable such as the number of employees of a certain company or business unit. That variable is not random, but can fluctuate from 1 month to another (or even from 1 week to another) along the year and it may not be easy to reach a single, crisp number to represent that variable. The same happens with typical output variables like the number of customers attended or the number of transactions processed in cases in which a strict and exhaustive counting of all inquiries and requests attended may not exist and the corresponding numbers are only known approximately. Fuzzy sets have proved very helpful to model these type of uncertainties. Hence, applying DEA to fuzzy data has been the subject of many studies [see Hatami-Marbini et al. (2011) and Emrouznejad et al. (2014) for a survey and a taxonomy], although many of the existing approaches have drawbacks (Soleimani-damaneh et al. 2006). A number of Fuzzy DEA (FDEA) approaches are based on the α-level set (i.e., α-cut) concept [e.g., Kao and Liu (2000), Saati et al. (2002)]. Another large group of FDEA studies use a fuzzy ranking approach [e.g., Guo and Tanaka (2001), León et al. (2003)], Ghasemi et al. (2015)]. Other researchers use a possibility [e.g., Lertworasirikul et al. (2003), or a fuzzy arithmetic approach [e.g., Wang et al. (2005)] or consider fuzzy random/type-2 fuzzy sets [e.g., Tavana et al. (2013)]. Another category of FDEA approaches, and which has been thoroughly reviewed in Peykani et al. (2021), is Fuzzy Chance-constrained DEA. Thus, it can be said that FDEA is a fruitful and growing area of efficiency analysis under uncertainty, with many new approaches and applications being reported [e.g., Kachouei et al. (2020), Ebrahimnejad and Amani (2021)]. And not only can FDEA be applied in efficiency analysis, but also it can be used to solve multiobjective fuzzy optimization problems (e.g., Bagheri et al. 2020Bagheri et al. , 2021a. Most existing FDEA approaches use a radial metric and an input or an output orientation. There are many situations, however, in which both inputs and outputs should be improved and not necessarily in the same proportions. In those cases, a non-radial and nonoriented approach, like the proposed fuzzy ERM (FERM) model, is adequate. In this paper, an axiomatic approach to derive a Fuzzy PPS (FPPS) is presented and an FERM model is proposed. Only a few researchers have attempted to explicitly build a FPPS from the observations. Thus, Allahviranloo et al. (2007) use the Extension principle to define an FPPS, while Raei Nojehdehi et al. (2012) use a geometrical approach based on t-norms. Table 1 summarizes the existing fuzzy ERM/SBM approaches together with other fuzzy DEA methods for comparison. As regards FERM approaches, note that a fuzzy ranking approach is used in Jahanshahloo et al. (2004), Ahmady et al. (2015) and Izadikhah et al. (2017). Hsiao et al. (2011) and Puri and Yadav (2013) use the α-level set approach of Kao and Liu (2000), while Saati and Memariani (2009) use the α-level set approach of Saati et al. (2002). Chen et al. (2013) also use α-cuts and the Extension Principle. Both Hsiao et al. (2011) and Chen et al. (2013) also formulate a Fuzzy Super SBM model. Wu et al. (2015) proposed the α-level set FERM approach with undesirable outputs. Izadikhah and Khoshroo (2018) also considered undesirable outputs and proposed a possibility approach to solve a super-efficiency FERM model. Finally, Azadi et al. (2015) proposed a possibility approach based on a multiplier ERM formulation.
As it can be seen in Table 1, the FERM approach proposed in this paper uses polygonal fuzzy numbers, an LU-fuzzy partial order, and an axiomatic derivation of the fuzzy PPS. With respect to polygonal fuzzy numbers, Stefanini et al. (2006) proposed using a uniform subdivision of the interval [0, 1] to get a finite number of α-cuts. Báez-Sánchez et al. (2012) study the polygonal fuzzy numbers as a particular case of parametric representation of fuzzy numbers with linear interpolation. As an application of polygonal fuzzy numbers, Chen and Adam (2018) has recently proposed a new transformation-based weighted fuzzy interpolative reasoning method. In this paper, we consider polygonal fuzzy number to model or approximate inputs and outputs. No ranking functions or expected values on the polygonal fuzzy numbers are needed to formulate the corresponding crisp model.
