You are on page 1of 27

Soft Computing

Pitfalls and solutions in inverse models of data envelopment analysis with network
structures
--Manuscript Draft--

Manuscript Number:

Full Title: Pitfalls and solutions in inverse models of data envelopment analysis with network
structures

Article Type: Original Research

Keywords: Inverse models of data envelopment analysis, Series network, Estimation of inputs and
outputs, Efficiency.

Corresponding Author: Morteza Yazdani


Universidad Internacional de Valencia
SPAIN

Corresponding Author Secondary


Information:

Corresponding Author's Institution: Universidad Internacional de Valencia

First Author: Zohreh Moghaddas

First Author Secondary Information:

Order of Authors: Zohreh Moghaddas

Farhad Hosseinzadeh Lotfi

Morteza Yazdani

Funding Information:

Abstract: Today, data envelopment analysis models with network structures are widely used to
evaluate the performance of production systems and activities in various fields. the
relationships between the internal stages of the network provides more information
about the performance of each stage as well as their effects on the performance of the
entire network to the managers and decision-makers. The inverse data envelopment
analysis model is introduced to estimate useful information to system decision-makers,
about analyzing the sensitivity of system inputs or outputs as long as the efficiency
score is kept unchanged or improved according to the managers preferences.
Managers can apply their important preferences and policies on resources, including
input and output when analyzing production, resource allocation process, increasing
resource efficiency, etc. on the system to use the results for future decisions. In this
article, we will discuss the problems of infeasibility that can occur in theory and
application for the inverse model of data envelopment analysis with network structure. 
After introducing these problems, an innovative idea is presented to prevent these
shortcomings.  Then, various problems are supposed, in terms of theory and
applications, and are solved with case studies.

Section/Category: Mathematical methods in data science

Powered by Editorial Manager® and ProduXion Manager® from Aries Systems Corporation
Click here to access/download;Manuscript;final 2022.9.14.docx

