+ All Categories
Home > Documents > A concept of fuzzy input mix-efficiency in fuzzy DEA and its application in banking sector

A concept of fuzzy input mix-efficiency in fuzzy DEA and its application in banking sector

Date post: 27-Dec-2016
Category:
Upload: shiv-prasad
View: 214 times
Download: 0 times
Share this document with a friend
14
A concept of fuzzy input mix-efficiency in fuzzy DEA and its application in banking sector Jolly Puri , Shiv Prasad Yadav Department of Mathematics, I.I.T. Roorkee, Roorkee 247667, India article info Keywords: Fuzzy data envelopment analysis Fuzzy CCR input efficiency Fuzzy SBM input efficiency Fuzzy input mix-efficiency Fuzzy correlation coefficient Fuzzy ranking approach Banking performance evaluation abstract Data envelopment analysis (DEA) is a linear programming based non-parametric technique for evaluating the relative efficiency of homogeneous decision making units (DMUs) on the basis of multiple inputs and multiple outputs. There exist radial and non-radial models in DEA. Radial models only deal with propor- tional changes of inputs/outputs and neglect the input/output slacks. On the other hand, non-radial mod- els directly deal with the input/output slacks. The slack-based measure (SBM) model is a non-radial model in which the SBM efficiency can be decomposed into radial, scale and mix-efficiency. The mix- efficiency is a measure to estimate how well the set of inputs are used (or outputs are produced) together. The conventional mix-efficiency measure requires crisp data which may not always be available in real world applications. In real world problems, data may be imprecise or fuzzy. In this paper, we propose (i) a concept of fuzzy input mix-efficiency and evaluate the fuzzy input mix-efficiency using a – cut approach, (ii) a fuzzy correlation coefficient method using expected value approach which calculates the expected intervals and expected values of fuzzy correlation coefficients between fuzzy inputs and fuzzy outputs, and (iii) a new method for ranking the DMUs on the basis of fuzzy input mix-efficiency. The proposed approaches are then applied to the State Bank of Patiala in the Punjab state of India with districts as the DMUs. Ó 2012 Elsevier Ltd. All rights reserved. 1. Introduction Data envelopment analysis (DEA), proposed by Charnes, Cooper, and Rhodes (1978), is a linear programming based non-parametric method for evaluating the relative efficiency of homogeneous deci- sion making units (DMUs) on the basis of multiple inputs and mul- tiple outputs. The popularity of DEA is due to its ability to measure relative efficiencies of DMUs without prior weights on the inputs and outputs. There are two types of models in DEA: radial and non-radial. Radial model is represented by the CCR model (Charnes et al., 1978), the first DEA model. Basically, it deals with propor- tional changes of inputs or outputs. The CCR efficiency score re- flects the proportional maximum input (output) reduction (expansion) rate which is common to all inputs (outputs). How- ever, in real world businesses, not all inputs (outputs) behave in the proportional way. Also a radial model neglects slacks in in- puts/outputs while reporting the efficiency score. In many cases, we find a lot of remaining non-radial slacks. So, if these slacks have an important role in evaluating the efficiency, the radial ap- proaches may mislead the decision when we utilize the efficiency score as the only index for evaluating performance of DMUs. In contrast, the non-radial model is represented by the slack-based measure (SBM) (Tone, 2001), which put aside the assumption of proportionate changes in inputs and outputs, and deal with slacks directly. Also the SBM model assesses the efficiency of the input or output mix as well as it assesses the overall level of efficiency. It has three variations (i) input-oriented, (ii) output-oriented, and (iii) non-oriented. For details of comparison between radial and non-radial measure, see Avkiran, Tone, and Tsutsui (2008) along with the shortcomings for both the CCR and the SBM models. Tone (1998) suggests that the results from both the CCR and the SBM models can be used to evaluate the mix-efficiency. The mix-effi- ciency is a measure to estimate how well the set of inputs are used (or outputs are produced) together (Herrero, Pascoe, & Mardle, 2006; Asbullah, 2010). Tone (1998) presented input mix-efficiency and output mix-efficiency by using input-oriented and output-ori- ented variations of both CCR and SBM models. Conventional mix-efficiency measure requires crisp input and output data, which may not always be available in real world applications. Actually, in real world problems, inputs and outputs are often imprecise or fuzzy. So, in order to calculate mix-efficiency with imprecise or fuzzy data, we propose the concept of fuzzy in- put mix-efficiency (FIME). For measuring FIME, we propose the in- put-oriented fuzzy CCR model (FCCR I ) and input-oriented fuzzy 0957-4174/$ - see front matter Ó 2012 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2012.08.047 Corresponding author. E-mail addresses: [email protected] (J. Puri), [email protected] (S.P. Yadav). Expert Systems with Applications 40 (2013) 1437–1450 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa
Transcript
Page 1: A concept of fuzzy input mix-efficiency in fuzzy DEA and its application in banking sector

Expert Systems with Applications 40 (2013) 1437–1450

Contents lists available at SciVerse ScienceDirect

Expert Systems with Applications

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

A concept of fuzzy input mix-efficiency in fuzzy DEA and its applicationin banking sector

Jolly Puri ⇑, Shiv Prasad YadavDepartment of Mathematics, I.I.T. Roorkee, Roorkee 247667, India

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

Keywords:Fuzzy data envelopment analysisFuzzy CCR input efficiencyFuzzy SBM input efficiencyFuzzy input mix-efficiencyFuzzy correlation coefficientFuzzy ranking approachBanking performance evaluation

0957-4174/$ - see front matter � 2012 Elsevier Ltd. Ahttp://dx.doi.org/10.1016/j.eswa.2012.08.047

⇑ Corresponding author.E-mail addresses: [email protected] (J. P

(S.P. Yadav).

Data envelopment analysis (DEA) is a linear programming based non-parametric technique for evaluatingthe relative efficiency of homogeneous decision making units (DMUs) on the basis of multiple inputs andmultiple outputs. There exist radial and non-radial models in DEA. Radial models only deal with propor-tional changes of inputs/outputs and neglect the input/output slacks. On the other hand, non-radial mod-els directly deal with the input/output slacks. The slack-based measure (SBM) model is a non-radialmodel in which the SBM efficiency can be decomposed into radial, scale and mix-efficiency. The mix-efficiency is a measure to estimate how well the set of inputs are used (or outputs are produced) together.The conventional mix-efficiency measure requires crisp data which may not always be available in realworld applications. In real world problems, data may be imprecise or fuzzy. In this paper, we propose (i) aconcept of fuzzy input mix-efficiency and evaluate the fuzzy input mix-efficiency using a – cut approach,(ii) a fuzzy correlation coefficient method using expected value approach which calculates the expectedintervals and expected values of fuzzy correlation coefficients between fuzzy inputs and fuzzy outputs,and (iii) a new method for ranking the DMUs on the basis of fuzzy input mix-efficiency. The proposedapproaches are then applied to the State Bank of Patiala in the Punjab state of India with districts asthe DMUs.

� 2012 Elsevier Ltd. All rights reserved.

1. Introduction

Data envelopment analysis (DEA), proposed by Charnes, Cooper,and Rhodes (1978), is a linear programming based non-parametricmethod for evaluating the relative efficiency of homogeneous deci-sion making units (DMUs) on the basis of multiple inputs and mul-tiple outputs. The popularity of DEA is due to its ability to measurerelative efficiencies of DMUs without prior weights on the inputsand outputs. There are two types of models in DEA: radial andnon-radial. Radial model is represented by the CCR model (Charneset al., 1978), the first DEA model. Basically, it deals with propor-tional changes of inputs or outputs. The CCR efficiency score re-flects the proportional maximum input (output) reduction(expansion) rate which is common to all inputs (outputs). How-ever, in real world businesses, not all inputs (outputs) behave inthe proportional way. Also a radial model neglects slacks in in-puts/outputs while reporting the efficiency score. In many cases,we find a lot of remaining non-radial slacks. So, if these slacks havean important role in evaluating the efficiency, the radial ap-proaches may mislead the decision when we utilize the efficiency

ll rights reserved.

uri), [email protected]

score as the only index for evaluating performance of DMUs. Incontrast, the non-radial model is represented by the slack-basedmeasure (SBM) (Tone, 2001), which put aside the assumption ofproportionate changes in inputs and outputs, and deal with slacksdirectly. Also the SBM model assesses the efficiency of the input oroutput mix as well as it assesses the overall level of efficiency. Ithas three variations (i) input-oriented, (ii) output-oriented, and(iii) non-oriented. For details of comparison between radial andnon-radial measure, see Avkiran, Tone, and Tsutsui (2008) alongwith the shortcomings for both the CCR and the SBM models. Tone(1998) suggests that the results from both the CCR and the SBMmodels can be used to evaluate the mix-efficiency. The mix-effi-ciency is a measure to estimate how well the set of inputs are used(or outputs are produced) together (Herrero, Pascoe, & Mardle,2006; Asbullah, 2010). Tone (1998) presented input mix-efficiencyand output mix-efficiency by using input-oriented and output-ori-ented variations of both CCR and SBM models.

Conventional mix-efficiency measure requires crisp input andoutput data, which may not always be available in real worldapplications. Actually, in real world problems, inputs and outputsare often imprecise or fuzzy. So, in order to calculate mix-efficiencywith imprecise or fuzzy data, we propose the concept of fuzzy in-put mix-efficiency (FIME). For measuring FIME, we propose the in-put-oriented fuzzy CCR model (FCCRI) and input-oriented fuzzy

Page 2: A concept of fuzzy input mix-efficiency in fuzzy DEA and its application in banking sector

Table 1Input and output data with hk

I ; qkI and wk

I .

DMUS Input 1 Input 2 Output 1 Output 2 hkI

Rank qkI

Rank wkI

Rank

1 20 151 100 90 1.000000 1 1.000000 1 1.000000 12 19 131 150 50 1.000000 1 1.000000 1 1.000000 13 25 160 160 55 0.882708 8 0.852165 8 0.965399 84 27 168 180 72 1.000000 1 1.000000 1 1.000000 15 22 158 94 66 0.763499 12 0.755612 11 0.989670 56 55 255 230 90 0.834771 10 0.703764 12 0.843062 127 33 235 220 88 0.901961 7 0.894835 6 0.992100 48 31 206 152 80 0.796334 11 0.773958 10 0.971901 79 30 244 190 100 0.960392 4 0.904642 5 0.941950 9

10 50 268 250 100 0.870647 9 0.780509 9 0.896471 1111 53 306 260 147 0.955098 6 0.866137 7 0.906856 1012 38 284 250 120 0.958204 5 0.936020 4 0.976848 6

Source of input and output data: Cooper, Seiford and Tone, 2007.

Table 2The fuzzified data in terms of TFNs.