Another difference between the existing and the proposed FERM approach is that we compute a crisp efficiency score instead of a fuzzy efficiency score. Fuzzy efficiency scores are more consistent with the fuzzy nature of the data, but the crisp efficiency scores are simpler to understand and apply for practitioners. In this paper, we have opted for crisp efficiency scores. Computing fuzzy efficiency scores using a fully fuzzy approach and LUfuzzy partial orders is also possible (e.g., Arana-Jiménez et al. 2020), but the process is necessarily much more involved. In this paper, we have searched for a compromise between fuzziness/information loss and simplicity.
Therefore, the contributions of the proposed approach are several. One of them is the use of polygonal fuzzy numbers and a LU-fuzzy partial order. Another is the axiomatic derivation of the fuzzy production possibility set that contains all the fuzzy operating points that are deemed feasible. Using this fuzzy DEA technology, a simple fuzzy optimization To solve the proposed FERM model, a crisp optimization model is formulated that, although in principle is non-linear, can be appropriately linearized. In the end, we have a simple and effective FDEA approach for assessing the efficiency and projecting the production units. The paper is organized as follows. In Sect. 2, the crisp DEA PPS and ERM model are reviewed. To consider and model the uncertainty in the data, in Sect. 3, polygonal fuzzy numbers are presented. Suitable notations and rules to deal with arithmetic operations and LU-fuzzy partial orders are also derived. Then, in Sect. 4, assuming that inputs and outputs are given by polygonal fuzzy numbers, the corresponding fuzzy PPS is derived, a fuzzy ERM model is formulated and a method to solve it is presented. Section 5 illustrates the proposed approach and compares it with existing approaches using different datasets from the literature. Finally, in Sect. 6, conclusions are drawn and further research outlined.

Crisp production possibility set and ERM model
Let us consider a set of n DMUs. For j ∈ J = {1, . . . , n}, each DMU j has m inputs X j = (x 1 j , . . . , x m j ) ∈ R m , produces s outputs Y j = (y 1 j , . . . , y s j ) ∈ R s . In the classic Charnes et al. (1978) DEA model, the production possibility set (PPS) or technology, denoted by T , satisfies the following axioms: Following the minimum extrapolation principle (see Banker et al. 1984), the DEA PPS, which contains all the feasible input-output bundles, is the intersection of all the sets that satisfy axioms (A1)-(A4), and can be expressed as In radial DEA models, see Charnes et al. (1978), the efficiency of a DMU p is measured in two ways. One way is reducing all the inputs equi-proportionally without decreasing the outputs (input-oriented model), and then, the problem is formulated as min{θ p ∈ R + |(θ p x p , y p ) ∈ T }. The second one consists in expanding all the outputs equi-proportionally without increasing the inputs (output-oriented model), i.e., max{θ p ∈ R + |(x p , θ p y P ) ∈ T }.
There exist other DEA approaches in which the reductions of inputs and outputs are not equi-proportional. In particular, Pastor et al. (1999) proposed the following Enhanced Russell Graph Efficiency Measure (ERM): λ j y r j ≥ γ r y r p , r = 1, . . . , s, where λ j , j = 1, . . . , n, are the intensity variable of each DMU j for defining the corresponding efficient target of DMU p . R is interpreted as the the ratio between the average input reduction and the average output increase. Pastor et al. (1999) showed some interesting properties of ERM. Thus, 0 < R ≤ 1, R is units invariant and a DMU j is efficient if and only if R = 1.

Polygonal fuzzy numbers
In this paper, we model the uncertainty in the data, and hence on the production possibility set, by means of polygonal fuzzy numbers. In this regard, in this section, we present the definition of polygonal fuzzy numbers, corresponding arithmetic operations, and inequality relationships using LU-partial orders on fuzzy sets. Furthermore, we derive certain properties that facilitate the modelling of the corresponding PPS which we propose in Sect. 4. To this end, it is necessary to establish the following notation and results.