Click here to view linked References


1
2
3
4
5 Pitfalls and solutions in inverse models of data envelopment analysis with
6 network structures
7
8
9
10
11
12
13
14
15
16 Zohreh Moghaddas1, Farhad Hosseinzadeh Lotfi2, Morteza Yazdani3*
17
1
18 Department of Mathematics and Statistics, Qazvin Branch, Islamic Azad University, Qazvin,
19 Iran(zmoghaddas@qiau.ac.ir)
20 2
21 Department of Mathematics, Science and Research Branch, Islamic Azad University, Tehran,
22 Iran(Farhad@Hosseinzadeh.ir)
3*
23 Universidad Internacional de Valencia, Spain (*corresponding: my.173021@gmail.com)
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
1
2
3
4
5 Pitfalls and solutions in inverse models of data envelopment analysis with
6 network structures
7
8
9
10
11
12 Abstract:
13 Today, data envelopment analysis models with network structures are widely used to evaluate
14
15 the performance of production systems and activities in various fields. the relationships between
16 the internal stages of the network provides more information about the performance of each
17
18
stage as well as their effects on the performance of the entire network to the managers and
19 decision-makers. The inverse data envelopment analysis model is introduced to estimate useful
20 information to system decision-makers, about analyzing the sensitivity of system inputs or
21
22 outputs as long as the efficiency score is kept unchanged or improved according to the managers
23 preferences. Managers can apply their important preferences and policies on resources,
24
25
including input and output when analyzing production, resource allocation process, increasing
26 resource efficiency, etc. on the system to use the results for future decisions. In this article, we
27 will discuss the problems of infeasibility that can occur in theory and application for the inverse
28
29 model of data envelopment analysis with network structure. After introducing these problems,
30 an innovative idea is presented to prevent these shortcomings. Then, various problems are
31
32
supposed, in terms of theory and applications, and are solved with case studies.
33 Keywords: Inverse models of data envelopment analysis, Series network, Estimation of inputs and outputs,
34 Efficiency.
35
36
37 1.Introduction:
38 Data Envelopment Analysis (DEA) is a non-parametric method for determining performance
39
40
analysis of production systems or activities organization Data envelopment analysis, which is
41 based on mathematical programming, is used to evaluate the relative efficiency of Decision-
42 Making Units (DMUs). It should be noted that the relative term is due to the fact that the obtained
43
44 efficiency score is the result of comparing units with each other. An important feature of this
45 technique is to consider multiple inputs and outputs in activities, which are usually
46
47
immeasurable. The basis of data envelopment analysis method for performance and efficiency
48 measurement is based on these two basic assumptions that if unit A can produce more output
49 than unit B but with the same amount of input, unit A is more efficient than unit B. Another is
50
51 that, if unit A can produce a certain amount of output with a certain amount of input, it is
52 expected that other similar units can produce the same amount of output with the same amount
53
54
of input. Similarly, if unit B with a certain amount of input be able to produce a certain amount
55 of output again, it is expected that other units will be able to do the same. The initial idea for the
56 Inverse Data Envelopment Analysis (IDEA) model was introduced by Zhang and Cui (1999). In this
57
58 way, they examined, if the output values (input) are changed, what should be the input (output)
59 values, provided that the value of efficiency is maintained or improved. Wei et al. (2000), also
60
61
62
63
64
65
1
2
3
4 introduced an inverse model of data envelopment analysis. They posed a fundamental question;
5
6 if we increase the input to a certain extent and want to keep the efficiency level constant, how
7 much output should we produce. Then the model in response to this question was called the
8
9 inverse model of data envelopment analysis. In general, depending on which parameters change
10 and which parameters are estimated, the inverse envelopment analysis model is introduced.
11 Resource allocation model and the investment analysis model are applications of inverse DEA
12
13 models. In the resource allocation model, the allocation of output resources is increased,
14 provided that the amount of efficiency is maintained or improved, the model seeks the minimum
15
16 amount of system input, Wei et al. (2000). In the investment analysis model, the input is reduced,
17 provided that the value of efficiency is maintained or improved, the model follows the maximum
18 value of the system output, Wei et al. (2000). Inverse data envelopment analysis models allow
19
20 Decision Makers (DMs), managers to apply their important preferences and policies to resources,
21 both input and output, when analyzing production, the resource allocation process, and
22
23 increasing resource efficiency. Inverse data envelopment analysis models used the concepts of
24 multi-objective linear programming, weighted sum, and single-objective DEA linear programming
25 to solve the problem. Another important issue that is dealt with is the extra input that is discussed
26
27 in inverse DEA modelling by Jahanshahloo et al.(2004).
28 After introducing the initial inverse models of data envelopment analysis, the researchers
29
introduced more advanced and modified models in the subject literature. Inverse data
30
31 envelopment analysis models can be introduced in constant and variable returns to scale form of
32 the technology. On basis of the paper introduced by Lertworasirikul et al. (2011), and Ghiyasi
33 (2015), provided an article that discussed about some difficulties may exist in use of Multiple
34
35 Objective Linear Programming (MOLP). Then he corrects the use of MOLP in a proposed inverse
36 model with variable returns to scale for of the technology. Moreover, Chen and Wang (2021),
37 also discussed about the issue of infeasibility of the inverse DEA method VRS technology. Adimi
38
39
et al. (2021), proposed a DEA model that provides the conditions for stability or improvement of
40 the returns to scale for the DMU under assessment.
41 One of the applications of inverse models of data envelopment analysis is in resource allocation
42 issues. In the meantime, the issue of merging decision-making units was raised, and many articles
43
44 are provided, Amin et al. (2017), Amin et al. (2019), and Lin et al. (2020). Ghobadi (2021),
45 mentioned that for the aim of merging units with interval data when having the imprecise or
46 ambiguous data, the inverse DEA concept and linear programming models could be employed to
47
48 estimate the inherited inputs/outputs of the merged units, and identify the least and most
49 achievable efficiency targets from the merged unit. Zeinodin and Ghobadi (2020), provided
50 researches that deals with merging of decision-making units under the inter-temporal
51
dependence assumption.
52
53 One of the features of the inverse model of data envelopment analysis is considering the
54 assumptions that can be added to the model with the opinion of the system decision makers or
55 managers. In the first models, these preconditions were to keep the amount of technical
56
57 efficiency constant. But, as more advanced models were introduced, in addition to maintaining
58 or improving the amount of technical efficiency, researchers discussed theoretically about
59 maintaining or improving cost, revenue, and scale efficiency scores and proposed different
60
61
62
63
64
65
1
2
3
4 models. Ghiyasi (2017), provided model deals with the inverse DEA problem when price
5
6 information is available. Its proposed models guarantee both technical efficiency and cost
7 efficiency to be kept fixed while input is estimated with a perturbed output. Ghiyasi and
8 Dehnokhalaji (2021) introduced an inverse model for resource allocation having price
9
10 information. Ghiyasi and Zhu (2020) presented an inverse semi-oriented radial data envelopment
11 analysis measure for dealing with negative data. Daryani et al. (2021), investigated considered
12 inverse DEA models and considered allocative efficiency remained fixed. Soleimani-
13
14
Chamkhorami et al. (2020), assumed considered cost or revenue efficiency scores unchanged
15 when data are altered is important for decision makers. Amin and Boamah (2020), introduced a
16 new inverse data envelopment analysis according to the cost efficiency model in order to
17 estimate potential gains from mergers.
18
19 In addition to introducing advanced and modified models of inverse data envelopment analysis,
20 the researchers also examined various data types in these models. Yousefi et al. (2019),