DMUS Input 1 (I1) Input 2 (I2) Output 1 (O1) Output 2 (O2)

1 (16,20,22) (150, 151,152) (95,100,102) (87,90,94)2 (18,19,20) (130, 131,132) (149,150,151) (46,50,52)3 (23,25,28) (158,160,162) (158,160,163) (53,55,56)4 (26,27,29) (165,168,169) (177,180,181) (70,72,75)5 (20,22,25) (155,158,162) (90,94,98) (63,66,68)6 (52,55,59) (250, 255,259) (222,230,235) (83,90,95)7 (30,33,34) (234,235,236) (210,220,225) (81,88,90)8 (27,31,33) (202, 206,208) (151,152,155) (75,80,84)9 (26,30,35) (240, 244,247) (188,190,193) (99,100,101)

10 (47,50,54) (262,268,271) (246,250,252) (94,100,108)11 (50,53,56) (300,306,309) (255,260,264) (143,147,152)12 (30,38,42) (283,284,285) (246,250,254) (116,120,123)

1438 J. Puri, S.P. Yadav / Expert Systems with Applications 40 (2013) 1437–1450

SBM model (FSBMI) with fuzzy input and fuzzy output data. Sev-eral approaches have been developed to deal with imprecise orfuzzy data in DEA. Sengupta (1992) applied principle of fuzzy settheory to introduce fuzziness in the objective function and theright-hand side vector of the conventional DEA model. Guo and Ta-naka (2001) used the ranking method and introduced a bi-levelprogramming model. Lertworasirikul (2001) developed a methodin which the inputs and outputs were firstly defuzzified and thenthe model was solved using a-cut approach. There are some otherapproaches based on a-cut which can be found in Meada, Entani,and Tanaka (1998), Kao and Liu (2000a) and Saati Mohtadi,Memariani, and Jahanshahloo (2002). Lertworasirikul, Fang, Jeffrey,Joines, and Nuttle (2003) proposed a possibility DEA model forfuzzy DEA (FDEA). Kao and Liu (2000a, 2000b, 2003, 2005) trans-formed fuzzy input and fuzzy output into intervals by using a-levelsets and built a family of crisp DEA models for the intervals. Liu(2008) and Liu and Chuang (2009) developed a fuzzy DEA/AR mod-el for the selection of flexible manufacturing systems and theassessment of university libraries respectively. Zhou, Lui, Ma, Liu,and Liu (2012) proposed a generalized fuzzy data envelopmentmodel with assurance regions, whose lower and upper bounds atgiven levels could be obtained. Entani, Maeda, and Tanaka (2002)

Table 3The expect intervals and the corresponding expected values of the fuzzy correlation coeffi

rL rR

I1 I2 O1 O2 I1 I2

I1 1.0000 0.8404 0.8314 0.6505 1.0000 0.8809I2 0.8404 1.0000 0.8862 0.8737 0.8809 1.0000O1 0.8314 0.8862 1.0000 0.6753 0.8472 0.8882O2 0.6505 0.8737 0.6753 1.0000 0.7175 0.8839

and Wang, Greatbanks, and Yang (2005) also changed fuzzy inputand fuzzy output data into intervals by usinga-level sets, but sug-gested two different interval DEA models. Dia (2004) proposed aFDEA model based on fuzzy arithmetic operations and fuzzy com-parisons between fuzzy numbers. The model requires the decisionmaker to specify a fuzzy aspiration level and a safety a-level so thatthe FDEA model could be transformed into a crisp DEA model forsolution. Wang, Luo, and Liang (2009) constructed two FDEA mod-els from the perspective of fuzzy arithmetic to deal with fuzzinessin input and output data in DEA. The two FDEA models were bothformulated as linear programs and could be solved to determinefuzzy efficiencies of DMUs. Jahanshahloo, Soleimani-damaneh,and Nasrabadi (2004) extended a slack-based measure (SBM) ofefficiency in DEA to fuzzy settings and developed a bi-objectivenonlinear DEA model for FDEA. Among all the approaches to solveFDEA, the most popular approach is a-cut approach. Hatami-Mar-bini, Saati, and Makui (2010) introduced two virtual DMUs calledideal DMU (IDMU) and anti-ideal DMU (ADMU) with fuzzy in-puts-outputs, and evaluated efficiency of DMUs by FDEA. Hatam-i-Marbini, Saati, and Tavana (2010) presented a four-phase FDEAframework based on the theory of displaced ideal. Wang and Chin(2011) proposed a ‘‘fuzzy expected value approach’’ for DEA inwhich fuzzy inputs and fuzzy outputs are first weighted respec-tively, and their expected values then used to measure the optimis-tic and pessimistic efficiencies of DMUs in fuzzy environments.Hsiao, Chern, Chiu, and Chiu (2011) proposed the fuzzy super-effi-ciency slack-based measure DEA model using a-cut approach andanalyze the operational performance of 24 commercial banks fac-ing problems on loan and investment parameters with vague char-acteristics. Majid Zerafat Angiz, Emrouznejad, and Mustafa (2012)introduced an alternative linear programming model that can in-clude some uncertainty information from the intervals within thea-cut approach and proposed the concept of ‘‘local a-level’’ to de-velop a multi-objective linear programming to measure the effi-ciency of DMUs under uncertainty.

In this paper, we use a-cut approach to solve FCCRI and FSBMI.Then, the results of these models are applied to calculate FIME.We propose a new method for calculating the fuzzy correlation

cients.

rEV

O1 O2 I1 I2 O1 O2

0.8472 0.7175 1.0000 0.8606 0.8393 0.68400.8882 0.8839 0.8606 1.0000 0.8872 0.87881.0000 0.6955 0.8393 0.8872 1.0000 0.68540.6955 1.0000 0.6840 0.8788 0.6854 1.0000

Page 3: A concept of fuzzy input mix-efficiency in fuzzy DEA and its application in banking sector

J. Puri, S.P. Yadav / Expert Systems with Applications 40 (2013) 1437–1450 1439

coefficients between fuzzy inputs and fuzzy outputs by using ex-pected value approach. Further, we also propose a new rankingmethod to rank the DMUs on the basis of FIME. All the proposedapproaches are then applied to the banking sector.

The paper is organized as follows: Section 2 presents an over-view of DEA with input-oriented CCR and SBM models, and inputmix-efficiency. Section 3 presents the description of FDEA withFCCRI and FSBMI. Section 4 presents the methodology for solvingFCCRI and FSBMI. Section 5 gives the definition of FIME. Section 6proposes a new method for evaluating the fuzzy correlation coeffi-cients between fuzzy inputs and fuzzy outputs by using expectedvalue approach. A numerical illustration is presented in Section7. Section 8 describes a new ranking method for DMUs. Section 9presents an application of the proposed approaches to the bankingsector. The last Section 10 concludes the findings of our study.

2. Data envelopment analysis (DEA)

Data envelopment analysis (DEA), proposed by Charnes et al.(1978), is a linear programming based non-parametric methodfor evaluating the relative efficiency of DMUs which uses multipleinputs to produce multiple outputs. Since 1978, it has got compre-hensive attention both in theory and applications. Based on theoriginal DEA model (Charnes et al., 1978), various theoreticalextensions have been developed (Banker, Charnes, & Cooper,1984; Charnes, Cooper, Seiford, & Stutz, 1982; Petersen, 1990;Tone, 2001; Cooper, Seiford, & Tone, 2007). DEA is a non-paramet-ric technique to construct a piecewise frontier (surface) over thedata. Efficiency measure is then calculated relative to this frontier.Using this frontier, DEA computes a maximal performance measurefor each DMU relative to that of all other DMUs with the restrictionthat each DMU lies on the efficient (extremal) frontier or is envel-oped by the frontier. There are two types of models in DEA: radialand non-radial. Radial model is represented by the CCR model(Charnes et al., 1978) and non-radial model is represented bySBM model (Tone, 2001). Both CCR and SBM models are orientedmodels in terms of input-oriented and output-oriented. In thispaper, we are taking input-oriented models. Assume that theperformance of a set of n homogeneous DMUs (DMUj; j = 1, . . . ,n)is to be measured. The performance of DMUj is characterized by aproduction process of m inputs (xij; i = 1, . . . , m) to yield s outputs(yrj;r = 1, . . . ,s). Let yrk be the amount of the rth output producedby the kth DMU and xik be the amount of the ith input used bythe kth DMU. Assume that input and output data is positive.

2.1. Input-oriented CCR model (CCRI model)

The CCRI model evaluates the CCR input efficiency of the DMUs.The CCR input efficiency of the kth DMU is denoted by hk

I and is de-fined as

hkI ¼min h

subject to hxik ¼Xn

j¼1

xijgjk þ s�ik 8i;

yrk ¼Xn

j¼1

yrjgjk � sþrk 8r;

gjk P 0; s�ik P 0; sþrk P 0 8i; r; j;

h unrestricted in sign;k ¼ 1; . . . ;n:

ð1Þ

where sþrk is slack in the rth output of the kth DMU; s�ik is slack in theith input of the kth DMU; gjk’s i.e. (gj1,gj2, . . . ,gjn) are non negativevariables forj = 1, 2, . . . ,n. Due to nonzero assumption of the data,we have 0 < hk

I 6 1 for k ¼ 1;2; . . . ;n (Cooper et al., 2007).

A DMUk is CCR input efficient if (i) hkI ¼ 1 and (ii) all slacks are

zero, i.e., s�ik ¼ 0 for i ¼ 1;2; . . . ;m and sþrk ¼ 0 for r ¼ 1;2; . . . ; s.

2.2. Input-oriented SBM model (SBMI model)

The SBMI model evaluates the SBM input efficiency of theDMUs. The SBM input efficiency of the kth DMU is denoted by qk

I

and is defined as

qkI ¼min 1� 1

m

Xm

i¼1

S�ikxik

subject to xik ¼Xn

j¼1

xijkjk þ S�ik 8i;

yrk ¼Xn

j¼1

yrjkjk � Sþrk 8r;

kjk P 0; S�ik P 0; Sþrk P 0 8i; r; j;

k ¼ 1; . . . ;n:

ð2Þ

where Sþrk is slack in the rth output of the kth DMU; S�ik is slack in theith input of the kth DMU; kjk’s i.e. (kj1,kj2, . . . ,kjn) are non negativevariables for j = 1, 2, . . . ,n.

A DMUk is SBM input efficient if (i) qkI ¼ 1 and (ii) all output

slacks are zero, i.e., Sþrk ¼ 0 for r ¼ 1;2; . . . ; s. However, input slacksmay be nonzero.

2.3. Input mix-efficiency (IME)

The IME of the kth DMU is defined as the ratio of SBM input effi-ciency of the kth DMU to CCR input efficiency of the kth DMU. TheIME of the kth DMU is denoted by wk

I and is defined by

wkI ¼

qkI

hkI

: ð3Þ

Due to nonzero assumption of the data, we have 0 < qkI 6 1

and 0 < hkI 6 1 for k ¼ 1;2; . . . ;n. Also qk

I 6 hkI (Tone, 1998). This

implies that 0 < wkI 6 1; and wk

I ¼ 1 if and only if qkI ¼ hk

I holds. Ifwk

I ¼ 1, it shows that DMUk has the most efficient combination of in-puts, even though it may be technically inefficient.

3. Fuzzy data envelopment analysis (FDEA)

Conventional DEA requires crisp input and output data, whichmay not always be available in real world applications. However,in real-world problems, inputs and outputs are often imprecise.To deal with imprecise data, the notion of fuzziness has been intro-duced. The DEA is extended to FDEA in which the imprecision isrepresented by fuzzy sets or fuzzy numbers. Various efforts havebeen made to handle fuzzy input and fuzzy output data in FDEA.Based on fuzzy input and fuzzy output data, both CCR and SBMmodels were extended to fuzzy CCR model (Guh, 2001; Guo &Tanaka, 2001) and fuzzy SBM model (Jahanshahloo et al., 2004;Saati & Memariani, 2009). Since in this paper, we are taking in-put-oriented models, we are defining FCCRI and FSBMI. Assumethat the performance of a set of n homogeneous DMUs (DMUj;j = 1,. . ., n) is to be measured. The performance of DMUj is charac-terized by a production process of m fuzzy inputsð~xij; i ¼ 1; . . . ;mÞ to yield s fuzzy outputs ð~yrj; r ¼ 1; . . . ; sÞ. Let ~yrk

be the amount of the rth fuzzy output produced by the kth DMUand ~xik be the amount of the ith fuzzy input used by the kthDMU. Assume that the fuzzy inputs and fuzzy outputs are positivefuzzy numbers (Nasseri, 2008). In this paper, we have taken thefuzzy inputs and fuzzy outputs as the triangular fuzzy numbers(TFNs) (Chen, 1994).