We denote by K C = a, a | a, a ∈ R and a ≤ a the family of all bounded closed intervals in R. A fuzzy set on R n is a mapping u : R n → [0, 1]. For each fuzzy set u, we denote its α-level set as [u] α = {x ∈ R n | u(x) ≥ α} for any α ∈ (0, 1], and its support as supp(u) = {x ∈ R n | u(x) > 0}. The closure of supp(u) defines the 0-level of u, i.e., [u] 0 = cl(supp(u)) where cl(M) means the closure of the subset M ⊂ R n . Following Prade (1978, 1980), a fuzzy set u on R is said to be a fuzzy number if (i) u is normal, this is there exists x 0 ∈ R, such that u(x 0 ) = 1, (ii) upper semi-continuous function, (iii) convex, and (iv) [u] 0 is compact. F C denotes the family of all fuzzy numbers. The α-levels of a fuzzy number can be represented as Although there are many parametrical families of fuzzy numbers [see Báez-Sánchez et al. (2012) and Hanss (2005) for a complete description of these families], two of the most used families of fuzzy numbers are triangular and trapezoidal fuzzy numbers, because of their easy modeling and interpretation [see, for instance, Prade (1978, 1980), Kaufmann and Gupta (1985), Khan et al. Definition 1 Given a partition of the interval [0, 1], P k = {α i : i = 0, 1 . . . , k}, with 0 = α 0 < α 1 <, . . . , < α k = 1, a fuzzy numberã is said to be a k-polygonal fuzzy In terms of their membership function instead of their α-levels, a polygonal fuzzy number can alternatively be defined as follows.

Proposition 1 Given a partition of the interval
(2) A proof of Proposition 1 can be found in Appendix A. As a result of Proposition 1, we can denote a k-polygonal fuzzy number with respect to P k = {α i : . For the sake of simplicity, in the sequel, we will assume that α i = i k , and thenã is said to be a regular k-polygonal fuzzy number. We denote the set of all regular k-polygonal fuzzy numbers as R P F N k . Thus,ã = (a − 0 , a − 1 , a + 1 , a + 0 ) corresponds to a trapezoidal fuzzy number. And, if a − 1 = a + 1 , thenã is a triangular fuzzy number, and it can be noted asã = (a − , a, a + ). Finally, we denote0 and1 as the regular polygonal fuzzy numbers whose components are all 0 and 1, respectively. Figure 1 illustrates examples of both, a general k-polygonal fuzzy number, on the top, and regular 1-polygonal and 2-polygonal fuzzy numbers, on the bottom.
Let us give the following definitions for two arithmetic operations, namely sum and multiplication by scalar, necessary for the derivation of the fuzzy PPS.
, the following basic arithmetical operations can be defined: Let us recall the LU-fuzzy partial orders, which are well known in the literature [see, e.g., Wu (2009), Stefanini and Arana-Jiménez (2019) and the references therein].
and some examples of k = 1, k = 2 and k = 3 regular polygonal fuzzy numbers (bottom) The relationships μ ν and μ ν mean ν μ and ν ≺ μ, respectively. In Arana-Jiménez (2018), a reformulation of the previous definition for triangular fuzzy numbers by means of the relationship between their parameters is presented. The following extends this result to regular k-polygonal fuzzy numbers.

Proposition 2 Given two regular k-polygonal fuzzy numbersã
Proof The proof is similar to that given in Arana-Jiménez (2018).

Proposed fuzzy PPS and fuzzy ERM model
Let us consider a set of n DMUs, j ∈ J = {1, . . . , n}, each DMU j has m inputs Let us consider the following axioms, which are analogous to (A1)-(A4) in Section 2 but considering fuzzy inputs and outputs and using the corresponding partial order introduced in Definition 3: Following the minimum extrapolation principle, the fuzzy production possibility set is the intersection of all sets that satisfy axioms (B1)-(B4): Theorem 1 Under axioms (B1), (B2), (B3), and (B4), T FDEA is the fuzzy production possibility set that results from the minimum extrapolation principle.
Proof See Appendix B.
Given the above fuzzy production possibility set, we can present the following definition of efficiency.