21 presented an inverse input and output-oriented RDM model in presence of negative data. Çakır
22
23 (2017), studied a resource allocation problem under a fuzzy environment and proposed an
24 integrated Shannon’s entropy–inverse data envelopment analysis. Ghobadi et al. (2014), applied
25
26 dynamic IDEA while fuzzy data is being considered. Hassanzadeh et al. (2018), demonstrated two
27 new inverse DEA models for resource allocation and investment analysis problems with negative
28 data.
29
30 In classical models of data envelopment analysis, efficiency score can be evaluated based on
31 radial or non-radial changes, Tone (2001). Thus, this is an important issue that is also considered
32 in inverse data envelopment analysis models by, Zhang and Cui (2020), Hosseininia and Saen
33
34 (2020), and Hu et al. (2020).
35 Undesirable factors in inverse models of data envelopment analysis is considered from various
36 aspects. Ghiyasi (2017), developed a theoretical background of the inverse DEA with pollution
37
38 generating technology that is capable of dealing with undesirable outputs. Chen et al. (2017),
39 provided a new inverse data envelopment analysis method with undesirable outputs to make
40 several scientifically based investment schemes. Eyni et al. (2017), provided a paper deal with
41
the inverse data envelopment analysis (DEA) with the preference of cone constraints in a way
42
43 that in the decision-making units, the undesirable inputs and outputs exist simultaneously. Lin et
44 al. (2019), provided an inverse data envelopment analysis model for measuring container ports’
45 efficiency and analyzing their resource consumption while having undesirable outputs. Wegener
46
47 and Amin (2019), introduced a new inverse data envelopment analysis model for optimizing GHG
48 emissions is developed with an application in the oil and gas industry. Emrouznejad et al. (2019),
49 provided a new inverse data envelopment analysis model to allocate CO2 emission. Chen et al.
50
51
(2021), considered undesirable outputs in IDEA models as they believed this is necessary for
52 ensuring the realization of safety objective under the current technical level. Zhang et al (2021),
53 developed a three-stage empirical system to identify the CO2 emissions allocation scheme while
54 calculating the overall efficiency score. Zhang et al. (2020), developed a new approach based on
55
56 the concept of utilizing both DEA and inverted-DEA to enhance the discrimination power of DMUs
57 with undesirable outputs.
58 More improvements and modifications in IDEA models have been performed. Zhang and Cui
59
60 (2016), analyzed all the possible changing relationships that need to be solved by the inverse DEA
61
62
63
64
65
1
2
3
4 method and developed different models for both the output and input orientations,
5
6 accomplishing the extension and integration of the inverse DEA model. Orisaremi et al. (2021),
7 tried to improve sustainability measures across the petroleum industry. Thus, they were intended
8 to estimate accurate potential reductions in gas flaring using inverse DEA models which uses
9
10 mathematical optimization problems. Lim (2016), Inverse data envelopment analysis (DEA) is a
11 reversed optimization problem that can serve as a useful planning tool for managerial decisions.
12 In this paper an inverse DEA problem considering expected changes of the production frontier in
13
14
the future by integrating the inverse optimization problem with a time series application of DEA
15 so that it can be an ex-ante decision support tool for the new product target setting practices is
16 presented. Soleimani-Chamkhorami et al (2020), utilized IDEA models for ranking system for
17 ordering extreme efficient units. Sayar et al. (2021), proposed inverse DEA models to help
18
19 decision-makers finding the required value of each input and each output's income share to meet
20 the income or budget constraint. Le et al. (2021) also presented a model about inverse Frontier-
21 based Benchmarking for Investigating the Efficiency and Achieving the Targets.
22
23
The purpose of this study is to investigate the problems and obstacles in inverse models of data
24 envelopment analysis with network structure. As mentioned above, in the literature of data
25 envelopment analysis, inverse models of data envelopment analysis have recently been
26
27 considered in theory and application in various fields. Because these models allow managers and
28 decision makers to assess their preferences and perform sensitivity analysis for the system under
29
30
their control, and thus can better act for future decision making.
31 Modhej et al. (2017) presented a paper to integrate inverse DEA and artificial neural network for
32 a large dataset with multiple DMUs for determining the best possible values of inputs for a large
33
34
number of DMUs when their output levels are changed and their efficiency values remain
35 unchanged. Kalantary et al. (2018) proposed a network dynamic range adjusted measure (RAM)
36 model. Then, improved it to an inverse version of network dynamic RAM model. Also changing
37 both inputs and outputs of decision-making units so that current efficiency scores of DMUs
38
39 remain unchanged are investigated. Kalantary and Farzipoor (2019) introduced a network
40 dynamic DEA model to assess sustainability of supply chains in multiple periods while designing
41 an inverse network DEA model in dynamic context. Farzipoor et al. (2020) presented a new INDEA
42
43
model to allocate resources while the inputs are not increased and the efficiency scores of all
44 DMUs remain constant. Khiavi and Eskandari (2021) presented a study that constructed a model
45 of Inverse Network Data Envelopment Analysis, to introduce the relative and interval scores of
46 the bullwhip effect magnitude, when a series of uncertain demands are made in a specific time
47
48 interval. Many studies on the modeling of inverse network data envelopment analysis have not
49 been done so far. As mentioned in the above articles, few articles have investigated this issue.
50 But what is more important is that no one has done a review of the model mathematically and
51
52 the shortcomings that may happen inside the network model of inverse data envelopment
53 analysis.
54
55
56
Since network models represent many activities in different fields in real world applications, so
57 inverse data envelopment analysis models with network structure are also of great importance.
58 However, since problems in the use of these models may occur, this article for the first time
59
60 introduces these difficulties and provides a solution to prevent these problems from occurring.
61
62
63
64
65
1
2
3
4 The contributions of the provided study are:
5
6
7  For the first time, the problems that may occur in the inverse model of data envelopment
8
9 analysis are introduced and discussed in details.
10  Mentioned problems are illustrated in the inverse model of data envelopment analysis
11
12
with numerical examples.
13  Solutions are provided to solve the mentioned problems.
14  Theorems are provided to discuss about the introduced models in detail.
15
16  Moreover, some protocols are provided to avoid these pitfalls.
17
18
19 The research impacts of the provided study which can aid experts to improve their decision for
20 enhancing a more accurate decision process are, first, the inverse model of data envelopment
21
analysis has been developed for a network system with intermediate relation. Secondly, by the
22
23 problems that appear in reverse network models, decision makers can overcome the issue of
24 infeasibility of the outcomes. Third, by this model, experts and decision makers would be able to
25
26 confirm the model utility and ultimately this enables them to configure multiple strategies for
27 future decision making based on obtained results.
28
29
30 The paper unfolds as follows. In section 2, pitfalls and difficulties in inverse DEA models are
31 discussed in details along with numerical examples. In section 3, solutions for overcoming these
32
33 difficulties are provided with some theorems. Section 4 provides a numerical example and
34 section 5 concludes the paper.
35
36
37 2. Pitfalls in Inverse DEA Network Models
38 Many methods have been introduced to measure the efficiency and performance of decision-
39
40
making units. But in comparison with the introduced models, the models introduced in the data
41 envelopment analysis technique are among the most efficient models for efficiency
42 measurement and data analysis. Data envelopment analysis technique is a method based on
43
44 mathematical programming that does not require any presuppositions and statistical information
45 about the data. In this technique, the efficient frontier is estimated according to the observed
46
47