Page 4: A concept of fuzzy input mix-efficiency in fuzzy DEA and its application in banking sector

1440 J. Puri, S.P. Yadav / Expert Systems with Applications 40 (2013) 1437–1450

Definition 1. A TFN eA is denoted by (a1,a2,a3) and is defined by themembership function leAðxÞ given by

leAðxÞ ¼x�a1

a2�a1; a1 < x 6 a2;

1; x ¼ a2;x�a3

a2�a3; a2 6 x < a3;

0; otherwise:

8>>>><>>>>:

3.1. Input-oriented fuzzy CCR model (FCCRI model)

The FCCRI model evaluates the fuzzy CCR input efficiency of theDMUs. The fuzzy CCR input efficiency of the kth DMU, denoted by~hk

I ; is defined by~hk

I ¼ min ~h

subject to ~h~xik ¼Xn

j¼1

~xijgjk þ ~s�ik 8i;

~yrk ¼Xn

j¼1

~yrjgjk � ~sþrk 8r;

gjk P 0; ~s�ik P ~0; ~sþrk P ~0 8; i; r; j;k ¼ 1; . . . ;n:

ð4Þ

where ~sþrk is the fuzzy slack in the rth fuzzy output of the kth DMU;~s�ik is the fuzzy slack in the ith fuzzy input of the kth DMU;gjk’s, i.e.,(gj1,gj2, . . . ,gjn) are non-negative variables for j = 1,2, . . . ,n.

3.2. Input-oriented fuzzy SBM model (FSBMI model)

The FSBMI model evaluates the fuzzy SBM input efficiency ofthe DMUs. The fuzzy SBM input efficiency of the kth DMU, denotedby ~qk

I ; is defined by

~qkI ¼min 1� 1

m

Xm

i¼1

eS�i~xik

subject to ~xik ¼Xn

j¼1

~xijkjk þ ~S�ik 8i;

~yrk ¼Xn

j¼1

~yrjkjk � eSþrk 8r;

kjk P 0; eS�ik P ~0; eSþrk P ~0 8i; r; j;

k ¼ 1; . . . ;n:

ð5Þ

where eSþrk is the fuzzy slack in the rth fuzzy output of the kth DMU;eS�ik is the fuzzy slack in the ith fuzzy input of the kth DMU; kjk’s, i.e.,(kj1,kj2, . . . ,kjn) are non-negative variables for j = 1, 2, . . . ,n.

4. Methodology for solving FCCRI model and FSBMI model

Kao and Liu (2000a) developed a procedure to transform a FDEAmodel to a family of crisp DEA models by applying the a-cuts andZadeh’s extension principle (Zadeh, 1975). Now, we will describethe procedure to convert FCCRI and FSBMI models into a familyof crisp DEA models. Let Sð~xijÞ and Sð~yrjÞ be the support of m fuzzyinputs ð~xij; i ¼ 1;2; . . . ;mÞ and s fuzzy outputs ð~yrj; r ¼ 1;2; . . . ; sÞof DMUj (j = 1, . . . ,n) respectively, given bySð~xijÞ ¼ fxijjl~xij

ðxijÞ > 0g and Sð~yrjÞ ¼ fyrjjl~yrjðyrjÞ > 0g: ð6Þ

The a- cuts of ~xij and ~yrj are respectively defined as

ð~xijÞa ¼ fxij 2 Sð~xijÞjl~xijðxijÞP ag ¼ ðxijÞLa; ðxijÞUa

h i8i; j ð7aÞ

¼ minxij

fxij 2 Sð~xijÞjl~xijðxijÞP ag;

�max

xij

fxij 2 Sð~xijÞjl~xijðxijÞP ag

�8i; j ð7bÞ

and

ð~yrjÞa ¼ fyrj 2 Sð~yrjÞjl~yrjðyrjÞP ag ¼ ðyrjÞ

La; ðyrjÞ

Ua

h i8r; j ð8aÞ

¼ minyrj

fyrj 2 Sð~yrjÞjl~yrjðyrjÞP ag;

�max

yrj

fyrj 2 Sð~yrjÞjl~yrjðyrjÞP ag

�8r; j; ð8bÞ

where 0 < a 6 1.Further, FCCRI and FSBMI models can easily be transformed into

crisp models by using a-cuts given in (7b) and (8b). Owing to theinput and output data being fuzzy numbers, the efficiency scoresare also fuzzy numbers. Let them be represented by ~hk

I and ~qkI with

the membership functions l~hkI

and l~qkI

respectively. Let S ~hkI

� �and

S ~qkI

� �be the support of the fuzzy efficiency scores ~hk

I and ~qkI of

the kth DMU respectively, given by

S ~hkI

� �¼ hk

I jl~hkI

hkI

� �> 0

n oand S ~qk

I

� �¼ qk

I jl~qkI

qkI

� �> 0

n oð9Þ

The a-cuts of ~hkI and ~qk

I are respectively defined as

~hkI

� �a¼ hk

I 2 S ~hkI

� �jl~hk

Ihk

I

� �P a

n o¼ hk

I

� �L

a; hk

I

� �U

a

� �8i; j ð10aÞ

¼ minhk

I

hkI 2 S ~hk

I

� �jl~hk

Ihk

I

� �P a

n o;

"

maxhk

I

hkI 2 S ~hk

I

� �jl~hk

Ihk

I

� �P a

n o#8i; j ð10bÞ

and

~qkI

� �a ¼ qk

I 2 S ~qkI

� �jl~qk

Iqk

I

� �P a

n o¼ qk

I

� �L

a; qkI

� �U

a

h i8i; j ð11aÞ

¼ minqk

I

qkI 2 S ~qk

I

� �jl~qk

Iqk

I

� �P a

n o;

"

maxqk

I

qkI 2 S ~qk

I

� �jl~qk

Iqk

I

� �P a

n o#8i; j; ð11bÞ

where 0 < a 6 1.

hkI

� �L

a¼ min

ðxij ÞLa6xij6ðxij Þ

Ua

ðyrj ÞLa6yrj6ðyrj Þ

Ua

8i;r;j

min h

subject to hxik¼Xn

j¼1

xijgjkþs�ik 8i;

yrk¼Xn

j¼1

yrjgjk�sþrk 8r;

gjk P0; s�ik P0; sþrk P0 8i;r;j;

8>>>>>>>><>>>>>>>>:

9>>>>>>>>=>>>>>>>>;ð12aÞ

hkI

� �U

a¼ max

ðxij ÞLa6xij6ðxij Þ

Ua

ðyrj ÞLa6yrj6ðyrj Þ

Ua

8i;r;j

min h

subject to hxik¼Xn

j¼1

xijgjkþs�ik 8i;

yrk¼Xn

j¼1

yrjgjk�sþrk 8r;

gjk P0; s�ik P0; sþrk P0 8i;r;j

8>>>>>>>>><>>>>>>>>>:

9>>>>>>>>>=>>>>>>>>>;ð12bÞ

and

qkI

� �L

a¼ minðxij Þ

La6xij6ðxij Þ

Ua

ðyrj ÞLa6yrj6ðyrj Þ

Ua

8i;r;j

min 1� 1m

Xm

i¼1

S�ikxik

subject to xik¼Xn

j¼1

xijkjkþS�ik 8i;

yrk¼Xn

j¼1

yrjkjk�Sþrk 8r;

kjk P0; S�ik P0; Sþrk P0 8i;r;j;

8>>>>>>>>>>>>><>>>>>>>>>>>>>:

9>>>>>>>>>>>>>=>>>>>>>>>>>>>;ð13aÞ

Page 5: A concept of fuzzy input mix-efficiency in fuzzy DEA and its application in banking sector

J. Puri, S.P. Yadav / Expert Systems with Applications 40 (2013) 1437–1450 1441

qkI

� �U

a ¼ maxðxij Þ

La6xij6ðxij Þ

Ua

ðyrj ÞLa6yrj6ðyrjÞ

Ua

8i;r;j

min 1� 1m

Xm

i¼1

S�ikxik

subject to xik¼Xn

j¼1

xijkjkþS�ik 8i;

yrk¼Xn

j¼1

yrjkjk�Sþrk 8r;

kjk P0; S�ik P0; Sþrk P0 8i;r; j:

8>>>>>>>>>>>>>><>>>>>>>>>>>>>>:

9>>>>>>>>>>>>>>=>>>>>>>>>>>>>>;

ð13bÞ

Further, following the Pareto’s efficiency concept, we can find the‘minimum efficiency’ of specific or targeted DMU. For this, we usetargeted DMU with lower bound outputs and other DMUs withupper bound outputs, and targeted DMU with upper bound inputsand other DMUs with lower bound inputs. Similarly, if we findthe targeted DMU to have the ‘maximum efficiency’, we use tar-geted DMU with upper bound outputs and other DMUs with lowerbound outputs, and targeted DMU with lower bound inputs andother DMUs with upper bound inputs. Thus, models 12a,12b,13aand 13b reduce to the following models:

hkI

� �L

a¼min h

subject to hðxikÞUa ¼Xn

j¼1;j–k

ðxijÞLagjk þ ðxikÞUagjk þ s�ik� �U 8i;

ðyrkÞLa ¼

Xn

j¼1;j–k

ðyrjÞUagjk þ ðyrkÞ

Lagjk � sþrk

� �L 8r;

gjk P 0; s�ik� �U P 0; sþrk

� �L P 0 8i; r; j;

ð14aÞ

hkI

� �U

a¼ min h

subject to hðxikÞLa ¼Xn

j¼1;j–k

ðxijÞUagjk þ ðxikÞLagjk þ s�ik� �L8i;

ðyrkÞUa ¼

Xn

j¼1;j–k

ðyrjÞLagjk þ ðyrkÞ

Uagjk � sþrk

� �U 8r;

gjk P 0; s�ik� �L P 0; sþrk

� �U P 0 8i; r; j

ð14bÞ

and

qkI

� �L

a ¼min 1� 1m

Xm

i¼1

S�ik� �U

ðxikÞUa

subject to ðxikÞUa ¼Xn

j¼1;j–k

ðxijÞLakjk þ ðxikÞUa kjk þ S�ik� �U 8i;

ðyrkÞLa ¼

Xn

j¼1;j–k

ðyrjÞUa kjk þ ðyrkÞ

Lakjk � Sþrk

� �L 8r;

kjk P 0; S�ik� �U P 0; Sþrk

� �LP 0 8i; r; j;

ð15aÞ

qkI

� �U

a ¼ min 1� 1m

Xm

i¼1

S�ik� �L

ðxikÞLa

subject to ðxikÞLa ¼Xn

j¼1;j–k

ðxijÞUa kjk þ ðxikÞLakjk þ S�ik� �L 8i;

ðyrkÞU ¼

Xn

j¼1;j–k

ðyrjÞLakjk þ ðyrkÞ

Ua kjk � Sþrk

� �U 8r;

kjk P 0; S�ik� �L P 0; Sþrk

� �UP 0 8i; r; j;

ð15bÞ

The sets of intervals hkI

� �L

a; hk

I

� �U

a

� �ja 2 ð0;1�; k ¼ 1;2; . . . ; n

� and

qkI

� �L

a; qkI

� �U

a

h ija 2 ð0;1�; k ¼ 1;2; . . . ;n

n oreveal the shape of l~hk

I

and l~qkI, respectively, although the exact form of the membership

functions are not known explicitly. In our study, we are taking fuzzyinputs ð~xij; i ¼ 1;2; . . . ;mÞ and fuzzy outputs ð~yrj; r ¼ 1;2; . . . ; sÞ asTFNs. Therefore, the membership functions l~hk

Iand l~qk

Ican be

approximated by the triangular membership functions whose a-cuts are represented by the sets of intervals

hkI

� �L

a; hk

I

� �U

a

� �ja 2 ð0;1�; k ¼ 1;2; . . . ; n

� and qk

I

� �L

a; qkI

� �U

a

h ija 2

nð0;1�; k ¼ 1;2; . . . ;ng; respectively. Thus, the fuzzy efficiencies ~hk

I

and ~qkI can be approximated by TFNs and are defined as:

Definition 2. The fuzzy CCR input efficiency ~hkI of the kth DMU is

defined by its a-cut ~hkI

� �a

which is given by

~hkI

� �a¼ hk

I

� �L

a; hk

I

� �U

a

� �; a 2 ð0;1�; k ¼ 1;2; . . . ;n:

where hkI

� �L

aand hk

I

� �U

aare obtained from the optimal values of

(14a) and (14b) respectively. Further, ~hkI by using a-cuts

~hkI

� �a; a 2 ð0;1� can be approximated by a TFN hk

a; hkb; h

kc

� �.