After the characterization result for the fuzzy PPS, given in Theorem 1, and to provide a measure for the efficiency of each DMU, we can formulate the following fuzzy ERM (FERM) model, as an extension of the corresponding crisp ERM model: where the different inputsx i j and outputsỹ r j belong to (RPFN k ) + , that is Note that one of the advantages of using k-polygonal fuzzy numbers is that you can have all the flexibility you need for modeling the fuzzy input and output data and yet be able to describe the whole membership function using a finite set of alpha values. Since the linear combination of the observed inputs and outputs is also k-polygonal fuzzy numbers, the above feature also applies to the constraints of the proposed model (5). That is why, according to Proposition 2, the following crisp model, which only considers a finite number of alpha values, is equivalent to model which considers all alpha values.
The above model above can be written in parameterized form as follows: Note that FERM inherits some ERM properties such as that 0 < R F ≤ 1, and that is not a sufficient condition for efficiency. Hence, model (6) needs to be modified to correct this, as argued below.
In fact, all convex combinations of the two DMUs have R F (X p ,Ỹ p ) = 1, for p = 1, 2. Therefore, an optimal value of unity for model (6) does not imply the efficiency of the DMU. Moreover, among all the alternative optimal targets for DMU 2 ,X target 2 = λ 1X1 + λ 2X2 , only one is truly efficient, namelyX target 2 =X 1 , which corresponds to λ 1 = 1 and λ 2 = 0. Hence, model (6) is not an adequate extension of the crisp case.
An initial attempt to solve this discrepancy between a ratio value R F (X p ,Ỹ p ) = 1 and the DMU's efficiency characterization, as discussed in Example 1, might be to consider a different variable θ il and γ rl , for each level l = 0, . . . , k, that is  Fig. 2 Consider two DMUs that consume a single input and produce a single and unit output. The inputs of DMU 1 and DMU 2 areX 1 = (1, 1.75, 2.5, 3.5, 4.75, 6), andX 2 = (2, 3, 3, 3.5, 4.75, 6), respectively. Left: DMU 2 is clearly inefficient,X 1 X 2 ,X 1 =X 2 and they have the same output, but when (6) is solved, we get that θ = 1 and hence R F (X 2 ,Ỹ 2 ) = 1. This is in contrast with the crisp case, where only efficient DMUs have an ERM efficiency score equal to unity. Center: If the variables θ l and γ l are let free for each l-level, the issue remains. Since the right limits of bothX 1 andX 2 coincide, the solution of model (7) would be θ 0 = θ 0.5 = θ 1 = 1 and R F m (X 2 ,Ỹ 2 ) = 1. Right: Using separate left and right variables θ L l and θ R l , the modified (MFERM) model (8) However, coming back to Example 1, the solution of model (7) for DMU 2 input would be θ 0 = θ 0.5 = θ 1 = 1, i.e., since the right limits for bothX 1 andX 2 coincide, this leads to all three θ l (and γ l ) variables being equal to unity, and therefore to a modified (FERM) value R F m (X 2 ,Ỹ 2 ) = 1, which would be misleading as DMU 2 is not efficient. See central plot of Fig. 2.
We have checked that in a fuzzy framework, both efficient and inefficient DMUs could have the same objective value of 1 for the straightforward extension of ERM model (6), or the modified (7). For example, if the upper/lower limits of the l-levels coincide, may lead to this case. This motivates the inclusion of left and right variables {θ L l ; θ R l ; γ L l ; γ R l }, for each l-level The advantage of model (8) over (5) is that, as shown below in Theorem 2, it has the efficiency indication property, i.e., R F M (X p ,Ỹ p )=1 if and only if the DMU p is efficient. Note that, using separate left and right input reduction factors for each input and left and right output expansion factors for each output, all feasible input reductions and output increases are exhausted. Back to Example 1, not all the above variables θ L l and θ R l are equal to one for the inefficient DMU 2 , and R F M (X 2 ,Ỹ 2 ) < 1. This can be seen in the right panel of Fig. 2, which shows the optimal left and right values of the input reduction variables for the different alpha values.
Despite the objective function of (8) is non-linear, it is possible, following a similar procedure to that proposed by Pastor et al. (1999), to reformulate the problem and get an equivalent problem with a linear objective function as follows: Reformulating (8) using (9), we have The above optimization problem (LMFERM) is now a linear program whose feasibility region and objective function are equivalent to those given in (8), with the change of variables given in (9).