units and the efficiency of the units is inversely related to the distance from the efficient frontier.
48 Note that despite the possibility of data turbulations or data changes over time, the efficiency
49 threshold may also change, as a result of which efficiency values will also change. IDEA models
50
51 are from a series of applied models and are of interest to researchers and decision makers.
52 Because these models allow organizations to be as efficient as before by making new decisions
53
54
and defining new strategies for their activities, and are able to examine the optimal amount of
55 consumption (production) input (output) to improve efficiency. In general, the primary purpose
56 of DEA models is to evaluate efficiency, and the primary purpose of IDEA models is to determine
57
58 the best possible input (output) corresponds to the output (input), under the conditions specified
59
60
61
62
63
64
65
1
2
3
4 by the manager. These conditions are mainly to keep the amount of efficiency constant or even
5
6 improve it.
7 In this study, we will examine this issue on decision-making units that have a network structure,
8
9 and we will point out the problems that may occur in this issue and protocols to solve them from
10 theoretical and practical aspects. In data envelopment analysis, in order to better evaluate the
11 problems that had a network structure, the relationship between the internal stages of the
12
13 network and the linking between them were also considered, Li et al (2018). It has been
14 suggested that in a network system, different stages of the network affect the performance of
15
16 the entire network. Therefore, the amount of efficiency obtained for the whole network (chain)
17 is the result of overlapping the efficiency of each component of the network. In evaluating a
18 network using IDEA models, different stages of the network must also be considered properly.
19
20 Therefore, it is important to properly consider the stages in the network, and to know the
21 problems that can occur for the IDEA model in theory and practical examples.
22
23 Consider Figure 1, a series network with two series-connected stages. Stage 1 uses the vector 𝑥
24 with m elements as its inputs and produce vector z with f elements as its output. The vector z is
25 the intermediate product which is considered also as the input of stage 2. The final output of the
26
27 network is vector y with s elements which finally, leaves the chain. Note that in black-box
28 evaluation, DEA models only consider the input (x) and the output (y) without considering the
29
30
intermediate product (z). As Wang et al. (1997) mentioned the necessity of considering
31 intermediate products in network systems.
32
33
34
35
36
37
38
39
40
41
42
43 FIGURE 2.1: A two-stage series network
44
45
46 Now consider optimization model (2.1) corresponds to Figure 1. Model (2.1) evaluates the
47 efficiency score of the entire chain in output orientation. Consider 𝜑 as the maximum radial
48
49 increase in output y, 𝜆𝑗1 and 𝜆𝑗2 as the intensifier variable for stage 1 and 2 related to each 𝐷𝑀𝑈𝑗 ,
50 for all j.
51
52
53 𝑚𝑎𝑥 𝜑
54
55
s.t.
56 ∑𝑛𝑗=1 𝜆𝑗1 𝑥𝑖𝑗 ≤ 𝑥𝑖𝑜 , 𝑖 = 1, … , 𝑚,
57
58 ∑𝑛𝑗=1 𝜆𝑗1 𝑧𝑓𝑗 ≥ 𝑧𝑓𝑜 , 𝑓 = 1, … , 𝑘, (2.1)
59 ∑𝑛𝑗=1 𝜆𝑗2 𝑧𝑓𝑗 ≤ 𝑧𝑓𝑜 , 𝑓 = 1, … , 𝑘,
60
61
62
63
64
65
1
2
3
4 ∑𝑛𝑗=1 𝜆𝑗2 𝑦𝑟𝑗 ≤ 𝜑𝑦𝑟𝑜 , 𝑟 = 1, … , 𝑠,
5
6 𝜆𝑗1 ≥ 0 , 𝜆𝑗2 ≥ 0, 𝑗 = 1, … , 𝑛.
7
8
9 𝜑 ∗ is the optimal objective function value of model (2.1) that represents the relative efficiency
10
11
score of 𝐷𝑀𝑈𝑜 under evaluation.
12 Now if the manager or decision maker increases 𝑦 to 𝑦 𝑁 , and wants to estimate the minimum
13
14
amount of system input so that technical performance remains constant, or improves according to
15 the manager (depending on the possible interval). In this case, we recommend model (2.2). This
16 model is an IDEA model with network structure, which is formulated as follows.
17
18 According to the increased output Wei et al. (2000) provided following models. The aim of model
19 (2.2) is to find the minimum increase in each element of the input vector x, 𝜎⃗ =
20
21
(𝜎1 , … , 𝜎𝑚 ), condition to keep the efficiency score unchanged or improved according to managers
22 decision.
23
24
25 min(𝜎𝑖 , … , 𝜎𝑚 )
26 s.t.
27
28
∑𝑛𝑗=1 𝜆𝑗1 𝑥𝑖𝑗 ≤ 𝜎𝑖 , 𝑖 = 1, … , 𝑚,
29 ∑𝑛𝑗=1 𝜆𝑗1 𝑧𝑓𝑗 ≥ 𝑧𝑓𝑜 , 𝑓 = 1, … , 𝑘, (2.2)
30
31 ∑𝑛𝑗=1 𝜆𝑗2 𝑧𝑓𝑗 ≤ 𝑧𝑓𝑜 , 𝑓 = 1, … , 𝑘,
32 ∑𝑛𝑗=1 𝜆𝑗2 𝑦𝑟𝑗 ≤ ( 𝜑 ∗ − 𝛼)𝑦𝑟𝑜
𝑁
, 𝑟 = 1, … , 𝑠,
33
34 𝜆𝑗1 ≥0 , 𝜆𝑗2 ≥ 0, 𝑗 = 1, … , 𝑛,
35
36 𝜎𝑖 ≥ 𝑥𝑖𝑜 , 𝑖 = 1, … , 𝑚.
37
38 In this study efficiency improvement is also considered with the term of 𝜑 ∗ − 𝛼.
39
40 Remark 1: If 𝜑 ∗ = 1, 𝐷𝑀𝑈𝑜 is efficienct. Thus, there is no room for efficiency improvement. Its
41 evident that 𝜑 ∗ − 𝛼 ≤ 𝜑 ∗ ≤ 1, therefore 𝜑 ∗ = 1 and 𝛼 = 0. In other cases, if 𝜑 ∗ > 1 then
42
43 there is room for efficiency improvement as 0 < 𝜑 ∗ − 1 and therefore 𝜑 ∗ − 𝛼 ≤ 𝜑 ∗ that 𝛼 ∈
44 [ 0, 𝜑 ∗ − 1 ].
45 In this way, if the manager wants, he can consider a value from the range [ 0, 𝜑 ∗ − 1 ] to increase
46
47 efficiency score. In this case, the value of the improved performance is 𝜑 ∗ − 𝛼.
48 If the manager wants to improve the efficiency of an inefficient unit, so that its efficiency is equal
49
50
to 1, it must set 𝛼 = 𝜑 ∗ − 1 which yield 𝜑 ∗ − 1 = 𝜑 ∗ − (𝜑 ∗ − 1 ) = 1. Otherwise, set 𝛼 = 0,
51 which means no performance improvement is assumed, 𝜑 ∗ − 𝛼 = 𝜑 ∗ . □
52
53
54 As model (2.2) is a multiple objective optimization problem. Single objective optimization
55 problem counterpart to model (2.2) is formulated as model (2.3) using goal programming
56
57 technique. According to the basics introduced by Wei et al. (2000) and model (2.2), following
58 model (2.3) is presented in IDEA literature. Thus, the weighted sum of variables 𝜎𝑖 introduce the
59
60 objective function.
61
62
63
64
65
1
2
3
4
5 𝑚
6
7 min ∑ 𝜎𝑖
8 𝑖=1
9 s.t.
10
11 ∑𝑛𝑗=1 𝜆𝑗1 𝑥𝑖𝑗 ≤ 𝜎𝑖 , 𝑖 = 1, … , 𝑚,
12
13
∑𝑛𝑗=1 𝜆𝑗1 𝑧𝑓𝑗 ≥ 𝑧𝑓𝑜 , 𝑓 = 1, … , 𝑘, (2.3)
14 ∑𝑛𝑗=1 𝜆𝑗2 𝑧𝑓𝑗 ≤ 𝑧𝑓𝑜 , 𝑓 = 1, … , 𝑘,
15 𝑛 2 ∗ 𝑁
16 ∑𝑗=1 𝜆𝑗 𝑦𝑟𝑗 ≤ (𝜑 − 𝛼)𝑦𝑟𝑜 , 𝑟 = 1, … , 𝑠,
17
𝜆𝑗1 ≥ 0 , 𝜆𝑗2 ≥ 0, 𝑗 = 1, … , 𝑛,
18
19 𝜎𝑖 ≥ 𝑥𝑖𝑜 , 𝑖 = 1, … , 𝑚.
20
21
22 Similar to model (2.2), in model (2.3) efficiency
23 Problems that can occur in solving model (2.3) are:
24
25
26 Pitfall 1 - Model (2.3) becomes infeasible when 𝑦 is changed to 𝑦 𝑁 .
27
28
29 In this case when 𝑦 is changed to 𝑦 𝑁 there may be situations where model (2.3) may not be able
30
31
to find a feasible solution for 𝜆2 that can handle both the third and fourth constraints. Therefore,
32 model (2.3) may be infeasible. According to the third and fourth constraints in model (2.3),
33 consider the case when 𝑦 is changed to 𝑦 𝑁 , it is clear that in the fourth constraint there exist at
34
35 least a 𝐷𝑀𝑈𝑙 for which 𝜆2𝑙 ≠ 0 that can satisfy this constraint for all 𝑟 = 1, … , 𝑠. Considering
36 this amount for 𝜆2𝑙 ≠ 0 it is not a guarantee that it can satisfy the third constraint for all 𝑓 =
37
38 1, … , 𝑘. The reason is that 𝑧𝑓𝑜 is a constant value which is the fth value for the intermediate
39 product for 𝐷𝑀𝑈𝑜 . Thus, model (2.3) might be infeasible.
40
41
42 Pitfall 2 - Model (2.3) is feasible when 𝑦 is changed to 𝑦 𝑁 , but the estimate parameters from
43 solving this model are the same as the initial values.
44
45
46 Consider model (2.3) with following presentation shown in model (2.4).
47
48
49
𝑚
50
51 min ∑ 𝜎𝑖 + 0 𝜆1 + 0 𝜆2
52
𝑖=1
53
54 s.t.
55 ∑𝑛𝑗=1 𝜆𝑗1 𝑥𝑖𝑗 + 0 𝜆2 ≤ 𝜎𝑖 , 𝑖 = 1, … , 𝑚,
56 2
57 ∑𝑛𝑗=1 𝜆𝑗1 𝑧𝑓𝑗 + 0𝜆 ≥ 𝑧𝑓𝑜 , 𝑓 = 1, … , 𝑘, (2.4)
1
58
0𝜆 + ∑𝑛𝑗=1 𝜆𝑗2 𝑧𝑓𝑗 ≤ 𝑧𝑓𝑜 , 𝑓 = 1, … , 𝑘,
59
60 0 𝜆1 + ∑𝑛𝑗=1 𝜆𝑗2 𝑦𝑟𝑗 ≤ (𝜑 ∗ − 𝛼)𝑦𝑟𝑜
𝑁
, 𝑟 = 1, … , 𝑠,
61
62
63
64
65
1
2
3
4
𝜆𝑗1 ≥ 0, 𝑗 = 1, … , 𝑛,
5
6 𝜆𝑗2 ≥ 0, 𝑗 = 1, … , 𝑛,
7
8 𝑥𝑖𝑜 ≤ 𝜎𝑖 𝑖 = 1, … , 𝑚.
9
10
11
The third, fourth, and fifth Constraints are redundant as in objective function multiplier of 𝜆2 is
12 zero. Thus model (2.4) can be reduced to model (2.5).
13
14 𝑚
15
16 min ∑ 𝜎𝑖 + 0 𝜆1
17 𝑖=1
18 s.t.
19
20 ∑𝑛𝑗=1 𝜆𝑗1 𝑥𝑖𝑗 ≤ 𝜎𝑖 , 𝑖 = 1, … , 𝑚,
21 ∑𝑛𝑗=1 𝜆𝑗1 𝑧𝑓𝑗 ≥ 𝑧𝑓𝑜 , 𝑓 = 1, … , 𝑘, (2.5)
22
23 𝜆𝑗1 ≥ 0, 𝑗 = 1, … , 𝑛,
24
25 𝑥𝑖𝑜 ≤ 𝜎𝑖 , 𝑖 = 1, … , 𝑚.
26
27
28
29 Theorem 1: Consider model (2.5), the fourth constraint is binding in the optimal solution. That
30 means, in any optimal solution of model (2.5) we have 𝜎𝑖∗ = 𝑥𝑖𝑜 for all 𝑖 = 1, … , 𝑚.
31
32
Proof: Consider model (2.5), a feasible solution for this model is as following.
33
34 𝜎𝑖∗ = 𝑥𝑖𝑜 , ∀𝑖 = 1, … , 𝑚; 𝜆𝑗1 = 0, ∀𝑗 = 1, … , 𝑛, 𝑗 ≠ 𝑜 (2.6)
35
36
37 Since, model (2.5) is minimization thus introduced feasible solution in (2.6) is optimal solution.
38
39 Therefore, the fourth constraint is binding in the optimal solution.
40
41 Now, consider the following model (2.7) which is the dual of model (2.4).
42
43
44 max ∑𝑘𝑓=1 𝑤𝑓′ 𝑧𝑓𝑜 + ∑𝑠𝑟=1 𝑢𝑟 (𝜑 ∗ − 𝛼)𝑦𝑟𝑜𝑁
− ∑𝑘𝑓=1 𝑤𝑓′′ 𝑧𝑓𝑜 + ∑𝑚 ′
𝑖=1 𝑣𝑖 𝑥𝑖𝑜
45
46 s.t.
𝑁
47 ∑𝑠𝑟=1 𝑢𝑟 (𝜑 ∗ − 𝛼)𝑦𝑟𝑗 − ∑𝑘𝑓=1 𝑤𝑓′′ 𝑧𝑓𝑗 ≤ 0, 𝑗 = 1, … , 𝑛,
48
49 ∑𝑘𝑓=1 𝑤𝑓′ 𝑧𝑓𝑗 − ∑𝑚
𝑖=1 𝑣𝑖 𝑥𝑖𝑗 ≤ 0, 𝑗 = 1, … , 𝑛, (2.7)
50
51 𝑣𝑖 𝑣𝑖′
+ ≤ 1, 𝑖 = 1, … , 𝑚,
52 𝑣𝑖 ≥ 0, 𝑣𝑖′ ≥ 0, 𝑖 = 1, … , 𝑚,
53