Definition 3. The fuzzy SBM input efficiency ~qkI of the kth DMU is

defined by its a-cut ~qkI

� �awhich is given by

~qkI

� �a ¼ qk

I

� �L

a; qkI

� �U

a

h i; a 2 ð0;1�; k ¼ 1;2; . . . ;n;

where qkI

� �L

a and qkI

� �U

a are obtained from the optimal values of (15a)and (15b) respectively. Further, ~qk

I by usinga-cuts ~qkI

� �a; a 2 ð0;1�

can be approximated by a TFN qka;qk

b;qkc

� �.

5. Fuzzy input mix-efficiency (FIME)

The FIME of the kth DMU, denoted by ~wkI ; is defined as the ratio

of fuzzy SBM input efficiency of the kth DMU to the fuzzy CCR in-put efficiency of the kth DMU. By using arithmetic operations onTFNs (Chen, 1994), ~wk

I can be defined as

~wkI ¼

~qkI

~hkI

; ~hkI – ~0 ð16Þ

¼qk

a;qkb;q

kc

� �hk

a; hkb; h

kc

� � ¼ qka;q

kb;q

kc

� �� hk

a; hkb; h

kc

� ��1

¼ qka;q

kb;q

kc

� �� 1

hkc

;1hk

b

;1hk

a

!; hk

a > 0

� qka

hkc

;qk

b

hkb

;qk

c

hka

!; hk

a > 0: ð17Þ

6. The proposed method for evaluating the fuzzy correlationcoefficients between fuzzy inputs and fuzzy outputs by usingexpected value approach

To ensure the validity of the crisp DEA model specification, thecorrelation coefficients between inputs and outputs are calculated.If positive inter-correlations are found, the inclusion of the inputsand outputs is justified. This test is called the isotonicity test (Avki-ran, 2006; Tsai, Chen, & Tzeng, 2006). It identifies whether increas-ing amounts of inputs lead to greater outputs. In the present study,we are concerned with the evaluation of ~hk

I ; ~qkI and ~wk

I on the basisof the fuzzy inputs and fuzzy outputs. To ensure the validity of thepresent FDEA models specification, it is mandatory to calculate thefuzzy correlation coefficients between the fuzzy inputs and fuzzyoutputs. However, in the literature of FDEA, nobody has thrownlight on this important part of the analysis. Therefore, we are pro-posing a method to calculate fuzzy correlation coefficients between

Page 6: A concept of fuzzy input mix-efficiency in fuzzy DEA and its application in banking sector

Table 4a- cuts ~hk

I

� �a

of the fuzzy CCR input efficiency ~hkI for different values of a 2 (0,1].

DMU a = 0(L,U)

a = 0.1[L,U]

a = 0.2[L,U]

a = 0.3[L,U]

a = 0.4[L,U]

a = 0.5[L,U]

a = 0.6[L,U]

a = 0.7[L,U]

a = 0.8[L,U]

a = 0.9[L,U]

a = 1[L,U]

1 L 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000U 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000

2 L 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000U 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000

3 L 0.8398 0.8430 0.8463 0.8496 0.8530 0.8569 0.8619 0.8670 0.8721 0.8774 0.8827U 0.9909 0.9717 0.9530 0.9348 0.9250 0.9160 0.9073 0.9007 0.8948 0.8887 0.8827

4 L 0.9582 0.9638 0.9694 0.9751 0.9808 0.9865 0.9923 0.9982 1.0000 1.0000 1.0000U 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000

5 L 0.6813 0.6891 0.6969 0.7048 0.7128 0.7209 0.7290 0.7373 0.7456 0.7541 0.7635U 0.9604 0.9362 0.9127 0.8898 0.8675 0.8458 0.8246 0.8039 0.7838 0.7725 0.7635

6 L 0.7563 0.7635 0.7709 0.7784 0.7860 0.7938 0.8017 0.8097 0.8179 0.8263 0.8348U 0.9077 0.8999 0.8921 0.8845 0.8675 0.8693 0.8624 0.8555 0.8486 0.8417 0.8348

7 L 0.7974 0.8062 0.8151 0.8241 0.8332 0.8424 0.8516 0.8610 0.8704 0.8852 0.9020U 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9810 0.9604 0.9404 0.9209 0.9020

8 L 0.7277 0.7343 0.7409 0.7476 0.7544 0.7612 0.7681 0.7750 0.7820 0.7891 0.7963U 0.9853 0.9570 0.9298 0.9036 0.8782 0.8538 0.8364 0.8257 0.8154 0.8056 0.7963

9 L 0.7880 0.7951 0.8023 0.8096 0.8169 0.8243 0.8460 0.8731 0.9011 0.9302 0.9604U 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9854 0.9604

10 L 0.8063 0.8122 0.8182 0.8243 0.8305 0.8369 0.8434 0.8500 0.8567 0.8636 0.8707U 0.9578 0.9486 0.9395 0.9305 0.9217 0.9129 0.9042 0.8957 0.8872 0.8789 0.8707

11 L 0.8855 0.8922 0.8990 0.9058 0.9127 0.9196 0.9266 0.9336 0.9407 0.9479 0.9551U 1.0000 1.0000 1.0000 1.0000 1.0000 0.9950 0.9869 0.9788 0.9708 0.9629 0.9551

12 L 0.8426 0.8500 0.8575 0.8650 0.8726 0.8803 0.8880 0.8957 0.9104 0.9339 0.9582U 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9910 0.9582

Table 5a-cut ~qk

I

� �a of the fuzzy SBM input efficiency ~qk

I for different values of a 2 (0,1].

DMU a = 0(L,U)

a = 0.1[L,U]

a = 0.2[L,U]

a = 0.3[L,U]

a = 0.4[L,U]

a = 0.5[L,U]

a = 0.6[L,U]

a = 0.7[L,U]

a = 0.8[L,U]

a = 0.9[L,U]

a = 1[L,U]

1 L 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000U 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000

2 L 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000U 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000

3 L 0.7562 0.7640 0.7720 0.7802 0.7885 0.7976 0.8079 0.8185 0.8294 0.8406 0.8522U 0.9735 0.9603 0.9472 0.9345 0.9220 0.9098 0.8978 0.8860 0.8745 0.8632 0.8522

4 L 0.8531 0.8627 0.8725 0.8825 0.8928 0.9032 0.9139 0.9348 0.9542 0.9739 1.0000U 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000

5 L 0.5998 0.6134 0.6274 0.6418 0.6566 0.6719 0.6876 0.7038 0.7206 0.7378 0.7556U 0.9035 0.8872 0.8712 0.8557 0.8404 0.8255 0.8109 0.7966 0.7827 0.7690 0.7556

6 L 0.6102 0.6187 0.6274 0.6361 0.6451 0.6544 0.6638 0.6735 0.6833 0.6934 0.7038U 0.8071 0.7959 0.7850 0.7742 0.7636 0.7532 0.7429 0.7329 0.7230 0.7133 0.7038

7 L 0.7789 0.7895 0.8003 0.8113 0.8225 0.8340 0.8457 0.8576 0.8698 0.8822 0.8948U 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9475 0.9339 0.9206 0.9076 0.8948

8 L 0.6664 0.6759 0.6857 0.6957 0.7060 0.7166 0.7275 0.7387 0.7501 0.7619 0.7740U 0.9378 0.9192 0.9011 0.8835 0.8665 0.8500 0.8339 0.8183 0.8031 0.7883 0.7740

9 L 0.7530 0.7660 0.7795 0.7934 0.8077 0.8226 0.8379 0.8537 0.8701 0.8871 0.9046U 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9772 0.9203 0.9046

10 L 0.6860 0.6944 0.7031 0.7119 0.7210 0.7303 0.7399 0.7497 0.7597 0.7700 0.7805U 0.9052 0.8916 0.8783 0.8652 0.8524 0.8398 0.8275 0.8154 0.8035 0.7919 0.7805

11 L 0.7531 0.7635 0.7740 0.7848 0.7957 0.8069 0.8183 0.8299 0.8418 0.8538 0.8661U 1.0000 1.0000 1.0000 1.0000 1.0000 0.9400 0.9152 0.9026 0.8903 0.8781 0.8661

12 L 0.8029 0.8147 0.8268 0.8392 0.8520 0.8650 0.8785 0.8923 0.9065 0.9210 0.9360U 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9771 0.9556 0.9360

1442 J. Puri, S.P. Yadav / Expert Systems with Applications 40 (2013) 1437–1450

the fuzzy inputs and fuzzy outputs by using expected value ap-proach (Hung & Wu, 2001). In this approach, we calculate expectedintervals and expected values of fuzzy correlation coefficients.

6.1. Expected interval and expected value of a fuzzy number

Let eA be a fuzzy number with membership function leAðxÞ givenby

leAðxÞ ¼feAðxÞ; a1 6 x 6 a2;

1; a2 6 x 6 a3;

geAðxÞ; a3 6 x 6 a4;

0; otherwise;

8>>>><>>>>:

where feA is an increasing function and geA is a decreasing function.The expected interval of eA is a crisp interval EIðeAÞ given byEIðeAÞ ¼ ½ELðeAÞ; ERðeAÞ� (Hung and Wu, 2001), where

ELðeAÞ ¼ a2 �Z a2

a1

feAðxÞdx and ERðeAÞ ¼ a3 þZ a4

a3

geAðxÞdx:

The expected value is defined by

EVðeAÞ ¼ ELðeAÞ þ ERðeAÞ2

:

Let eA be a TFN denoted by (a1,a2,a3). The expected interval and ex-pected value of eA are given by

Page 7: A concept of fuzzy input mix-efficiency in fuzzy DEA and its application in banking sector

J. Puri, S.P. Yadav / Expert Systems with Applications 40 (2013) 1437–1450 1443

EIðeAÞ ¼ a1 þ a2

2;a2 þ a3

2

h iand

EVðeAÞ ¼ a1 þ 2a2 þ a3

4; respectively: ð18Þ

Fig. 1b. Shape of membership functions of ~hkI for k = 7 to 12.

Fig. 1c. Shape of membership functions of ~qkI for k =1 to 6.

Fig. 1d. Shape of membership functions of ~qkI for k =7 to 12.