Note that above (MFERM) model, and equivalently model (LMFERM), computes the efficiency score of a DMU p and this is indicated in the arguments of R F (X p ,Ỹ p ). It can, however, be used to project any feasible operating point (x,ỹ) to compute its efficiency R F M (x,ỹ). This will be useful for proving the following result.

Proof See Appendix D
In summary, extending the conventional ERM efficiency, the proposed MFERM approach uses a linearizable, non-oriented, non-radial optimization model that exhausts all feasible input reductions and output expansions at all l-levels, providing crisp efficiency scores (and corresponding efficiency ranking) as well as fuzzy efficient targets.

Triangular fuzzy numbers dataset
To illustrate the proposed fuzzy ERM model, consider the dataset from Arya and Yadav (2017) shown in Table 2 and which correspond to 12 Community Health Centers (CHCs) in Meerut district of Uttar Pradesh, India. The table shows the fuzzy input and output of each DMU. The first input is the total sum of doctors and staff nurses, while the second one is the number of pharmacists. The two outputs correspond to the number of inpatients and outpatients, respectively. All the inputs and outputs are given as triangular fuzzy numbers, which correspond to k = 1 regular polygonal fuzzy numbers.
The efficiency R F M (X p ,Ỹ p ) of each DMU p , p = 1, . . . , 12, has been computed using (10), and is shown in the second column of Tables 2 and 3. It can be seen that DMUs H1, H2, H3, H8, and H12 are labelled efficient. In Table 3, we also add a ranking, based on this efficiency measurement, so that this ranking can be compared with those of other methods. Efficient DMUs are all ranked equal. Table 2 shows the fuzzy targets computed for each DMU using (11) and (12). Due to horizontal space constraints, the structure of the table is unusual in the sense that for each DMU, there are two rows, each showing the data and the results for each of the two inputs and two outputs. Note that the target coincides with the observed data in the case of the efficient DMUs and dominate it, in the sense of the partial order introduced in Definition 3, in the case of inefficient DMUs.
For comparison purposes, Table 3 also shows the efficiency scores computed by other approaches from the literature. The fourth column shows the S B M E p efficiency score computed by the fuzzy SBM model of Arya and Yadav (2017). Note that their method labelled all DMUs as efficient except one, namely H11. Thus, it seems that the proposed FERM approach has, at least for this dataset, more discriminant power than this method. The fifth and sixth columns of Table 3 correspond to the results of the Fuzzy SBM model of Chen et al. (2013) and Hsiao et al. (2011). Since they use fuzzy slack variables, their efficiency scores are fuzzy. The corresponding α-cuts for α = 0, 1 levels are shown. Note that the FERM efficiency scores are within the corresponding α = 0 cut for all DMUs.
Finally, the two last columns of Table 3 correspond to the fuzzy ERM approach of Izadikhah et al. (2017), where a pair of models are used to generate upper and lower limits of interval efficiency score, based on the enhanced Russell model applied to interval data. The whole dataset is converted into intervals for applying this methodology, computing the nearest weighted interval approximation of fuzzy numbers, with the weighting function (4a 3 , 4a 3 ). We use a preference degree measurement ρ j , see Izadikhah et al. (2017) for more details, to establish some partial order between intervals and ranking the DMUs. In this case, although at first sight, it appears that this method has more discriminant power from Izadikhah et al. (2017), we also remark the different efficiency interpretations. Izadikhah et al. (2017) approach computes an efficiency score interval (converting or approximating all data to intervals). Whereas the proposed method is based on technical efficiency within a fuzzy technology framework. According to the latter, the efficient DMUs are characterized by R F M (X p ,Ỹ p ) = 1 (see Theorem 2). In this regard, we found in some cases that efficient DMUs are ranked after inefficient ones by Izadikhah et al. (2017) approach. However, despite these differences, we see agreement between the results of both methods. Thus, except for DMU 12, R F M (X p ,Ỹ p ) always lies within the corresponding efficiency score interval. The level of agreement between the results of our approach and Izadikhah et al. (2017) Arya and Yadav (2017)], together with FERM targets as well as the parameters  (1960, 1965, 1972) (1960, 1965, 1972)

has been
Interval efficiency andỹ target r 4 , are represented with black squares and dashed lines. To compute the MFERM input and output targets, the observed datax i4 andỹ r 4 are multiplied by the corresponding variables {θ L il ; θ R il ; γ L rl ; γ R rl }, at levels l = 0, 1, 2, thus exhausting all possible input and output improvements. In the case of output 1, the support of the observed data and the MFERM target differ significantly, and hence, the x-axis has been broken (the gap is marked with two vertical lines) tested using a Wilcoxon signed-rank test, which is a non-parametric paired difference test for matched pairs or dependent samples. With a p-value= 0.2146, we cannot reject the null hypothesis of homogeneity, i.e., that the median difference is zero.