54 𝑤𝑓′ ≥ 0, 𝑤𝑓′′ ≥ 0, 𝑓 = 1, … , 𝑘,
55 𝑢𝑟 ≥ 0, 𝑟 = 1, … , 𝑠.
56
57
58
59 Dual of model (2.5) is formulated as model (2.8).
60
61
62
63
64
65
1
2
3
4
5
6 max ∑𝑘𝑓=1 𝑤𝑓′ 𝑧𝑓𝑜 + ∑𝑚 ′
𝑖=1 𝑣𝑖 𝑥𝑖𝑜
7 s.t.
8
9 ∑𝑘𝑓=1 𝑤𝑓′ 𝑧𝑓𝑗 − ∑𝑚
𝑖=1 𝑣𝑖 𝑥𝑖𝑗 ≤ 0, 𝑗 = 1, … , 𝑛,
10
11 𝑣𝑖 + 𝑣𝑖′ ≤ 1, 𝑖 = 1, … , 𝑚, (2.8)
12 𝑣𝑖 ≥ 0, 𝑣𝑖′ ≥ 0, 𝑖 = 1, … , 𝑚,
𝑤𝑓′ ≥ 0,
13
14 𝑓 = 1, … , 𝑘.
15
16 The important feature of model (2.8) is that, it also shows that model (2.7) is separable. This is
17
18 the case that has been considered in enveloping form of models (2.4) and (2.5).
19
20 □
21
22
23 3. Solutions for pitfalls in IDEA network models
24 In this section a solution is presented to overcome the difficulties that may occur in the inverse
25
26 data envelopment analysis model with network structures as shown in section 2. Now, to
27 overcome two difficulties introduced above for which the correct result of model (2.4) is not
28
29 obtained, the following model (3.9) is suggested. In model (3.9), to consider the relationships
30 between network stages, we consider the second constraint in model (3.9) instead of the second
31 and third constraints in model (2.4). In this way, model (3.9) can be used to evaluate the output
32
33 efficiency of networks.
34
35 max 𝜑
36
37 s.t.
38 ∑𝑛𝑗=1 𝜆𝑗1 𝑥𝑖𝑗 ≤ 𝜎𝑖 , 𝑖 = 1, … , 𝑚,
39
40 ∑𝑛𝑗=1 𝜆𝑗1 𝑧𝑓𝑗 ≥ ∑𝑛𝑗=1 𝜆𝑗2 𝑧𝑓𝑗 , 𝑓 = 1, … , 𝑘, (3.9)
41
∑𝑛𝑗=1 𝜆𝑗2 𝑦𝑟𝑗 ≤ 𝜑 𝑁
𝑦𝑟𝑜 , 𝑟 = 1, … , 𝑠,
42
43 𝜆𝑗1 ≥ 0 , 𝜆𝑗2 ≥ 0, 𝑗 = 1, … , 𝑛,
44
45 𝜎𝑖 ≥ 𝑥𝑖𝑜 , 𝑖 = 1, … , 𝑚.
46
47
48 Consider this fact that efficiency score obtained from model (3.9) is the same as efficiency score
49 obtained from model (2.1). Considering the efficiency scores obtained from model (3.9), we
50 introduce model (3.10) for estimating input parameters of the network provided that the
51
52 technical efficiency remains constant, i.e., assuming α = 0. Model (3.10) is a multi-objective
53 mathematical model, which will be transformed into a single-objective model (3.11) with the help
54
55
of goal programming technique.
56
57 min(𝜎1 , … , 𝜎𝑚 )
58
59 s.t.
60 ∑𝑛𝑗=1 𝜆𝑗1 𝑥𝑖𝑗 ≤ 𝜎𝑖 , 𝑖 = 1, … , 𝑚,
61
62
63
64
65
1
2
3
4 ∑𝑛𝑗=1 𝜆𝑗1 𝑧𝑓𝑗 ≥ ∑𝑛𝑗=1 𝜆𝑗2 𝑧𝑓𝑗 , 𝑓 = 1, … , 𝑘, (3.10)
5
6 ∑𝑛𝑗=1 𝜆𝑗2 𝑦𝑟𝑗 ≤ (𝜑 ∗ − 𝛼)𝑦𝑟𝑜
𝑁
, 𝑟 = 1, … , 𝑠,
7
8 𝜆𝑗1 ≥ 0 , 𝜆𝑗2 ≥ 0, 𝑗 = 1, … , 𝑛,
9 𝜎𝑖 ≥ 𝑥𝑖𝑜 , 𝑖 = 1, … , 𝑚.
10
11
12 In model (3.11), the objective function is considered as the weighted sum of the components of
13
14
vector 𝜎. These weights are introduced by managers under the condition that 𝑤𝑖 ≥ 0, and
15 ∑𝑚𝑖=1 𝑤𝑖 = 1.
16
17 𝑚
18
19 min ∑ 𝑤𝑖 𝜎𝑖
20 𝑖=1
21 s.t.
22
23 ∑𝑛𝑗=1 𝜆𝑗1 𝑥𝑖𝑗 ≤ 𝜎𝑖 , 𝑖 = 1, … , 𝑚,
24 ∑𝑛𝑗=1 𝜆𝑗1 𝑧𝑓𝑗 ≥ ∑𝑛𝑗=1 𝜆𝑗2 𝑧𝑓𝑗 , 𝑓 = 1, … , 𝑘, (3.11)
25
𝑛 2 ∗ 𝑁
26 ∑𝑗=1 𝜆𝑗 𝑦𝑟𝑗 ≤ (𝜑 − 𝛼)𝑦𝑟𝑜 , 𝑟 = 1, … , 𝑠,
27
28 𝜆𝑗1 ≥ 0 , 𝜆𝑗2 ≥ 0, 𝑗 = 1, … , 𝑛,
29 𝜎𝑖 ≥ 𝑥𝑖𝑜 , 𝑖 = 1, … , 𝑚.
30
31
32
33 Theorem 1: Model (3.11) is always feasible.
34
35 Proof: Without loss of generality let 𝛼 = 0. This means the efficiency score is considered to be
36 fixed and no efficiency improvement is assumed. Considering the increased amount of output,
37
38
i.e., 𝑦 𝑁 . As 𝑦 𝑁 is greater than y, thus:
39 𝜆2𝑜 = 1, 𝜆𝑗2 = 0 ∀𝑗 ≠ 𝑜, 𝜆1𝑜 = 1, 𝜆𝑗1 = 0 ∀𝑗 ≠ 𝑜, and 𝜎𝑖 = 𝑥𝑖𝑜 , ∀𝑖, (3.12)
40
is a feasible solution for model (3.11). □
41
42
43 According to Theorem 1, difficulty of infeasibility cannot occur in model (3.11).
44
45 Remark2: In estimating the input parameter, i.e., the optimal solution for the vector 𝜎⃗ =
46 (𝜎1 , … , 𝜎𝑚 ), the estimated values are not necessarily equal to the initial values of the inputs as
47 in model (2.6). The reason is that model (3.11) in contrast to model (2.4) is not a separable
48
49 optimization problem according to 𝜆2 and 𝜆1 .
50 □
51
52
53 The reason that the above two points are always satisfied in model (3.11) is that model (3.11)
54
55 tries to find a feasible value for 𝜆2 and 𝜆1 with the perturbed value of 𝑦 , that is 𝑦 𝑁 . With a
56 new value of 𝑦 𝑁 , a value is established for the intensifier variable 𝜆2 , then in regards of the third
constraint it also affects 𝜆1 until the second constraint is satisfied. Since the intermediate
57
58
59 product, Z, which has known and fixed numerical value is not considered to relates two stages
60
61
62
63
64
65
1
2
3
4 together, thus mentioned problems do not occur. Therefore, the input parameters, i.e., 𝜎⃗ =
5
6 (𝜎1 , … , 𝜎𝑚 ) can be estimated according to the first constraint, and the estimated value for the
7 vector σ is not necessarily the same as the initial values of the input vector x.
8
9
10 Theorem 4: It is proved that efficiency score obtained from model (3.13) and model (3.15) are
11 the same, i.e., 𝜑 ∗𝑡 = 𝜑 ∗ .
12
13 Proof: Consider Model (3.13), (3.14), and (3.15).
14
15
16 m𝑎𝑥 φ
17 s.t.
18
∑𝑛𝑗=1 𝜆𝑗1 𝑥𝑖𝑗 ≤ 𝑥𝑖𝑜 , 𝑖 = 1, … , 𝑚,
19
20 ∑𝑛𝑗=1 𝜆𝑗1 𝑧𝑓𝑗 ≥ ∑𝑛𝑗=1 𝜆𝑗2 𝑧𝑓𝑗 , 𝑓 = 1, … , 𝑘,
21
22 (3.13)
23 ∑𝑛𝑗=1 𝜆𝑗2 𝑦𝑟𝑗 ≤ (𝜑 )𝑦𝑟𝑜 , 𝑟 = 1, … , 𝑠,
24
25 𝜆𝑗1 ≥0 , 𝜆𝑗2 ≥ 0, 𝑗 = 1, … , 𝑛.
26
27 𝑚
28
29 min ∑ 𝜎𝑖
30 𝑖=1
31
s.t.
32
33 ∑𝑛𝑗=1 𝜆𝑗1 𝑥𝑖𝑗 ≤ 𝜎𝑖 , 𝑖 = 1, … , 𝑚,
∑𝑛𝑗=1 𝜆𝑗1 𝑧𝑓𝑗 ≥ ∑𝑛𝑗=1 𝜆𝑗2 𝑧𝑓𝑗 ,
34
35 𝑓 = 1, … , 𝑘, (3.14)
36 ∑𝑛𝑗=1 𝜆𝑗2 𝑦𝑟𝑗 ≤ (𝜑 ∗ )𝑦𝑟𝑜
𝑁
, 𝑟 = 1, … , 𝑠,
37
38 𝜆𝑗1 ≥ 0 , 𝜆𝑗2 ≥ 0, 𝑗 = 1, … , 𝑛,
39
𝜎𝑖 ≥ 𝑥𝑖𝑜 , 𝑖 = 1, … , 𝑚.
40
41
42 max 𝜑 𝑡
43
44
s.t.
45 ∑𝑛𝑗=1 𝜆𝑗1 𝑥𝑖𝑗 + 𝜆1𝑛+1 𝜎𝑖∗ ≤ 𝜎𝑖∗ , 𝑖 = 1, … , 𝑚, (3.15)
∑𝑛𝑗=1 𝜆𝑗1 𝑧𝑓𝑗 + 𝜆1𝑛+1 𝑧𝑓𝑜 ≥ ∑𝑛𝑗=1 𝜆𝑗2 𝑧𝑓𝑗 + 𝜆2𝑛+1 𝑧𝑓𝑜
46
47 𝑓 = 1, … , 𝑘,
48 ∑𝑛𝑗=1 𝜆𝑗2 𝑦𝑟𝑗 + 𝜆2𝑛+1 𝑦𝑟𝑜
𝑁
≥ (𝜑 𝑡 )𝑦𝑟𝑜
𝑁
, 𝑟 = 1, … , 𝑠,
49
50 𝜆𝑗1 ≥ 0 , 𝜆𝑗2 ≥ 0, 𝑗 = 1, … , 𝑛.
51
52
53 Model (3.13) has n DMUs to be evaluated, but Model (3.15) has n+1 DMUs. Consider 𝐷𝑀𝑈𝑜 is
54 under evaluation, the extra 𝐷𝑀𝑈 in model (3.15) is ( 𝜎 ∗ , 𝑦𝑜𝑁 ). Theorem 4 proves that the
55
56 efficiency the efficiencies of ( 𝑥𝑜 , 𝑦𝑜𝑁 ) and ( 𝜎 ∗ , 𝑦𝑜𝑁 ) are the same.
57 Consider Model (3.13) for efficiency evaluation of set of n DMUs. Model (3.14) tries to find the
58
59 minimum possible input increase when outputs are increased by decision makers. Model (3.15)
60 is similar to Model (3.13) but it has one more DMU as ( 𝜎 ∗ , 𝑦𝑜𝑁 ).
61
62
63
64
65
1
2
3
4
5
Consider optimal solution of model (3.13) and (3.14) respectively as (𝜆1∗ , 𝜆2∗ , φ∗ ) and
6 (𝜆1∗ , 𝜆2∗ , 𝜎 ∗ ). Optimal solution of model (3.15) is (𝜆1𝑡∗ , 𝜆2𝑡∗ , φ∗𝑡 ) where 𝜆1𝑡∗ = (𝜆1∗ , 𝜆1∗
𝑛+1 )1×𝑛+1
2𝑡∗ 2∗ 2∗
7 and 𝜆 = (𝜆 , 𝜆𝑛+1 )1×𝑛+1 .
8
9
10 According to the optimal solution of model (3.14), (𝜆1∗ , 𝜆2∗ , 𝜎 ∗ ), a feasible solution for model
11
12 (3.15) is constructed as((𝜆1∗ , 0)1×𝑛+1 , (𝜆2∗ 0)1×𝑛+1 , φ ).
13
14 Now, consider the first, second, and third constraints in model (3.14). According to this fact that
15 φ∗ ≥ 1 thus according to the third constraint:
16
17
∑𝑛𝑗=1 𝜆𝑗2∗ 𝑦𝑟𝑗 ≥ 𝜑 ∗ 𝑦𝑟𝑜
𝑁 𝑁
≥ 𝑦𝑟𝑜 , ∀𝑟 (3.16)
18 Moreover, according to model (3.15);
19
20
∑𝑛𝑗=1 𝜆𝑗2∗𝑡 𝑦𝑟𝑗 + 𝜆2∗𝑡 𝑁 𝑛 2∗𝑡 2∗𝑡 𝑛 2
𝑛+1 𝑦𝑟𝑜 ≤ ∑𝑗=1 𝜆𝑗 𝑦𝑟𝑗 + 𝜆𝑛+1 (∑𝑗=1 𝜆𝑗 𝑦𝑟𝑗 ) =
21 (3.17)
𝑛 2∗𝑡 𝑛 2∗ 2∗𝑡 𝑛 2∗𝑡 2∗ 2∗𝑡
22
= ∑𝑗=1 𝜆𝑗 𝑦𝑟𝑗 + ∑𝑗=1 𝜆𝑗 𝜆𝑛+1 𝑦𝑟𝑗 = ∑𝑗=1(𝜆𝑗 +𝜆𝑗 𝜆𝑛+1 ) 𝑦𝑟𝑗 , ∀𝑟
23
24
25 According to first constraint in model (3.14) as 𝜎 ∗ ≥ 𝑥𝑜 thus;
26
27 ∑𝑛𝑗=1 𝜆𝑗∗1 𝑥𝑖𝑗 ≤ 𝑥𝑖𝑜 ≤ 𝜎𝑖∗ , ∀𝑖 (3.18)
28 Now according to model (3.15);
29
30 ∑𝑛𝑗=1 𝜆𝑗1∗𝑡 𝑥𝑖𝑗 + 𝜆1∗𝑡 ∗ 𝑛 1∗𝑡 1∗𝑡 𝑛 1∗𝑡 1∗𝑡 𝑛 ∗1
𝑛+1 𝜎𝑖 ≥ ∑𝑗=1 𝜆𝑗 𝑥𝑖𝑗 + 𝜆𝑛+1 𝑥𝑖𝑜 ≥ ∑𝑗=1 𝜆𝑗 𝑥𝑖𝑗 + 𝜆𝑛+1 (∑𝑗=1 𝜆𝑗 𝑥𝑖𝑗 ) =
31
32
= ∑𝑛𝑗=1(𝜆𝑗1∗𝑡 + 𝜆1∗𝑡 ∗1
𝑛+1 𝜆𝑗 )𝑥𝑖𝑗 , ∀𝑖 (3.19)
33 Therefore, feasible intensifies variables for model (3.15) is obtained which satisfies all
34 constraints;
35
36 𝜆̃1 = (𝜆1∗ , 𝜆1∗𝑡 ̃2 2∗ 2∗𝑡
𝑛+1 )1×𝑛+1 , 𝜆 = (𝜆 , 𝜆𝑛+1 )1×𝑛+1 (3.20)
37
38
39 Now it should be proved that 𝜑 ∗ = 𝜑 ∗𝑡 . By contradiction suppose that 𝜑 ∗ < 𝜑 ∗𝑡 it is possible to
40 have two cases:
41
42 a) 𝑥𝑜 = 𝜎 ∗
43 In this case there exist a feasible solution for model (3.13) such as (𝜆̃2 , 𝜑 ∗𝑡 ) which has a
44
45
better objective value than existing optimal objective function 𝜑 ∗ which is a contradiction
46 with optimality of model (3.13).
47 b) 𝜎 ∗ ≩ 𝑥𝑜
48
49 In this case there exist at least one index like l for which the equality of 𝑥𝑜 and 𝜎 ∗ does
50 not happen, 𝑥𝑙𝑜 < 𝜎𝑙∗ . It is possible to subtract a small positive number such as ε from 𝜎𝑙∗
51
52 and remains the other indexes with their values. Thus, a feasible solution is constructed;
53 𝜎𝑙∗ − ε = 𝜎̃𝑙∗ , 𝜎̃𝑖∗ = 𝑥𝑖𝑜 𝑖 = 1, … . . , 𝑙 − 1, 𝑙 + 1, … . , 𝑚 (3.21)
54 ∗ ̃
55 Therefore, we have a feasible solution (𝜎̃ , 𝜆 ) with lower efficiency function which is a
56 contradiction with the optimality of 𝜎 ∗ .
57
58
59
60 4. Case Study
61
62
63
64
65
1
2
3
4 4.1. Data sets and results
5
6
7 Considering this application, the authors try to show two pitfalls mentioned in previous section
8
9 and provide the solutions for overcoming these difficulties. Also, the obtained results are also
10 discussed in detains for setting strategies for managers and DMs.
11 Consider the data set in Table 4.1. Consider ten network production systems (DMUs) with two
12
13 sub-processes, supplier and manufacturer. Here, 10 tomato paste two-stage series network
14 (DMUs) is assessed. The dataset dates back to 2019, which is reported in Table 4.1. The list of
15 tomato paste supply chains are derived from Tehran Stock Exchange and Iran Fara-Bourse . This
1 2