6.2. Fuzzy correlation coefficient between fuzzy input ~x and fuzzyoutput ~y

The crisp correlation coefficient denoted by r(x,y) between twosets of crisp data x = {x1,x2, . . . ,xn} and y = {y1,y2, . . . ,yn} is calculatedby the following formula:

rðx; yÞ ¼ nPn

i¼1xiyi �Pn

i¼1xiPn

i¼1yiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffinPn

i¼1x2i �

Pni¼1xi

� �2q ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

nPn

i¼1y2i �

Pni¼1yi

� �2q : ð19Þ

However, in real world applications due to non-availability of exactdata, the data may be imprecise or fuzzy. Hence, the fuzzy correla-tion coefficient between two sets ~x and ~y of fuzzy dataf~x1; ~x2; . . . ; ~xng and f~y1; ~y2; . . . ; ~yng respectively, denoted by ~rð~x; ~yÞ;can be calculated by the following formula:

~rð~x; ~yÞ ¼nPn

i¼1ð~xi � ~yiÞ� �

HPn

i¼1~xi� �

�Pn

i¼1~yi� �� �ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

nPn

i¼1~x2i H

Pni¼1~xi

� �2q ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

nPn

i¼1~y2i H

Pni¼1~yi

� �2q : ð20Þ

With the assumption of positive fuzzy data and by using arithmeticoperations of addition, subtraction, multiplication and division onfuzzy numbers (Chen, 1994), we can find out the fuzzy correlationcoefficient between two sets of fuzzy data by solving (20). Bansal(2010) defined the square root of a TFN which is non linear arithme-tic operation. Thus, we can find out the square root operation in(20). As all the fuzzy inputs and fuzzy outputs are positive andare represented by TFNs, the square of a positive TFN eA ¼ ða; b; cÞcan be defined as

eA2 ¼ eA � eA ¼ ða; b; cÞ � ða; b; cÞ¼ ðminða2; ac; c2Þ; b2

;maxða2; ac; c2ÞÞ ¼ ða2; b2; c2Þ:

But, it is a difficult task to apply (20) if data set is large. So, in orderto obtain the fuzzy correlation coefficient between the sets of fuzzydata, we propose a method to calculate the fuzzy correlation coeffi-cient using the expected value approach. Firstly, find the expectedintervals of fuzzy data sets, i.e., find EIð~xiÞ ¼ xi

L; xiR

� �,

EIð~yiÞ ¼ yiL; y

iR

� �; i ¼ 1;2; . . . n and then by using these expected

intervals find the expected interval of the fuzzy correlation coeffi-cient ~rð~x; ~yÞ denoted by rEIð~x; ~yÞ ¼ ½rLð~x; ~yÞ; rRð~x; ~yÞ�; where

rLð~x; ~yÞ ¼nPn

i¼1xiLyi

L �Pn

i¼1xiL

Pni¼1yi

LffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffinPn

i¼1ðxiLÞ

2 �Pn

i¼1xiL

� �2q ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

nPn

i¼1ðyiLÞ

2 �Pn

i¼1yiL

� �2q

ð21Þ

Fig. 1a. Shape of membership functions of ~hkI for k = 1 to 6.

and

rRð~x; ~yÞ ¼nPn

i¼1xiRyi

R �Pn

i¼1xiR

Pni¼1yi

RffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffinPn

i¼1 xiR

� �2 �Pn

i¼1xiR

� �2q ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

nPn

i¼1ðyiRÞ

2 �Pn

i¼1yiR

� �2q

ð22Þ

The lower and upper bounds of rEIð~x; ~yÞ satisfy the followingproperties:

1. �1 6 rLð~x; ~yÞ 6 1 and �1 6 rRð~x; ~yÞ 6 1.2. rLð~x; ~yÞ ¼ rLð~y; ~xÞ and rRð~x; ~yÞ ¼ rRð~y; ~xÞ.3. rLð~x; ~yÞ ¼ 1 and rRð~x; ~yÞ ¼ 1 if ~x ¼ ~y.

The expected value of the fuzzy correlation coefficient is de-noted by rEV ð~x; ~yÞ and is defined by

rEV ð~x; ~yÞ ¼ rLð~x; ~yÞ þ rRð~x; ~yÞ2

ð23Þ

Page 8: A concept of fuzzy input mix-efficiency in fuzzy DEA and its application in banking sector

Table 7The defuzzified values of ~hk

I , ~qkI and ~wk

I by using the COG method and ranking of theDMUs.

DMU dCOG~hk

I

� �Rank dCOG ~qk

I

� �Rank dCOG

~wkI

� �Rank

1 1.000000 1 1.000000 1 1.000000 -2 1.000000 1 1.000000 1 1.000000 -3 0.904413 8 0.860624 7 0.962657 -4 0.986404 3 0.951356 3 0.965559 -5 0.801730 11 0.752967 12 0.980094 -6 0.832920 12 0.707026 11 0.860847 -7 0.899795 5 0.891256 8 1.008355 -8 0.836413 9 0.792696 10 0.979001 -9 0.916128 6 0.885870 6 0.987985 -

10 0.878263 10 0.790567 9 0.911769 -

1444 J. Puri, S.P. Yadav / Expert Systems with Applications 40 (2013) 1437–1450

and it also satisfies the following properties:

1. �1 6 rEV ð~x; ~yÞ 6 1:2. rEV ð~x; ~yÞ ¼ rEVð~y; ~xÞ:3. rEV ð~x; ~yÞ ¼ 1if ~x ¼ ~y:

If the value of rEV ð~x; ~yÞ for each ~x and ~y is positive, then the FDEAmodel is said to be consistent and inclusion of fuzzy inputs andfuzzy outputs is justified.

Remark 1. If the data is crisp, the proposed approach gives thesimilar results as given by the conventional correlation coefficientformula defined in Eq. (19). In case of crisp data,

11 0.946851 7 0.873083 4 0.929775 -12 0.933581 4 0.912983 5 0.988885 -

rLð~x; ~yÞ ¼ rRð~x; ~yÞ ¼ rEV ð~x; ~yÞ ¼ rðx; yÞ:

Table 8Ranking of the DMUs on the basis of ~wk

I .

DMU dCOG~hk

I

� �Rank dCOG ~qk

I

� �Rank dCOG

~wkI

� �Rank

1 1.000000 1 1.000000 1 1.000000 12 1.000000 1 1.000000 1 1.000000 13 0.904413 8 0.860624 7 0.962657 54 0.986404 3 0.951356 3 0.965559 35 0.801730 11 0.752967 12 0.980094 96 0.832920 12 0.707026 11 0.860847 127 0.899795 5 0.891256 8 1.008355 68 0.836413 9 0.792696 10 0.979001 109 0.916128 6 0.885870 6 0.987985 8

10 0.878263 10 0.790567 9 0.911769 1111 0.946851 7 0.873083 4 0.929775 312 0.933581 4 0.912983 5 0.988885 7

7. Numerical illustration

The CCR input efficiency hkI ; SBM input efficiency qk

I and Mix in-put efficiency wk

I of the twelve DMUs with crisp data are evaluatedand results are shown in Table 1.

The results show that DMU1, DMU2 and DMU4 are CCR input effi-cient, SBM input efficient and also Mix input efficient. There arenine DMUs (DMU3,DMU5,DMU6,DMU7,DMU8,DMU9,DMU10, -DMU11, and DMU12) which are input mix-inefficient. The inputmix-inefficiency represents the degree to which the input mixshould change to become fully efficient. The DMU6 is the mostmix-inefficient DMU with rank 12. The results shown in Table 1are for crisp data. However, in real world applications data maynot always be available crisply. Actually, in real world problems,inputs and outputs are often imprecise or fuzzy. Thus to deal withactual problems, we fuzzify the crisp data given in Table 1 and rep-resent the fuzzy data in terms of TFNs. The fuzzified data is shownin Table 2.

7.1. The fuzzy correlation coefficient between fuzzy inputs and fuzzyoutputs

To ensure the validity of the FDEA model specification, the ex-pected intervals and expected values of the fuzzy correlation coef-ficients between the fuzzy inputs and fuzzy outputs are calculatedby using the proposed method given in Section 6. The expectedinterval values and the corresponding expected values of the fuzzycorrelation coefficients between the sets of fuzzy data (given in Ta-ble 2) are shown in the matrix form in Table 3.

It can be seen from Table 3 that the left and right bounds of eachexpected interval and the corresponding expected value are posi-tive. Thus, the inclusion of the fuzzy inputs and fuzzy outputs is

Table 6The values of ~hk

I , ~qkI and ~wk

I approximated as the TFNs.

DMU ~hkI

1 (1.0000,1.0000,1.0000)2 (1.0000,1.0000,1.0000)3 (0.8397,0.8827,0.9909)4 (0.9582,1.0000,1.0000)5 (0.6813,0.7635,0.9604)6 (0.7563,0.8348,0.9077)7 (0.7974,0.9020,1.0000)8 (0.7277,0.7963,0.9853)9 (0.7880,0.9604,1.0000)

10 (0.8063,0.8707,0.9578)11 (0.8855,0.9551,1.0000)12 (0.8426,0.9582,1.0000)

justified and the FDEA model which is taken in our present studyis consistent.

7.2. FCCRI, FSBMI and FIME evaluations

The a-cuts ~hkI

� �a

and ~qkI

� �aof fuzzy CCR input efficiency ~hk

I andfuzzy SBM input efficiency ~qk

I of the twelve DMUs are evaluatedby using Models 14a, 14b, 15a and 15b at different values of aand are shown in Tables 4 and 5 respectively.

The graphical representations of the fuzzy efficiencies ~hkI and ~qk

I

for the kth DMU which are obtained by using a-cuts ~hkI

� �a

and

~qkI

� �a are shown in Figs. 1a, 1b, 1c and 1d. It can be seen from

the figures that the shape of the membership functions of ~hkI and

~qkI are approximated as triangular membership functions. There-

fore, the fuzzy efficiencies ~hkI and ~qk

I for the kth DMU are obtained

by using a-cuts ~hkI

� �a

and ~qkI

� �a given in Tables 4 and 5, and the

~qkI

~wkI

(1.0000,1.0000,1.0000) (1.00000,1.00000,1.00000)(1.0000,1.0000,1.0000) (1.00000,1.00000,1.00000)(0.7562,0.8522,0.9735) (0.76314,0.96540,1.15943)(0.8531,1.0000,1.0000) (0.85307,1.00000,1.04360)(0.5998,0.7556,0.9035) (0.62458,0.98967,1.32604)(0.6102, 0.7038,0.8071) (0.67223,0.84306,1.06725)(0.7789,0.8948,1.0000) (0.77893,0.99211,1.25403)(0.6664,0.7740,0.9378) (0.67633,0.97191,1.28877)(0.7530, 0.9046,1.0000) (0.75297,0.94195,1.26904)(0.6860, 0.7805,0.9052) (0.71618,0.89647,1.12266)(0.7531,0.8661,1.0000) (0.75311,0.90686,1.12935)(0.8029, 0.9360,1.0000) (0.80293,0.97685,1.18687)

Page 9: A concept of fuzzy input mix-efficiency in fuzzy DEA and its application in banking sector

J. Puri, S.P. Yadav / Expert Systems with Applications 40 (2013) 1437–1450 1445

value of ~wkI is obtained by using Eq. (17). Table 6 presents the val-

ues of ~hkI , ~qk

I and ~wkI approximated as TFNs.

Table 6 indicates that ~0 < ~hkI 6

~1 and ~0 < ~qkI 6

~1. But due to thedivision operation of TFNs, the right part of ~wk

I take the valuesgreater than or equal to 1. So, the FIME for each DMU does notlie between ~0 and ~1; and thus the ranking of the DMUs on the basisof FIME becomes very difficult.

8. A new ranking method for DMUs

For ranking the DMUs on the basis of ~hkI , ~qk

I and ~wkI ; we use

defuzzification method. The aim of the defuzzification is to deter-mine a real value which corresponds to the fuzzy number. In liter-ature, there are various methods to defuzzify a fuzzy number(Kataria, 2010).

Let eA be a fuzzy number defined on R. Then the height of a fuzzy

number eA denoted by hðeAÞ is defined as hðeAÞ ¼ supx2RleAðxÞ. Let

MðeAÞ ¼ x 2 RjleAðxÞ ¼ hðeAÞn obe the set of all those points for

which membership value is equal to the height of eA: Then someof the popularly known deffuzification methods (Kataria, 2010)are defined as follows:

(i) Middle/Mean of maximum (MOM) method – The MOMmethod gives the mean of all those points where member-ship value is maximum. Mathematically, the defuzzifiedvalue dMOMðeAÞ of eA is given by

Table 9Input an

DMU

123456789

1011121314151617

Source:

Table 1The exp

I1I2I3O1O2

dMOMðeAÞ¼X

x2MðeAÞxjMðeAÞj where j:j stands for the cardinality of the set:

d output data of SBOP in various districts of Punjab for the period 2010–2011.

District Name Inputs

Labour Fixed assets

Amritsar (198,201,202) 2592.68Bathinda (553,559,563) 5679.69Faridkot (130,152,160) 1268.55Fatehgarh Sahib (218,221,222) 2325.50Ferozepur (170,173,175) 1724.56Gurdaspur (125,134,135) 1188.89Hoshiarpur (157,160,165) 1985.85Jalandhar (287,291,293) 3443.90Kapurthala (140,145,148) 1662.71Ludhiana (777,781,784) 7326.59Mansa (150,157,159) 1268.78Moga (62,66,70) 681.55Muktsar (106,111,113) 787.72Nawan Shahar (86,115,125) 810.04Patiala (1999,2004,2006) 42939.53Ropar (260,276,280) 2503.95Sangrur (570,576,579) 5955.68

IT Services Department, State Bank of Patiala, Head Office, The Mall, Patiala.