Modified dataset with 2-polygonal fuzzy numbers
In this section, we illustrate the proposed approach with fuzzy data that are not triangular fuzzy numbers. Actually, polygonal fuzzy numbers allow a great deal of flexibility in modelling the uncertainty in the input and output data. To keep things simple, however, we will just consider 2-polygonal fuzzy numbers, instead of triangular (1-polygonal). We have modified the inputs and outputs of the DMUs for getting RPFN 2 data. The modifications have been randomly generated keeping the same closure; see Table 4.
As an example, Fig. 3 shows the input/output data for DMU H4, as well as the corresponding input and output targets. The black circles and solid lines show the observed 2-polygonal fuzzy inputs and outputs. The MFERM targets are represented with dashed lines and black squares. For example, for input 2, the corresponding variables are {θ L 2,0 ; θ L 2,1 ; θ L 2,2 ; θ R 2,2 ; θ R 2,1 ; θ R 2,0 } = {0.525; 0.849; 1.000; 1.000; 0.683; 0.445}. In the case of output 1, the target support is significantly to the right of the observed data, implying a large inefficiency as regards that output. For the other inefficient DMUs, this separation between the targets and observed data values is even greater and happens mainly for output 2.
The lower part of Table 4 also shows the computed MFERM targets. The efficient DMUs are projected onto themselves. For the inefficient DMUs, as graphically shown in Fig. 2 for DMU H4, the fuzzy targets always dominate the observed inputs and outputs and sometimes by a large amount.
The interpretation of the fuzzy targets is similar to that of the fuzzy data from which they derive. Thus, for each α-level, the corresponding α-cuts of the observed input and the corresponding target can be compared, and the same occurs with the observed output and the target output. The proposed approach guarantees that, for each α-level, the lower limit of the target input is less than the lower limit of the observed input and that the upper limit of the target input is less than the upper limit of the observed input. Moreover, the relative difference between the two lower limits and between the upper limits must be at least the value of the corresponding θ L il , θ R il , γ L rl or γ R rl variable computed by the model (LMFERM) (10), and Eq. (9), which represents the maximum margin for improvement for that input or output.
By contrast, for output 2, the number of outpatients, {γ L 2,0 ; γ L 2,0.5 ; γ L 2,1 ; γ R 2,1 ; γ R 2,0.5 ; γ R 2,0 } = {1.0005; 1.0005; 1.0002; 1.0002; 1.0001; 1.000} because the upper limit of one of the 0.0set cannot be increased, as it can be seen in Fig. 3. This leads to the corresponding MFERM In summary, what the example in Sect. 5.2 illustrates is that the use of k-polygonal fuzzy numbers provides all the flexibility that be needed to represent the uncertainty in the data and that the proposed FERM approach can handle all those situations providing appropriate efficiency scores and the corresponding fuzzy targets.