16
17 network has with two inputs, one intermediate, and two outputs. Note that the increased
18 outputs (𝑦 𝑁 ) are also specified in Table 1, but no performance improvement is considered, that
19 is 𝛼 = 0. Inputs are labor cost in 1000$, and material costs in 10000$, outputs are revenue in
20
21
100000$, and production in 1000 Ton; and intermediate product is supplied material in 1000 Ton.
22 Managers and Decision Makers (DMs) assumed the increased outputs that are shown in Table
23 4.1.
24
25
26 TABLE 4.1. Data set
27
28 Inputs Outputs Improved Outputs Intermediate
29 DMUs
30 𝑥1 𝑥2 𝑦1 𝑦2 𝑦1𝑁 𝑦2𝑁 𝑧1
31
32 DMU1 9 9 100 80 103 85 45
33 DMU2 24 12 44 32 45 34 65
34
35
DMU3 31 11 75 65 78 67 34
36 DMU4 20 34 39 16 38 17 62
37 DMU5 18 23 86 61 88 63 87
38
DMU6 16 16 24 23 26 25 12
39
40 DMU7 14 76 24 54 25 57 62
41 DMU8 75 20 98 20 99 19 49
42
DMU9 25 17 73 24 75 27 82
43
44 DMU10 72 73 26 63 26 65 61
45
46 To evaluate 10 DMUs that have a network structure similar to Figure 2.1, the results obtained
47
48 from models (2.1) and (2.2) are as mentioned in Table 4.2 and Table 4.3. Note that the efficiency
49 values of each DMU obtained from Model (2.1) are shown in Table 4.2 .Table 4.2 and Table 4.3
50
51 summarizes the results of efficiency evaluation and parameters estimating for DMU 1 and DMU
52 2, using models (2.1) and (2.3).
53 As you can see, DMU 1 is efficient in output orientation using model (2.1). But when its output is
54
55 increased from 80 and 100 to 85 and 103, respectively, provided that the value of efficiency
56 remains constant, i.e., in model (2.3) (𝜑 ∗ − 𝛼) = 1, thus model (2.3) becomes infeasible. In
57
58 model (2.3) 𝑦 was changed to 𝑦 𝑁 , with this new value of the outputs model (2.3) could not find
59 a suitable value for𝜆2∗ thus could not establish constraints (2.3.b), (2.3.c), and (2.3.d). In this case,
60
61
62
63
64
65
1
2
3
4 by considering 𝑦 𝑁 , which exist in the corresponding constraints of the second stage of the
5
6 network, the corresponding relations of the second stage are not established, therefore, a value
7 for 𝑥 is not estimated. In this way, the optimal value for components of vector 𝜎 is not obtained.
8
9 In the iteration that model (2.3) becomes infeasible, the corresponding answer is given in Table
10 4.2. From the evaluation results of DMU 1 in Table 4.2, it is clear that the intensifier variable of
11 the second stage for DMU6 is negative and therefore model (2.3) was infeasible.
12
13 Now consider DMU, results of analyzing this DMU is shown in TABLE 4.3. In model (2.3), consider
14 the constraints (2.3.b) and (2.3.c) that establish the relationship between the first and second
15
16 stages of the network. The relationship between the first and second stages of the network is
17 defined by the fixed and known value of the intermediate product, 𝑧. In this case, if according to
18
the constraint (2.3.d), a new value of 𝜆2∗ is obtained, since 𝑧 is a known and fixed numerical value,
19
20 the changes made in the second stage will not affect the first stage. In this case, from model (2.3),
21 estimate the input parameter to be the same as the initial input value.
22
23
24 TABLE 4.2: Results of DMU1
25 Efficiency score obtained
26 Model (2.3) = Infeasible
from model (2.1) =1
27
28 𝑥1 =9 𝑥2 =9 𝜎1∗ =? 𝜎1∗ =?
29 𝜆1∗
1 1 𝜆12∗ 1 𝜆1∗
1 _ 𝜆12∗ _
30
31 𝜆1∗
2 0 𝜆2∗
2 0 𝜆1∗
2 _ 𝜆2∗
2 _
32 𝜆1∗
3 0 𝜆2∗
3 0 𝜆1∗
3 _ 𝜆2∗
3 _
33 DMU
𝜆1∗
4 0 𝜆2∗
4 0 𝜆1∗
4 _ 𝜆2∗
4 _
34 1
35 𝜆1∗
5 0 𝜆2∗
5 0 𝜆1∗
5 _ 𝜆2∗
5 _
36 𝜆1∗
6 0 𝜆2∗
6 0 𝜆1∗
6 _ 𝜆2∗
6 _
37
38 𝜆1∗
7 0 𝜆2∗
7 0 𝜆1∗
7 _ 𝜆2∗
7 _
39 𝜆1∗
8 0 𝜆2∗
8 0 𝜆1∗
8 _ 𝜆2∗
8 _
40
𝜆1∗
9 0 𝜆2∗
9 0 𝜆1∗
9 _ 𝜆2∗
9 _
41
42 𝜆1∗
10 0 2∗
𝜆10 0 𝜆1∗
10 _ 2∗
𝜆10 _
43
44
45
46 TABLE 4.3: Results of DMU2
47 Optimal objective
48 Efficiency score obtained
function value of
49 from model (2.1) =1
50 model (2.3) =36
51 𝑥1 =24 𝑥2 =12 𝜎1∗ =24 𝜎1∗ =12
52
53 DMU 𝜆1∗
1 0 2∗
𝜆1 1.44 𝜆1∗
1 0 𝜆12∗ 1.44
1∗ 2∗ 1∗
54 2 𝜆2 1 𝜆2 0 𝜆2 1 𝜆2∗
2 0
55 1∗ 2∗ 1∗ 2∗
𝜆3 0 𝜆3 0 𝜆3 0 𝜆3 0
56
1∗ 2∗ 1∗ 2∗
57 𝜆4 0 𝜆4 0 𝜆4 0 𝜆4 0
1∗ 2∗ 1∗ 2∗
58 𝜆5 0 𝜆5 0 𝜆5 0 𝜆5 0
59 1∗ 2∗ 1∗ 2∗
60 𝜆6 0 𝜆6 0 𝜆6 0 𝜆6 0
61
62
63
64
65
1
2
3
4
5
𝜆1∗
7 0 𝜆2∗
7 0 𝜆1∗
7 0 𝜆2∗
7 0
1∗ 2∗ 1∗ 2∗
6 𝜆8 0 𝜆8 0 𝜆8 0 𝜆8 0
7 𝜆1∗
9 0 𝜆2∗
9 0 𝜆1∗
9 0 𝜆2∗
9 0
8 1∗ 2∗ 1∗ 2∗
9 𝜆10 0 𝜆10 0 𝜆10 0 𝜆10 0
10
11 TABLE 4.4 and TABLE 4.5 summarizes the results of models (2.1) and (2.3) for DMU3 and DMU4.
12
13 Similar to DMU1 and DMU2, model (2.3) is not possible and the results of the last iteration to
14 solve the model are shown in TABLE 4.2. As it is clear, the intensifier variables have negative
15
16 values and therefore model (2.3) is infeasible. In DMU4, the optimal estimated values obtained,
17 is exactly equal to the initial input values. The points mentioned about DMU 1, DMU2, DMU3,
18 and DMU4 have also occurred in other units.