0ected intervals and the corresponding expected values of the fuzzy correlation coe

rL rR

I1 I2 I3 O1 O2 I1 I2 I3

1.0000 0.9710 0.9783 0.7084 0.9024 1.0000 0.9712 0.90.9710 1.0000 0.9935 0.5424 0.7879 0.9712 1.0000 0.90.9783 0.9935 1.0000 0.5787 0.8180 0.9784 0.9935 1.00.7084 0.5424 0.5787 1.0000 0.9370 0.7073 0.5424 0.50.9024 0.7879 0.8180 0.9370 1.0000 0.9015 0.7879 0.8

(ii) Largest of maximum (LOM) method – The LOM methodresults into the largest of all those points where membershipvalue is maximum. Mathematically, the defuzzified valuedLOMðeAÞ of eA is given by

fficients

784935000787183

dLOMðeAÞ ¼maxfxjx 2 MðeAÞg:

(iii) Smallest of maximum (SOM) method – The SOM method

gives the minimum value of all those points where member-ship value is maximum. Mathematically, the defuzzifiedvalue dSOMðeAÞ of eA is given by

dSOMðeAÞ ¼minfxjx 2 MðeAÞg:

(iv) Bisector/Centre of area (COA) method – The bisector method

gives the value that divides the region into two sub-regionsof equal area. Mathematically, the defuzzified value dCOAðeAÞof eA is given by

Z dCOAðeAÞxmin

leAðxÞdx ¼Z xmax

dCOAðeAÞ leAðxÞdx:

(v) Centroid/Centre of gravity (COG) method – The COG methoddetermines the centre of gravity of a fuzzy number. Mathe-matically, COG method defines the defuzzified value dCOGðeAÞas

dCOGðeAÞ ¼R

x x � leAðxÞdxRx leAðxÞdx

:

The results of MOM, LOM and SOM are the same in case of thefuzzy numbers having unique maximum of the membership func-tion. In this paper, we are using TFNs which have a unique maxi-

Outputs

Total expenses Interest income Other income

(55,58.71,63) 53.27 4.5(101, 105.68,109) 113.04 12.77(29,35.80,39) 36.23 3.89(48,52.08,55) 64 6.56(30,35.00,38) 55.42 6.37(33,36.04,40) 26.65 2.4(47,52.47,55) 20.5 2.88(112,118.55,122) 77.25 9.79(42,49.13,55) 19.56 4.39(156,160.71,165) 403.31 32.31(24,30.09,37) 34 4.45(12,15.76,20) 15.8 1.71(17,20.45,24) 29.52 3.08(30,33.16,37) 9.28 0.85(697,701.63,704) 237.18 35.1(96,99.91,104) 50.53 7.83(140, 143.91,147) 158 15.53

.

rEV

O1 O2 I1 I2 I3 O1 O2

0.7073 0.9015 1.0000 0.9711 0.9784 0.7078 0.90200.5424 0.7879 0.9711 1.0000 0.9935 0.5424 0.78790.5787 0.8183 0.9784 0.9935 1.0000 0.5787 0.81811.0000 0.9370 0.7078 0.5424 0.5787 1.0000 0.93700.9370 1.0000 0.9020 0.7879 0.8181 0.9370 1.0000

Page 10: A concept of fuzzy input mix-efficiency in fuzzy DEA and its application in banking sector

Table 11a-cuts ~hk

I

� �a

of the fuzzy CCR input efficiency ~hkI for different values of a 2 (0,1].

DMU a = 0(L,U)

a = 0.1[L,U]

a = 0.2[L,U]

a = 0.3[L,U]

a = 0.4[L,U]

a = 0.5[L,U]

a = 0.6[L,U]

a = 0.7[L,U]

a = 0.8[L,U]

a = 0.9[L,U]

a = 1[L,U]

1 L 0.5357 0.5363 0.5368 0.5374 0.5379 0.5384 0.5390 0.5395 0.5401 0.5406 0.5412U 0.5515 0.5504 0.5494 0.5484 0.5473 0.5463 0.5453 0.5442 0.5432 0.5422 0.5412

2 L 0.5619 0.5669 0.5726 0.5760 0.5796 0.5831 0.5866 0.5902 0.5938 0.5974 0.6010U 0.6457 0.6410 0.6364 0.6319 0.6273 0.6229 0.6184 0.6140 0.6097 0.6053 0.6010

3 L 0.6954 0.6954 0.6954 0.6954 0.6954 0.6954 0.6954 0.6954 0.6954 0.6954 0.6954U 0.7261 0.7137 0.7018 0.6954 0.6954 0.6954 0.6954 0.6954 0.6954 0.6954 0.6954

4 L 0.7106 0.7113 0.7120 0.7127 0.7134 0.7141 0.7147 0.7154 0.7161 0.7168 0.7175U 0.7302 0.7289 0.7276 0.7263 0.7251 0.7238 0.7225 0.7213 0.7200 0.7188 0.7175

5 L 0.8754 0.8768 0.8783 0.8797 0.8812 0.8827 0.8841 0.8856 0.8871 0.8949 0.9053U 1.0000 1.0000 1.0000 1.0000 1.0000 0.9879 0.9704 0.9534 0.9369 0.9208 0.9053

6 L 0.4578 0.4578 0.4578 0.4578 0.4578 0.4578 0.4578 0.4578 0.4578 0.4578 0.4578U 0.4659 0.4624 0.4589 0.4578 0.4578 0.4578 0.4578 0.4578 0.4578 0.4578 0.4578

7 L 0.4198 0.4212 0.4228 0.4243 0.4258 0.4273 0.4289 0.4304 0.4320 0.4335 0.4351U 0.4451 0.4441 0.4431 0.4421 0.4411 0.4401 0.4391 0.4381 0.4371 0.4361 0.4351

8 L 0.8035 0.8045 0.8055 0.8064 0.8074 0.8084 0.8093 0.8103 0.8113 0.8122 0.8132U 0.8277 0.8263 0.8248 0.8233 0.8219 0.8204 0.8190 0.8175 0.8161 0.8146 0.8132

9 L 0.7133 0.7151 0.7170 0.7188 0.7206 0.7225 0.7243 0.7262 0.7281 0.7300 0.7318U 0.7609 0.7579 0.7549 0.7520 0.7490 0.7461 0.7432 0.7403 0.7375 0.7347 0.7318

10 L 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000U 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000

11 L 0.7953 0.7953 0.7953 0.7953 0.7953 0.7953 0.7953 0.7953 0.7953 0.7953 0.7953U 0.9469 0.9211 0.8965 0.8730 0.8507 0.8294 0.8090 0.7953 0.7953 0.7953 0.7953

12 L 0.5875 0.5912 0.5949 0.5986 0.6025 0.6063 0.6102 0.6142 0.6181 0.6222 0.6263U 0.7277 0.7038 0.6812 0.6600 0.6514 0.6471 0.6428 0.6386 0.6345 0.6303 0.6263

13 L 0.8866 0.8866 0.8866 0.8866 0.8866 0.8866 0.8866 0.8866 0.8866 0.8866 0.8866U 0.9252 0.9045 0.8866 0.8866 0.8866 0.8866 0.8866 0.8866 0.8866 0.8866 0.8866

14 L 0.2380 0.2380 0.2380 0.2380 0.2380 0.2380 0.2380 0.2380 0.2380 0.2380 0.2380U 0.2398 0.2380 0.2380 0.2380 0.2380 0.2380 0.2380 0.2380 0.2380 0.2380 0.2380

15 L 0.4208 0.4210 0.4213 0.4216 0.4218 0.4221 0.4223 0.4226 0.4229 0.4231 0.4234U 0.4261 0.4258 0.4255 0.4253 0.4250 0.4247 0.4245 0.4242 0.4239 0.4236 0.4234

16 L 0.7091 0.7091 0.7091 0.7091 0.7091 0.7091 0.7091 0.7091 0.7091 0.7091 0.7091U 0.7308 0.7260 0.7213 0.7167 0.7121 0.7091 0.7091 0.7091 0.7091 0.7091 0.7091

17 L 0.6450 0.6457 0.6464 0.6470 0.6477 0.6484 0.6490 0.6497 0.6504 0.6511 0.6517U 0.6611 0.6602 0.6592 0.6583 0.6573 0.6564 0.6555 0.6545 0.6536 0.6527 0.6517

Table 12a-cut ~qk

I

� �a of the fuzzy SBM input efficiency ~qk

I for different values of a 2 (0,1].

DMU a = 0(L,U)

a = 0.1[L,U]

a = 0.2[L,U]

a = 0.3[L,U]

a = 0.4[L,U]

a = 0.5[L,U]

a = 0.6[L,U]

a = 0.7[L,U]

a = 0.8[L,U]

a = 0.9[L,U]

a = 1[L,U]

1 L 0.4247 0.4260 0.4274 0.4287 0.4301 0.4315 0.4329 0.4343 0.4357 0.4372 0.4387U 0.4543 0.4527 0.4510 0.4494 0.4478 0.4463 0.4447 0.4432 0.4417 0.4402 0.4387

2 L 0.5403 0.5417 0.5431 0.5445 0.5458 0.5472 0.5487 0.5501 0.5515 0.5529 0.5544U 0.5720 0.5701 0.5683 0.5665 0.5647 0.5630 0.5612 0.5595 0.5578 0.5561 0.5544

3 L 0.5872 0.5901 0.5930 0.5960 0.5990 0.6021 0.6052 0.6084 0.6116 0.6148 0.6182U 0.7022 0.6922 0.6827 0.6735 0.6647 0.6562 0.6481 0.6402 0.6326 0.6252 0.6182

4 L 0.6421 0.6439 0.6457 0.6476 0.6495 0.6514 0.6533 0.6553 0.6573 0.6592 0.6612U 0.6893 0.6863 0.6833 0.6804 0.6776 0.6747 0.6720 0.6692 0.6665 0.6639 0.6612

5 L 0.8408 0.8442 0.8477 0.8513 0.8549 0.8585 0.8622 0.8660 0.8698 0.8737 0.8776U 1.0000 0.9865 0.9686 0.9519 0.9356 0.9184 0.9019 0.8956 0.8894 0.8835 0.8776

6 L 0.3917 0.3931 0.3946 0.3961 0.3976 0.3991 0.4007 0.4023 0.4039 0.4056 0.4073U 0.4317 0.4291 0.4265 0.4240 0.4215 0.4190 0.4166 0.4142 0.4119 0.4096 0.4073

7 L 0.3338 0.3350 0.3361 0.3373 0.3384 0.3396 0.3408 0.3420 0.3432 0.3444 0.3457U 0.3623 0.3605 0.3587 0.3570 0.3553 0.3536 0.3520 0.3503 0.3488 0.3472 0.3457

8 L 0.6119 0.6129 0.6140 0.6151 0.6162 0.6173 0.6184 0.6195 0.6206 0.6217 0.6229U 0.6396 0.6378 0.6361 0.6344 0.6327 0.6310 0.6294 0.6277 0.6261 0.6245 0.6229

9 L 0.5658 0.5682 0.5706 0.5731 0.5756 0.5782 0.5808 0.5834 0.5861 0.5889 0.5917U 0.6311 0.6267 0.6224 0.6182 0.6141 0.6101 0.6063 0.6025 0.5988 0.5952 0.5917

10 L 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000U 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000

11 L 0.6830 0.6877 0.6925 0.6976 0.7028 0.7082 0.7138 0.7196 0.7257 0.7321 0.7387U 0.8207 0.8109 0.8015 0.7925 0.7839 0.7756 0.7676 0.7600 0.7526 0.7455 0.7387