Case study: electric power distribution company
This numerical example is taken from Izadikhah et al. (2017), where 17 Iranian suppliers of self-supporting cable for Markazi Province Electric Power Distribution Company (MPEPDC) in Iran are evaluated. There are 2 inputs, x 1 = the overall suppliers ranking (ordinal variable) and x 2 = unit price by considering volume discount (interval type). Furthermore, there are 6 outputs: y 1 = production capacity and y 2 = financial potential, both of interval type, and y 3 = environmental standards and regulations, y 4 = research and developments for ecodesign product, y 5 = safety and health standards, and y 6 = customer satisfaction. These last four outputs are fuzzy triangular numbers (see cf. Tables 10-14 in Izadikak for more details). Note that, since interval data [a 1 , a 2 ] correspond with trapezoidal fuzzy numbers a 1 , a 1 , a 2 , a 2 ), we model such type of data as (R P F N 1 ) + . Recall also that the triangular fuzzy outputs are particular cases of the trapezoidal ones, where x − 1 = x + 1 .
We adopt the same approximation of the ordinal input x 1 as an interval as done in Izadikhah et al. (2017). Table 5 shows the results of applying the proposed (PLFERM) model to this dataset, and the comparison with the interval efficiency scores obtained by Izadikhah et al. (2017). The proposed approach identifies nine DMUs as efficient DMUs. This relatively high number is not surprising for a fuzzy problem with a such number of DMUs and variables. As already discussed above in Sect. 5.2, obviating the differences between efficiency interpretations in both methods, we find sufficient agreement, since most R F M (X p ,Ỹ p ) fall within the corresponding efficiency intervals. Moreover, if we apply a Wilcoxon signed-rank test of the results of our approach and Izadikhah et al. (2017), we get a p-value= 0.285. This means that we cannot reject the null hypothesis of homogeneity, i.e., that the median difference is equal to zero, which supports the existence of sufficient agreement between both efficiency measurements. As in Sect. 5.1, Izadikhah et al. (2017) provides a ranking even between the efficient DMUs and, again, we find that in some cases, efficient DMUs are ranked after inefficient ones. The proposed approach is not able to discriminate between the efficient DMUs, but can easily rank the inefficient DMUs just sorting by their R F M (X p ,Ỹ p ) value.

Conclusions
This paper presents a new approach for efficiency assessment and target setting when the input and output data are fuzzy. It is based on polygonal fuzzy numbers and LU-fuzzy partial orders. First, from the observed fuzzy data, and using simple axioms analogous to the ones considered in the crisp case, the fuzzy PPS containing all feasible operating points is inferred. Based on this PPS, a fuzzy ERM DEA model is proposed to compute, for each DMU, a crisp efficiency score, and a fuzzy target. The use of polygonal fuzzy numbers provides ample flexibility for modeling the uncertainty in the data. In addition, the non-radial approach used, which exhausts all possible input and output slacks, provides increased discriminant power. The proposed approach has two main limitations. One is that it computes crisp efficiency scores. The second one is that it does not work (i.e., it leads to unbounded solutions) when the left limit of any input or output of a DMU is zero.
As regards topics for continuing this research, one is formulating a super-efficiency approach, so that the efficient DMUs can be classified into extreme efficient and non-extreme efficient and the former can be ranked. Another topic is fully fuzzy approach for computing fuzzy efficiency scores. Other interesting extensions of the proposed approach include handling negative data and undesirable outputs given as fuzzy sets. Finally, a meaningful comparison between fuzzy DEA and stochastic DEA is also due (see, e.g., Wanke et al. (2018)).
article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Appendix B: Proof of Theorem 1
Proof Denote by T true the result of the minimum extrapolation principle axioms (B1), (B2), (B3), and (B4). To prove the theorem, it is necessary to show that T FDEA = T true . To this end, let us divide the proof into two parts.
(ii) T FDEA ⊆ T true . We need to prove that every element of T FDEA belongs to T true . To this purpose, consider (x,ỹ) ∈ T FDEA , which means that there exists λ 0, λ ∈ R n , such that We have that (X j ,Ỹ j ) ∈ T true by (B1), for all j ∈ J . Then, by (B4), it follows that (λ jX j , λ jỸ j ) ∈ T true , for all j ∈ J . Reasoning by induction, let us prove that • Check that in the case s = 1 it holds. This case is immediate, since (X 1 ,Ỹ 1 ) ∈ T true , and then, by (B3), (λ 1X1 , λ 1Ỹ1 ) ∈ T true . • Check that if cases s ≤ t are true, this implies that the case s = t + 1 is also true.