19
20
21 TABLE 4.4: Results of DMU 3
22
Efficiency score
23 model (2.3) =
24 obtained from model
(2.1) = 0.6
Infeasible
25
26 𝑥1 =31 𝑥2 =11 𝜎1∗ =? 𝜎1∗ =?
27
1∗
28 𝜆1 0 2∗
𝜆1 0 𝜆1∗
1 _ 𝜆12∗ _
1∗ 2∗ 1∗ 2∗
29 𝜆2 0.5 𝜆2 0 𝜆2 _ 𝜆2 _
30
31 DMU 𝜆1∗
3 0 2∗
𝜆3 1 1∗
𝜆3 _ 𝜆32∗ _
1∗ 2∗ 1∗ 2∗
32 3 𝜆4 0 𝜆4 0 𝜆4 _ 𝜆4 _
33 1∗ 2∗ 1∗ 2∗
𝜆5 0 𝜆5 0 𝜆5 _ 𝜆5 _
34
1∗ 2∗ 1∗ 2∗
35 𝜆6 0 𝜆6 0 𝜆6 _ 𝜆6 _
1∗ 2∗ 1∗ 2∗
36 𝜆7 0 𝜆7 0 𝜆7 _ 𝜆7 _
37 1∗ 2∗ 1∗ 2∗
38 𝜆8 0 𝜆8 0 𝜆8 _ 𝜆8 _
1∗ 2∗ 1∗ 2∗
39 𝜆9 0 𝜆9 0 𝜆9 _ 𝜆9 _
40 1∗ 2∗ 1∗ 2∗
41
𝜆10 0 𝜆10 0 𝜆10 _ 𝜆10 _
42
43
44
TABLE 4.5: Results of DMU 4
45
46 Optimal objective
47 Efficiency score obtained
function value of model
48 from model (2.1) = 0.6
(2.3) = 54
49
50 𝑥1 =20 𝑥2 =34 𝜎1∗ =20 𝜎2∗ =34
1∗
51 𝜆1 1.4 2∗
𝜆1 1.4 𝜆1∗
1 1.4 𝜆12∗ 1.38
52 1∗ 2∗ 1∗
53
DMU
𝜆2 0 𝜆2 0 𝜆2 0 𝜆2∗
2 0
4
1∗ 2∗ 1∗ 2∗
54 𝜆3 0 𝜆3 0 𝜆3 0 𝜆3 0
55 1∗ 2∗ 1∗ 2∗
56 𝜆4 0 𝜆4 0 𝜆4 0 𝜆4 0
1∗ 2∗ 1∗ 2∗
57 𝜆5 0 𝜆5 0 𝜆5 0 𝜆5 0
58 1∗ 2∗ 1∗ 2∗
𝜆6 0 𝜆6 0 𝜆6 0 𝜆6 0
59 1∗ 2∗ 1∗ 2∗
60 𝜆7 0 𝜆7 0 𝜆7 0 𝜆7 0
61
62
63
64
65
1
2
3
4
5
𝜆1∗
8 0 𝜆2∗
8 0 𝜆1∗
8 0 𝜆2∗
8 0
1∗ 2∗ 1∗ 2∗
6 𝜆9 0 𝜆9 0 𝜆9 0 𝜆9 0
7 𝜆1∗
10 0 2∗
𝜆10 0 𝜆1∗
10 0 2∗
𝜆10 0
8
9
10
11
12 Now consider the data set tabulated in TABLE 4.1 and solve models (3.9) and (3.11) respectively
13
14 for efficiency evaluation and parameter estimation. Model (3.9) calculates the efficiency of
15 network in the output orientation. Model (3.11) estimates the minimum input value with the
16
17 new output value, 𝑦 𝑁 , the efficiency score remains constant and assuming α = 0. Note that
18 efficiency score obtained from model (2.1) and (3.9) are the same. According to the data of Table
19
20
4.1 and models (3.9) and (3.11), the results obtained for estimating the values of the input vector
21 are shown in Table 4.6 and TABLE 4.7. According to TABLE 4.3, from the optimal answer of model
22 (3.11) for DMU1 and DMU 2, the estimated values of inputs 𝜎⃗ ∗ = (15.49, 9) , are higher than the
23
24 initial input values, 𝑥⃗ = (15, 9).
25
26
27
28
TABLE 4.6. Results for DMU 1
29 Efficiency score obtained Optimal objective function
30 from model (3.9) = 0.62 value of model (3.11) = 13.42
31
32 𝑥1 =15 𝑥2 =9 𝜎1∗ =15.49 𝜎2∗ =9
33
34 𝜆1∗
1 0 𝜆12∗ 1 𝜆1∗
1 0 𝜆12∗ 0.64
35 𝜆1∗
2 0 𝜆2∗
2 0 1∗
𝜆2 0 𝜆2∗
2 0
36
𝜆1∗
3 0 𝜆32∗ 0 1∗
𝜆3 0 𝜆2∗
3 0.52
37
38 DMU 𝜆1∗
4 0 𝜆42∗ 0 1∗
𝜆4 0 𝜆2∗
4 0
39 1
𝜆1∗
5 0.52 𝜆2∗
5 0 𝜆1∗
5 0.53 𝜆2∗
5 0
40
41 𝜆1∗
6 0 𝜆2∗
6 0 𝜆1∗
6 0 𝜆2∗
6 0
42
43 𝜆1∗
7 0 𝜆2∗
7 0 𝜆1∗
7 0 𝜆2∗
7 0
44
𝜆1∗
8 0 𝜆2∗
8 0 𝜆1∗
8 0 𝜆2∗
8 0
45
46 𝜆1∗
9 0 𝜆2∗
9 0 𝜆1∗
9 0 𝜆2∗
9 0
47 𝜆1∗
10 0 2∗
𝜆10 0 𝜆1∗
10 0 2∗
𝜆10 0
48
49
50
51
52
53
TABLE 4.7. Results for DMU 2
54 Optimal objective function
Efficiency score obtained from
55 value of model (3.11) =
model (3.9) = 0.3
56 36.47
DMU
57
2 𝑥1 =24 𝑥2 =12 𝜎1∗ =24 𝜎2∗ =12.47
58
59 𝜆1∗
1 0 𝜆12∗ 0.44 𝜆1∗
1 0 𝜆12∗ 0.45
60 𝜆1∗
2 0.3 𝜆2∗
2 0 𝜆1∗
2 0.27 𝜆2∗
2 0
61
62
63
64
65
1
2
3
4
5 𝜆1∗
3 0 𝜆2∗
3 0 𝜆1∗
3 0 𝜆2∗
3 0
6 𝜆1∗
4 0 𝜆2∗
4 0 𝜆1∗
4 0 𝜆2∗
4 0
7
8 𝜆1∗
5 0 𝜆2∗
5 0 𝜆1∗
5 0 𝜆2∗
5 0
1∗ 2∗ 1∗ 2∗
9 𝜆6 0 𝜆6 0 𝜆6 0 𝜆6 0
10
11 𝜆1∗
7 0 𝜆2∗
7 0 𝜆1∗
7 0 𝜆2∗
7 0
1∗ 2∗ 1∗ 2∗
12 𝜆8 0 𝜆8 0 𝜆8 0 𝜆8 0
13
14 𝜆1∗
9 0 𝜆2∗
9 0 𝜆1∗
9 0.03 𝜆2∗
9 0
15 𝜆1∗
10 0 2∗
𝜆10 0 𝜆1∗
10 0 2∗
𝜆10 0
16
17
18 As it turned out, model (3.11) is feasible and the problems mentioned in model (2.3) did not occur
19 in model (3.11).
20
21 TABLE 4.8 and TABLE 4.9 summarize the results obtained for the evaluation of DMU5 and DMU4.
22 As it is known, for DMU4 estimated values of the input is equal to 𝜎⃗ ∗ = (𝜎1∗ , 𝜎2∗ ) = (20 ,34)
23
24 which, is equal to the initial inputs 𝑥⃗ = (𝑥1 , 𝑥2 ) = (20,34). As it turns out, for assessing DMU4
25 model (2.7) is feasible but no more input values are estimated. In unit 5 for the first and second
26
27 inputs the estimated input values are 𝜎⃗ ∗ = (𝜎1∗ , 𝜎2∗ ) = (18.4,23.5) while the initial values are
28 𝑥⃗ = (𝑥1 , 𝑥2 ) = (18,23). As it turns out, model (3.11) is feasible and more input values is
29
30 estimated for the input parameter than their initial values.
31
32 TABLE 4.8. Results of DMU 3
33
34 Efficiency score obtained from Optimal objective function
35 model (3.9) = 0.44 value of model (3.11) = 41.95
36
𝑥1 =18 𝑥2 = 23 𝜎1∗ = 18.4 𝜎2∗ =23.5
37
38 𝜆1∗
1 0 𝜆12∗ 0.86 𝜆1∗
1 0 𝜆12∗ 0.88
39 𝜆1∗
2 0 𝜆2∗
2 0 1∗
𝜆2 0 𝜆2∗
2 0
40
41 𝜆1∗
3 0 𝜆2∗
3 0 𝜆1∗
3 0 𝜆2∗
3 0
42 DMU 𝜆1∗
4 0 𝜆2∗
4 0 𝜆1∗
4 0 𝜆2∗
4 0
43 5
1∗ 2∗ 1∗ 2∗
44 𝜆5 0.44 𝜆5 0 𝜆5 0.46 𝜆5 0
45 𝜆1∗
6 0 𝜆2∗
6 0 𝜆1∗
6 0 𝜆2∗
6 0
46
47 𝜆1∗
7 0 𝜆2∗
7 0 𝜆1∗
7 0 𝜆2∗
7 0
1∗ 2∗ 1∗ 2∗
48 𝜆8 0 𝜆8 0 𝜆8 0 𝜆8 0
49
50 𝜆1∗
9 0 𝜆2∗
9 0 𝜆1∗
9 0 𝜆2∗
9 0
51 𝜆1∗
10 0 2∗
𝜆10 0 𝜆1∗
10 0 2∗
𝜆10 0
52
53
54 TABLE 4.9. Results of DMU 4
55
Efficiency score obtained from Optimal objective function
56
model (3.9) = 0.18 value of model (3.11) = 54
57 DMU
58 4
𝑥1 =20 𝑥2 = 34 𝜎1∗ = 20 𝜎2∗ =34
59
60 𝜆1∗
1 0 𝜆12∗ 0.39 𝜆1∗
1 0 𝜆12∗ 0.39