12 L 0.5231 0.5277 0.5325 0.5375 0.5426 0.5480 0.5536 0.5594 0.5654 0.5717 0.5783U 0.6553 0.6458 0.6368 0.6282 0.6201 0.6123 0.6049 0.5978 0.5911 0.5845 0.5783

13 L 0.7206 0.7248 0.7292 0.7336 0.7382 0.7430 0.7478 0.7528 0.7580 0.7633 0.7688U 0.8390 0.8309 0.8230 0.8155 0.8082 0.8011 0.7942 0.7876 0.7811 0.7749 0.7688

14 L 0.1708 0.1718 0.1728 0.1738 0.1748 0.1758 0.1769 0.1780 0.1791 0.1802 0.1814U 0.2075 0.2042 0.2011 0.1982 0.1955 0.1928 0.1903 0.1879 0.1856 0.1835 0.1814

15 L 0.2823 0.2827 0.2830 0.2834 0.2837 0.2841 0.2844 0.2848 0.2851 0.2855 0.2859U 0.2895 0.2892 0.2888 0.2884 0.2881 0.2877 0.2873 0.2870 0.2866 0.2862 0.2859

16 L 0.5817 0.5830 0.5843 0.5856 0.5869 0.5882 0.5895 0.5908 0.5922 0.5935 0.5949U 0.6188 0.6163 0.6138 0.6114 0.6089 0.6065 0.6042 0.6018 0.5995 0.5972 0.5949

17 L 0.5821 0.5832 0.5843 0.5854 0.5865 0.5877 0.5888 0.5899 0.5910 0.5921 0.5933U 0.6063 0.6050 0.6036 0.6023 0.6010 0.5997 0.5984 0.5971 0.5958 0.5945 0.5933

1446 J. Puri, S.P. Yadav / Expert Systems with Applications 40 (2013) 1437–1450

Page 11: A concept of fuzzy input mix-efficiency in fuzzy DEA and its application in banking sector

Fig. 2a. Shape of membership functions of ~hkI for k = 1 to 9.

Fig. 2b. Shape of membership functions of ~hkI for k = 10 to 17.

Fig. 2c. Shape of membership functions of ~qkI for k =1 to 9.

Fig. 2d. Shape of membership functions of ~qkI for k =10 to 17.

J. Puri, S.P. Yadav / Expert Systems with Applications 40 (2013) 1437–1450 1447

mum at the modal value, so the results of MOM, LOM and SOM arethe same and are not very much useful. The results of the bisectormethod sometimes, but not always, coincide with the COG meth-od. The COG method is one of the best methods because it takesinto account all the information provided in the fuzzy numberand this method is popularly used in case of TFNs. Keeping in viewthe popularity of the COG method among all other defuzzificationmethods, we are using the COG method to defuzzify the fuzzy effi-ciencies and on the basis of these defuzzified values we are going

to rank the DMUs. Let dCOG~hk

I

� �; dCOG ~qk

I

� �and dCOG

~wkI

� �be the

defuzzified values of ~hkI , ~qk

I and ~wkI respectively which are shown

in Table 7.Table 7 indicates that the ranking of DMUs on the basis of ~hk

I and

~qkI is possible in conventional way, as the values of dCOG

~hkI

� �and

dCOG ~qkI

� �lie between 0 and 1. But there are some DMUs whose

dCOG~wk

I

� �> 1 and thus the ranking of DMUs on the basis of ~wk

I is

still unattainable. Therefore, for doing exact ranking of the DMUson the basis of ~wk

I ; we propose a new method of ranking based

on dCOG~wk

I

� �.

The proposed method for ranking the DMUs on the basis of ~wkI –

The algorithm of the new ranking method is as given below:Let J = {1, 2, . . . ,n} be the index set where n be the number of

DMUs; and let i = 1 and RKi¼ 1 be the initialization.

Step 1. Evaluate ~wkI for each DMU.

Step 2. Find the defuzzified value dCOG~wk

I

� �of ~wk

I for each DMU.

Step 3. If the value of dCOG~wk

I

� �2 ð0;1�for each DMU, STOP the

algorithm and evaluate the ranks of the DMUs in the con-ventional way starting from RKi

. Otherwise go to step 4.

Step 4. Let Ki ¼ j 2 J : dCOG~wj

I

� �¼ 1

n obe the set of all those indi-

ces j’s of DMUj’s for which the defuzzified value of ~wjI is

equal to 1. If Ki – /, give the rank RKito each DMUj such

that j 2 Kiand repeat the algorithm with J = J � {J \ Ki},i = i + 1 and RKi

¼ RKiþ jKij. Otherwise STOP the algorithm

and give the ranks to the remaining DMUs in the conven-tional way starting from RKi

.

Remark 2. The DMUj is known as the fuzzy input mix-efficient ifj 2 K1. This implies that such a DMU has the most efficient combi-nation of the fuzzy inputs, even though it may be technically fuzzyinefficient.

Using the above algorithm the ranking of the DMUs on the basisof ~wk

I is shown in Table 8.Table 8 shows that the FIME determines the ranking of the

DMUs at an order of R(1) � R(2) > R(4) � R(11) >R(3) > R(7) > R(12) > R(9) > R(5) > R(8) > R(10) > R(6) where R(k) =rank of DMUk on the basis of ~wk

I : Here K1 = {1,2}, so DMU1 andDMU2 are fuzzy input mix-efficient and have the most efficientcombination of fuzzy inputs.

9. An application to banking sector

Consider a performance assessment problem of the State Bankof Patiala (SBOP) in India where the performance of SBOP is evalu-ated in seventeen districts of the Punjab State in terms of three in-puts and two outputs. The DMUs in our study are the districts ofthe Punjab state. The inputs used in our study are (i) Labour, (ii)Fixed assets and (iii) Total expenses. Labour is the number ofemployees working in the bank in different districts. Fixed assetsare the physical capital of the bank in different districts. Total ex-penses include interest and non-interest expenses. Out of these

Page 12: A concept of fuzzy input mix-efficiency in fuzzy DEA and its application in banking sector

Table 13The values of ~hk

I , ~qkI and ~wk

I approximated as TFNs.

DMU ~hkI

~qkI

~wkI

1 (0.53573,0.54117,0.55148) (0.42473,0.43866,0.45429) (0.77016,0.81058,0.84798)2 (0.56185,0.60104,0.64568) (0.54032,0.55436,0.57195) (0.83682,0.92234,1.01798)3 (0.69536,0.69536,0.72608) (0.58721,0.61815,0.70215) (0.80874,0.88896,1.00977)4 (0.71062,0.71751,0.73017) (0.64205,0.66123,0.68926) (0.87932,0.92156,0.96994)5 (0.87536,0.90527,1.00000) (0.84077,0.87763,1.00000) (0.84077,0.96947,1.14239)6 (0.45776,0.45776,0.46589) (0.39166,0.40731,0.43168) (0.84067,0.88979,0.94303)7 (0.41975,0.43510,0.44512) (0.33381,0.34566,0.36230) (0.74993,0.79444,0.86313)8 (0.80353,0.81321,0.82771) (0.61186,0.62286,0.63957) (0.73922,0.76593,0.79595)9 (0.71332,0.73183,0.76088) (0.56580,0.59166,0.63112) (0.74361,0.80847,0.88476)

10 (1.00000,1.00000,1.00000) (1.00000,1.00000,1.00000) (1.00000,1.00000,1.00000)11 (0.79531,0.79531,0.94688) (0.68302,0.73868,0.82069) (0.72134,0.92880,1.03191)12 (0.58747,0.62628,0.72772) (0.52307,0.57830,0.65530) (0.71878,0.92339,1.11546)13 (0.88663,0.88663,0.92523) (0.72058,0.76883,0.83897) (0.77881,0.86714,0.94625)14 (0.23795,0.23795,0.23983) (0.17080,0.18137,0.20749) (0.71217,0.76222,0.87199)15 (0.42079,0.42337,0.42606) (0.28229,0.28586,0.28953) (0.66256,0.67520,0.68806)16 (0.70909,0.70909,0.73075) (0.58170,0.59489,0.61879) (0.79603,0.83895,0.87265)17 (0.64503,0.65172,0.66111) (0.58214,0.59326,0.60630) (0.88055,0.91030,0.93996)

Table 14The defuzzified values of ~hk

I , ~qkI and ~wk

I by using the COG method and ranking of the DMUs.

DMU dCOG~hk

I

� �Rank dCOG ~qk

I

� �Rank dCOG

~wkI

� �Rank dIE

COG~wk

I

� �(In %age)

1 0.439227 13 0.542794 13 0.809572 13 19.04282 0.555538 12 0.602858 12 0.925713 3 7.42873 0.635836 6 0.705695 9 0.902490 7 9.7514 0.664183 5 0.719430 7 0.923605 4 7.65 0.906133 2 0.926875 2 0.984210 2 1.5796 0.410216 14 0.460304 14 0.891161 9 10.88397 0.347259 15 0.433321 15 0.802501 14 19.74998 0.624767 7 0.814814 5 0.767031 16 23.29699 0.596193 9 0.735341 6 0.812280 12 18.772

10 1.000000 1 1.000000 1 1.000000 1 011 0.747465 4 0.845960 4 0.894016 8 10.598412 0.585558 11 0.647156 11 0.919210 5 8.07913 0.776125 3 0.899413 3 0.864067 10 13.593314 0.186557 17 0.238312 17 0.782126 15 21.787415 0.285902 16 0.423382 16 0.675266 17 32.473416 0.598458 8 0.716589 8 0.835879 11 16.412117 0.593904 10 0.652627 10 0.910266 6 8.9734

Note: Input mix-inefficiency in kth DMU (In %age form) = dIECOG

~wkI

� �¼ 1� dCOG

~wkI

� �� �� 100.

1448 J. Puri, S.P. Yadav / Expert Systems with Applications 40 (2013) 1437–1450

three inputs, two inputs, namely, labour and total expenses are ta-ken as fuzzy inputs and are represented as TFNs. The outputs usedin our study are (i) Interest income and (ii) Other income. Interestincome is the income earned by the bank in different districts fromadvances and investments. Other income accounts for the incomefrom off-balance sheet items such as commission, exchange andbrokerage etc. The objective of our study is to measure the FIMEof SBOP in various districts for the period 2010–2011 and to knowwhich districts are the best performers in terms of FIME and whichone are the worst performers. In the period 2010–2011, the totalnumbers of districts in the Punjab state were twenty. We areexcluding three districts in our study because these were estab-lished in the same year 2010. The data is shown in Table 9.

To ensure the validity of the FDEA model specification, the cor-relation coefficients between the fuzzy inputs and fuzzy outputsare calculated. The expected interval values and the correspondingexpected values of the fuzzy correlation coefficients between thesets of fuzzy data (given in Table 9) are shown in the matrix formin Table 10.

Table 10 indicates that the left and right bounds of each ex-pected interval and the corresponding expected value are positive.It means that all the inter-correlations between the inputs and out-puts are positive. Thus, the inclusion of the input and output data isjustified and the FDEA model which is taken in our present study isconsistent.

The a-cuts ~hkI

� �a

and ~qkI

� �a of ~hk

I and ~qkI of the seventeen DMUs

are evaluated by using models 14a, 14b, 15a and 15b at differentvalues of a and are shown in Tables 11 and 12, respectively. Thegraphical representations of the fuzzy efficiencies ~hk

I and ~qkI for

the kth DMU which are obtained by using a-cuts ~hkI

� �a

and ~qkI

� �a

are shown in Figs. 2a,2b, 2c and 2d. It can be seen from the figuresthat the shape of the membership functions of ~hk

I and ~qkI are

approximated as triangular membership functions. Further, Table13 presents the values of ~hk

I , ~qkI and ~wk

I approximated as TFNs.Let dCOG

~hkI

� �; dCOG ~qk

I

� �and dCOG

~wkI

� �be the defuzzified values

of ~hkI , ~qk

I and ~wkI respectively which are obtained by using COG

method of defuzzification and are shown in Table 14. The ranking

of DMUs on the basis of dCOG~hk

I

� �and dCOG ~qk

I

� �is done in a conven-

tional way. However, the ranking of DMUs on the basis of dCOG~wk

I

� �is done by applying the proposed method of ranking given in Sec-tion 8. In this case the algorithm terminates at the Step 3 since the

value of dCOG~wk

I

� �2 ð0;1� for each DMU.

Table 14 indicates that DMU10, i.e., Ludhiana district is the mostefficient district in terms of ~hk

I , ~qkI and ~wk

I . The order of performanceas well as level of inefficiency (in brackets) of the districts in termsof ~wk

I is given by DMU10 > DMU5 > DMU2 > DMU4 > DMU12 >

Page 13: A concept of fuzzy input mix-efficiency in fuzzy DEA and its application in banking sector

J. Puri, S.P. Yadav / Expert Systems with Applications 40 (2013) 1437–1450 1449

DMU17 > DMU3 > DMU11 > DMU6 > DMU13 > DMU16 > DMU9 > DMU1 >DMU7 > DMU14 > DMU8 > DMU15 i.e. Ludhiana (0%) > Ferozepur(1.579%) > Bathinda (7.4287%) > Fatehgarh Sahib (7.6%) > Moga(8.079%) > Sangrur (8.9734%) > Faridkot (9.751%) > Mansa(10.5984%) > Gurdaspur (10.8839%) > Muktsar (13.5933%) > Ropar(16.4121%) > Kapurthala (18.772%) > Amritsar (19.0428%) >Hoshiarpur (19.7499%) > Nawan Shahar (21.7874%) > Jalandhar(23.2969%) > Patiala (32.4734%). Thus the worst performer in termsof FIME is Patiala district with dCOGð~wk

I Þ ¼ 0:675266. This means thatPatiala district is not utilizing its inputs (fuzzy as well as non-fuzzy)efficiently. Table 14 also indicates that Bathinda district is inefficient

in terms of ~hkI and ~qk

I with dCOG~hk

I

� �and dCOG ~qk

I

� �equal to 0.555538

and 0.602858 respectively. However, Bathinda district is a good per-

former in terms of ~wkI with dCOG

~wkI

� �¼ 0:925713. Table 14 also re-

veals that the highest and lowest levels of fuzzy mix-inefficiencyhave been seen for Patiala (32.4734%) and Ferozepur (1.579%)respectively.

10. Conclusions

In view of the fact that precise input and output data are not al-ways available in real world applications, we have developed, inthis paper FIME model. For measuring the FIME, we have proposedthe FCCRI and FSBMI. These two FDEA models have been formu-lated as linear programming models using a-cut approach for easeof solution and implementation. To ensure the validity of the FDEAmodel specification, we have proposed a fuzzy correlation coeffi-cient method using expected value approach which calculatesthe expected interval and expected value of fuzzy correlation coef-ficient between fuzzy inputs and fuzzy outputs. If positive inter-correlations are found, the inclusion of the fuzzy inputs and fuzzyoutputs is justified. Further, a new ranking method based ondefuzzification approach has been developed for comparing andranking DMUs in terms of FIME, which provides not only a fullranking but also the information that to what degree a FIME is big-ger than another one. All of the proposed approaches have been ap-plied to evaluate the performances of seventeen districts of SBOPin the Punjab State of India in terms of FIME. It is shown thatLudhiana district is the most efficient district in terms of ~hk

I , ~qkI

and ~wkI . The highest and lowest levels of fuzzy mix-inefficiency

have been seen for Patiala (32.4734 %) and Ferozepur (1.579 %)respectively. The input mix-inefficiency represents the degree towhich the input mix should change to become fully efficient.According to the findings of our study, all the fuzzy input mix-inef-ficient districts are suggested to decrease their input mix in orderto become fully efficient.

Acknowledgments

The first author is thankful to the University Grants Commission(UGC), Government of India for financial assistance. The authorsare thankful to Ms. M. A. Malini, Manager (Systems), State Bankof Patiala, IT Services Deptt, H. O., Patiala, India for providing therelevant data of the bank.

References

Asbullah, M. A. (2010). A new approach to estimate the mix efficiency in dataenvelopment analysis. Applied Mathematical Sciences, 4(43), 2135–2143.

Avkiran, N. K. (2006). Productivity analysis in the services sector with dataenvelopment analysis (3rd ed.). Brisbane: University of Queensland BusinessSchool, The University of Queensland.

Avkiran, N. K., Tone, K., & Tsutsui, M. (2008). Bridging radial and non-radialmeasures of efficiency in DEA. Annals of Operations Research, 164, 127–138.

Banker, R. D., Charnes, A., & Cooper, W. W. (1984). Some models for estimatingtechnical and scale inefficiencies in DEA. Management Science, 30(9),1078–1092.

Bansal, A. (2010). Some non linear arithmetic operations on triangular fuzzynumbers (m,a,b). Advances in Fuzzy Mathematics, 5(2), 147–156.

Charnes, A., Cooper, W. W., & Rhodes, E. (1978). Measuring the efficiency of decisionmaking units. European Journal of Operational Research, 2, 429–444.

Charnes, A., Cooper, W. W., Seiford, L., & Stutz, J. (1982). A multiplicative model forefficiency analysis. Socio-Economic Planning Sciences, 6, 223–224.

Chen, S. M. (1994). Fuzzy system reliability analysis using fuzzy number arithmeticoperations. Fuzzy Sets and Systems, 66, 31–38.

Cooper, W. W., Seiford, L. M., & Tone, K. (2007). Data envelopment analysis: Acomprehensive text with models, applications, references and DEA-solver software(2nd ed.). New York: Springer.

Dia, M. (2004). A model of fuzzy data envelopment analysis. INFOR, 42, 267–279.Entani, T., Maeda, Y., & Tanaka, H. (2002). Dual models of interval DEA and its

extension to interval data. European Journal of Operational Research, 136,32–45.

Guh, Y. Y. (2001). Data envelopment analysis in fuzzy environment. Information andManagement Sciences, 12(2), 51–65.

Guo, P., & Tanaka, H. (2001). Fuzzy DEA: A perceptual evaluation method. Fuzzy Setsand Systems, 119, 149–160.

Hatami-Marbini, A., Saati, S., & Makui, A. (2010). Ideal and anti-ideal decisionmaking units: A fuzzy DEA approach. Journal of Industrial EngineeringInternational, 6(10), 31–41.

Hatami-Marbini, A., Saati, S., & Tavana, M. (2010). An ideal-seeking fuzzy dataenvelopment analysis framework. Applied Soft Computing, 10, 1062–1070.

Herrero, I., Pascoe, S., & Mardle, S. (2006). Mix efficiency in a multi-species fishery.Journal of Productivity Analysis, 25, 231–241.

Hsiao, B., Chern, C. C., Chiu, Y. H., & Chiu, C. R. (2011). Using fuzzy super-efficiencyslack-based measure data envelopment analysis to evaluate Taiwan,scommercial bank efficiency. Expert Systems with Applications, 38, 9147–9156.

Hung, W. L., & Wu, J. W. (2001). A note on the correlation of fuzzy numbers byexpected interval. International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, 9(4), 517–523.

Jahanshahloo, G. R., Soleimani-damaneh, M., & Nasrabadi, E. (2004). Measure ofefficiency in DEA with fuzzy input–output levels: A methodology for assessing,ranking and imposing of weights restrictions. Applied Mathematics andComputation, 156, 175–187.

Kao, C., & Liu, S. T. (2000a). Fuzzy efficiency measures in data envelopment analysis.Fuzzy Sets and Systems, 113, 427–437.

Kao, C., & Liu, S. T. (2000b). Data envelopment analysis with missing data: Anapplication to University libraries in Taiwan. Journal of the Operational ResearchSociety, 51, 897–905.

Kao, C., & Liu, S. T. (2003). A mathematical programming approach to fuzzyefficiency ranking. International Journal of Production Economics, 86, 45–154.

Kao, C., & Liu, S. T. (2005). Data envelopment analysis with imprecise data: Anapplication of Taiwan machinery firms. International Journal of Uncertainty,Fuzziness and Knowledge-Based Systems, 13(2), 225–240.

Kataria, N. (2010). A comparative study of the defuzzification methods in anapplication. The IUP Journal of Computer Sciences, IV(4), 48–54.

Lertworasirikul, S. (2001). Fuzzy data envelopment analysis for supply chain modelingand analysis, dissertation proposal in industrial engineering. North Carolina StateUniversity.

Lertworasirikul, S., Fang, S. C., Jeffrey, A., Joines, J. A., & Nuttle, H. L. W. (2003). Fuzzydata envelopment analysis (DEA): A possibility approach. Fuzzy Sets andSystems, 139(2), 379–394.

Liu, S. T. (2008). A fuzzy DEA/AR approach to the selection of flexible manufacturingsystems. Computers & Industrial Engineering, 54(1), 66–76.

Liu, S. T., & Chuang, M. (2009). Fuzzy efficiency measures in fuzzy DEA/AR withapplication to university libraries. Expert Systems with Applications, 36(2P1),1105–1113.

Majid Zerafat Angiz, L., Emrouznejad, A., & Mustafa, A. (2012). Fuzzy dataenvelopment analysis: A discrete approach. Expert Systems with Applications,39, 2263–2269.

Meada, Y., Entani, T., & Tanaka, H. (1998). Fuzzy DEA with interval efficiency.Proceedings of 6th european congress on intelligent techniques and soft computing,EUFIT ’98 (Vol. 2, pp. 1067–1071). Aachen, Germany: Verlag Mainz.

Nasseri, H. (2008). Fuzzy numbers: Positive and nonnegative. Internationalmathematical Forum, 3(36), 1777–1780.

Petersen, N. C. (1990). Data envelopment analysis on a relaxed set of assumptions.Management Science, 36(3), 305–313.

Saati, S., & Memariani, A. (2009). SBM model with fuzzy input-output levels in DEA.Australian Journal of Basic and Applied Sciences, 3(2), 352–357.

Saati Mohtadi, S., Memariani, S. A., & Jahanshahloo, G. R. (2002). Efficiency analysisand ranking of DMUs with fuzzy data. Fuzzy Optimization and Decision Making, 1,255–267.

Sengupta, J. K. (1992). A fuzzy systems approach in data envelopment analysis.Computers & Mathematics with Applications, 24, 259–266.

Tone, K. (1998). On mix efficiency in DEA. The Operations Research Society of Japan,14–15. Available from: <http://ci.nii.ac.jp/naid/110003478327/en>.

Tone, K. (2001). A slacks-based measure of efficiency in data envelopment analysis.European Journal of Operational Research, 130, 498–509.

Tsai, H. C., Chen, C. M., & Tzeng, G. H. (2006). The comparative productivityefficiency for global telecoms. International Journal of Production Economics, 103,509–526.

Page 14: A concept of fuzzy input mix-efficiency in fuzzy DEA and its application in banking sector

1450 J. Puri, S.P. Yadav / Expert Systems with Applications 40 (2013) 1437–1450

Wang, Y. M., & Chin, K. S. (2011). Fuzzy data envelopment analysis: A fuzzyexpected value approach. Expert Systems with Applications, 38,11678–11685.

Wang, Y. M., Greatbanks, R., & Yang, J. B. (2005). Interval efficiency assessment usingdata envelopment analysis. Fuzzy Sets and Systems, 153(3), 347–370.

Wang, Y. M., Luo, Y., & Liang, L. (2009). Fuzzy data envelopment analysis based uponfuzzy arithmetic with an application to performance assessment ofmanufacturing enterprises. Expert Systems with Applications, 36(3), 5205–5211.

Zadeh, L. A. (1975). The concept of linguistic variable and its application toapproximate reasoning-I. Information Science, 8, 199–249.

Zhou, Z., Lui, S., Ma, C., Liu, D., & Liu, W. (2012). Fuzzy data envelopment analysismodels with assurance regions: A note. Expert Systems with Applications, 39,2227–2231.


Recommended