61
62
63
64
65
1
2
3
4
5
𝜆1∗
2 0 𝜆2∗
2 0 𝜆1∗
2 0 𝜆2∗
2 0
1∗ 2∗ 1∗ 2∗
6 𝜆3 0 𝜆3 0 𝜆3 0 𝜆3 0
7 𝜆1∗
4 0 𝜆2∗
4 0 𝜆1∗
4 0 𝜆2∗
4 0
8
1∗ 2∗ 1∗ 2∗
9 𝜆5 0.2 𝜆5 0 𝜆5 0.2 𝜆5 0
10 𝜆1∗
6 0 𝜆2∗
6 0 𝜆1∗
6 0 𝜆2∗
6 0
11
12 𝜆1∗
7 0 𝜆2∗
7 0 𝜆1∗
7 0 𝜆2∗
7 0
13 1∗ 2∗ 1∗ 2∗
𝜆8 0 𝜆8 0 𝜆8 0 𝜆8 0
14
15 𝜆1∗
9 0 𝜆2∗
9 0 𝜆1∗
9 0 𝜆2∗
9 0
16 𝜆1∗
10 0 2∗
𝜆10 0 𝜆1∗
10 0 2∗
𝜆10 0
17
18
19
20
21
22
The rest of the evaluated units also have results similar to those discussed for the above four
23 units. To avoid repeating points and increasing the number of pages, they have been avoided. It
24 should be noted that data envelopment analysis models with various general (series-parallel)
25
26 network structure have been introduced in the literature.
27 The problems mentioned in this study can also occur in inverse models of data envelopment
28
analysis with general, series, and parallel network structures. Therefore, the method introduced
29
30 in this study can be used to solve pitfalls in these types of networks.
31
32
33 4.2. Discussion of analysis of results
34
35
Take TABLE 4.10 into account. The optimal estimated inputs of the network depicted in FIGURE
36
37 1 are shown in TABLE 4.10. It should be noted that having the estimated inputs, (𝜎1∗ , 𝜎2∗ ), managers
38 can set different strategies while having the efficiency score constant. By increasing the outputs according
39 to the managers and DM preferences, the estimated inputs are as mentioned in TABLE 4.10. An important
40
41 note is that all the estimated inputs are either equal or greater than the initial inputs values. DMUs 2, 3,
42 6, 8, 9, and 10 have witnessed an increase in their second inputs. Moreover, DMUs 1, and 7 have witnessed
43 an increase in their second inputs. Meanwhile, DMU5 witnessed an increase in both of the inputs. But,
44
45
DMU 4 has not increased its inputs and the estimated values are equal to its initial inputs. It can be said
46 that 90% DMUs have increased their inputs while the outputs are increased and the efficiency scores are
47 kept constant. Meanwhile 10% of DMUs do not increase their input values.
48
49
50
51 TABLE 4.10. Estimated inputs from model (3.11)
52
Optimal
53
54 DMUs objective 𝜎1∗ 𝜎2∗ 𝑥1 𝑥1
55 value
56 DMU1 113.49 15.49 9 9 9
57
DMU2 36.47 24 12.47 24 12
58
59 DMU3 42.43 31 11.43 31 11
60 DMU4 54.00 20 34 20 34
61
62
63
64
65
1
2
3
4 DMU5 41.95 18.42 23.53 18 23
5
6 DMU6 32.10 16 18.2 16 16
7 DMU7 90.78 14.78 76 14 76
8 DMU8 95.20 75 20.2 75 20
9
10 DMU9 42.86 25 17.86 25 17
11 DMU10 148.70 72 76.65 72 73
12
13
14
15 Consider FIGURE 2 where the trend of inputs alterations for all the DMUs are depicted. As is seen
16
17
there exist a difference between the initial and the estimate values of the first and second inputs.
18
19
20
21
22 80
23
24 70
25
26
27 60
28
29
30 50
σ1*
31
32 X1
40
33 σ2*
34
35 X2
30
36
37
38 20
39
40
41 10
42
43 0
44 DMU1 DMU2 DMU3 DMU4 DMU5 DMU6 DMU7 DMU8 DMU9 DMU10
45
46 FIGURE 2. Comparison of initial and estimated values
47
48
49 The estimated inputs along with the increased outputs can be used as alternative strategies for
50 managers and DMs. Consider DMU1. This unit considers the increased outputs (103, 85) as
51
52
managers preferences. Also, its efficiency is kept constant with the value of 1.00. Therefore, the
53 estimated inputs are calculated as (15.49, 9). Thus, DMU 1 have the same performance while the
54 outputs and inputs are increased. Consider DMU5. Its increased outputs are (88, 63) and the
55
56 efficiency score (0.44) is kept constant, thus the estimated inputs are obtained as (18.42, 23.53).
57 This procedure can also be considered for other DMUs and alternative strategies for input
58
59
consumption and output productions are suggested.
60
61
62
63
64
65
1
2
3
4 5. Managerial Implications
5
6
7 Sensitivity analysis is very important for managers and DMs. Because changes in the input and
8
9 output values of the system under their guidance can help analyze the results as well as
10 determine multiple strategies to guide the system in the future. In this study, the output of a
11 network system under the supervision of its managers and decision makers is increased and by
12
13 considering the efficiency value constant, it is tried to obtain the estimated values of the inputs
14 of the network system. By performing this sensitivity analysis, the percentage of input and
15
16 output changes can be checked by maintaining the system efficiency.
17 Consider DMU 5. This DMU is an efficient DMU. By 103% and 106.25% times increasing of the
18 outputs respectively and keeping the efficiency value constant, the estimated inputs are obtained
19
20 by 172.11% and 100% times increasing of the initial inputs respectively. This means by average
21 of 104.62% times increasing of the outputs, the estimated inputs are obtained by 136.05% times
22
23 increasing from the initial inputs. Consider DMU 1. This DMU is an efficient DMU. By 108.63%
24 and 108.39% times increasing of the outputs respectively and keeping the efficiency value
25 constant, the estimated inputs are obtained by 102.33% and 102.30% times increasing of the
26
27 initial inputs respectively. This means by average of 108.51% times increasing of the outputs, the
28 estimated inputs are obtained by 102.31% times increasing from the initial inputs. As a crucial
29
30
result that can be suggested to managers after solving the presented method in this application
31 is that the alternative strategy can be found for each DMUs. According to the results shown in
32 Table 4.10 and the analysis discussed above, the best strategy can be suggested for DMU5
33
34 because by increasing the output and keeping the efficiency constant, the percentage of increase
35 in system’s inputs is less than the percentage increase in system’s outputs, and this strategy is
36
37
very desirable and advantageous for the system under evaluation.
38
39
40
41
6. Conclusion:
42 Data envelopment analysis models were introduced and rapidly developed to evaluate the
43 efficiency and performance of a set of decision-making units . These models were used in various
44
45 fields and were developed by researchers in terms of application and theory. Since in many real-
46 world examples, production systems, organizations, or activities have a networked structure,
47
48
network envelopment analysis network models were introduced in the literature and then
49 developed. In addition to evaluating performance for managers and decision makers, they need
50 other information from the system under their control to be able to perform further analysis for
51
52 planning, risk-taking, and decision-making. Therefore, inverse models of data envelopment
53 analysis were introduced. These models are divided into two general categories: resource
54
55
allocation model and investment analysis model. It is clear that system managers and decision
56 makers are eager to apply their important preferences and policies to resources, including system
57 inputs and outputs when analyzing production, resource allocation process, increasing resource
58
59 efficiency on the network system. Therefore, this issue is of special importance. In this study,
60 theoretical and practical problems that can occur for the inverse model of data envelopment
61
62
63
64
65
1
2
3
4 analysis with network structure are introduced and these problems are introduced and
5
6 discussed. These problems are introduced for network systems with series structure. Solutions
7 to get rid of these problems are also proposed in the inverse model of network data envelopment
8
9 analysis. By considering the problems introduced and the solutions mentioned, we can get rid of
10 the pitfalls that may occur in the inverse model of network data envelopment analysis .The
11 mentioned shortcomings of a series network systems in this study might be exist in other types
12
13 of network systems such as general and parallel. The limitation of this study is that the mentioned
14 models and formulations are introduced for series networks, and in case of having other types of
15
16 network systems they are not immediately applicable. Also, for future research it is useful to
17 formulate other types of network system and try to find a solution for the mentioned
18 shortcomings in this study in those systems.
19
20
21
22
23
References:
24
25 M.E. Adimi, M. Rostamy-Malkhalifeh, F.H. Lotfi and R. Mehrjoo, A model to evaluate the effects of the
26 returns to scale on the inverse data envelopment analysis. Mathematical Sciences. 15 (2021) 111-121.
27
28 G.R. Amin, A. Emrouznejad and S. Gattoufi, Minor and major consolidations in inverse DEA: Definition and
29
30 determination. Computers & Industrial Engineering. 103 (2017) 193-200.
31
32 G.R. Amin, S. Al-Muharrami and M. Toloo, A combined goal programming and inverse DEA method for
33 target setting in mergers. Expert Systems with applications. 115 (2019) 412-417.
34
35 G.R. Amin and M. Ibn Boamah, A new inverse DEA cost efficiency model for estimating potential merger
36 gains: a case of Canadian banks. Annals of Operations Research. 295 (2020) 21-36.
37
38 L. Chen, Y. Wang, F. Lai and F. Feng, An investment analysis for China's sustainable development based on
39
inverse data envelopment analysis. Journal of cleaner production. 142 (2017) 1638-1649.
40
41 L. Chen and Y.M. Wang, Limitation and optimization of inputs and outputs in the inverse data
42
43 envelopment analysis under variable returns to scale. Expert Systems with Applications. 183 (2021)
44 115344.
45
46 L. Chen, Y. Gao, M.J. Li, Y.M. Wang and L.H. Liao, A new inverse data envelopment analysis approach to
47 achieve China’s road transportation safety objectives. Safety Science. 142 (2021) 105362.
48
49 S. Çakır, Proposing integrated Shannon’s entropy–inverse data envelopment analysis methods for
50
51
resource allocation problem under a fuzzy environment. Engineering Optimization. 49 (2017) 1733-1749.
52
Z. Shiri Daryani, G. Tohidi, B. Daneshian, S. Razavyan and F.H. Lotfi, Inverse DEA in two-stage systems
53
54 based on allocative efficiency. Journal of Intelligent & Fuzzy Systems. 40 (2021) 591-603.
55
56 M. Eyni, G. Tohidi and S. Mehrabeian, Applying inverse DEA and cone constraint to sensitivity analysis of
57 DMUs with undesirable inputs and outputs. Journal of the Operational Research Society. 68 (2017) 34-40.
58
59
60
61
62
63
64
65
1
2
3
4 A. Emrouznejad, G.L. Yang and G.R. Amin, A novel inverse DEA model with application to allocate the CO2
5
6 emissions quota to different regions in Chinese manufacturing industries. Journal of the Operational
7 Research Society. 70 (2019) 1079-1090.
8
9 S. Ghobadi, Merging decision-making units with interval data. RAIRO-Operations Research. 55 (2021)
10 S1605-S1631.
11
12 S. Ghobadi, G.R. Jahanshahloo, F.H. Lotfi and M. Rostami-Malkhalifeh, Dynamic inverse DEA in the
13 presence of fuzzy data. Advances in Environmental Biology. 8 (2014) 139-151.
14
15 M. Ghiyasi, Inverse DEA based on cost and revenue efficiency. Computers & Industrial Engineering. 114
16
17 (2017) 258-263.
18
19 M. Ghiyasi, A. Dehnokhalagy, A scenario-based model for resource allocation with price information.
20 Foundations of Computing and Decision Sciences. 46 (2021) 339-360.
21
22 M. Ghiyasi, On inverse DEA model: The case of variable returns to scale. Computers & Industrial
23 Engineering. 87 (2015) 407-409.
24
25 M. Ghiyasi, Industrial sector environmental planning and energy efficiency of Iranian provinces. Journal of
26
cleaner production. 142 (2017) 2328-2339.
27
28
M. Ghiyasi and N. Zhu, An inverse semi-oriented radial data envelopment analysis measure for dealing
29
30 with negative data. IMA Journal of Management Mathematics. 31 (2020) 505-516.
31
32 G. Zhang and J. Cui, A general inverse DEA model for non-radial DEA. Computers & Industrial Engineering.
33 142 (2020) 106368.
34
35 A. Hassanzadeh, S. Yousefi, R.F. Saen and S.S.S. Hosseininia, How to assess sustainability of countries via
36 inverse data envelopment analysis?. Clean Technologies and Environmental Policy. 20 (2018) 29-40.
37
38 S.S. Hosseininia and R.F. Saen, Developing a novel inverse data envelopment analysis (DEA) model for
39
40
evaluating after‐sales units. Expert Systems. 37 (2020) e12579.
41
42
X. Hu, J. Li, X. Li and J. Cui, A revised inverse data envelopment analysis model based on radial models.
43 Mathematics. 8 (2020) 803.
44
45 G.R. Jahanshahloo, F.H. Lotfi, N. Shoja, G. Tohidi and S. Razavyan, Input estimation and identification of
46 extra inputs in inverse DEA models. Applied Mathematics and Computation. 156 (2004) 427-437.
47
48 M.H. Le, M. Afsharian and H. Ahn, Inverse Frontier-based Benchmarking for Investigating the Efficiency
49 and Achieving the Targets in the Vietnamese Education System. Omega. 103 (2021) 102427.
50
51 S. Lertworasirikul, P. Charnsethikul and S.C. Fang, Inverse data envelopment analysis model to preserve
52
53 relative efficiency values: The case of variable returns to scale. Computers & Industrial Engineering. 61
54 (2011) 1017-1023.
55
56 H. Li, C. Chen, W.D. Cook, J. Zhang and J. Zhu, Two-stage network DEA: Who is the leader?. Omega. 74
57 (2018) 15-19.
58
59
60
61
62
63
64
65
1
2
3
4 Y. Lin, L. Yan and Y.M. Wang, Performance evaluation and investment analysis for container port
5
6 sustainable development in china: An inverse DEA approach. Sustainability. 11 (2019) 4617.
7
8 Y. Lin, Y.M. Wang and H.L. Shi, Mergers and acquisitions matching for performance improvement: a DEA-
9 based approach. Economic research-Ekonomska istraživanja. 33 (2020) 3545-3561.
10
11 D.J. Lim, Inverse DEA with frontier changes for new product target setting. European Journal of
12 Operational Research. 254 (2016) 510-516.
13
14 D. Modhej, M. Sanei, N. Shoja and F. HosseinzadehLotfi, Integrating inverse data envelopment analysis
15 and neural network to preserve relative efficiency values. Journal of Intelligent & Fuzzy Systems. 32 (2017)
16
17 4047-4058.
18
19 K.K. Orisaremi, F.T. Chan and N.S. Chung, Potential reductions in global gas flaring for determining the
20 optimal sizing of gas-to-wire (GTW) process: An inverse DEA approach. Journal of Natural Gas Science and
21 Engineering. 93 (2021) 103995.
22
23 T. Sayar, M. Ghiyasi and J. Fathali, New inverse DEA models for budgeting and planning. RAIRO-Operations
24
Research. 55 (2021) 1933-1948.
25
26
K. Soleimani-Chamkhorami, F. Hosseinzadeh Lotfi, G. Jahanshahloo and M. Rostamy-Malkhalifeh, A
27
28 ranking system based on inverse data envelopment analysis. IMA Journal of Management Mathematics.
29 31 (2020) 367-385.
30
31 K. Soleimani-Chamkhorami, F. Hosseinzadeh Lotfi, G.R. Jahanshahloo and M. Rostamy-Malkhalifeh,
32 Preserving cost and revenue efficiency through inverse data envelopment analysis models. INFOR:
33
34 Information Systems and Operational Research. 58 (2020) 561-578.
35
36 K. Tone, A slacks-based measure of efficiency in data envelopment analysis. European Journal of
37 Operational Research. 130 (2001) 498-509.
38
39 Q. Wei, J. Zhang and X. Zhang, An inverse DEA model for inputs/outputs estimate. European Journal of
40 Operational Research. 121 (2000) 151-163.
41
42 M. Wegener and G.R. Amin, Minimizing greenhouse gas emissions using inverse DEA with an application
43 in oil and gas. Expert Systems with Applications. 122 (2019) 369-375.
44
45 S. Yousefi, R.F. Saen and S.S.S. Hosseininia, Developing an inverse range directional measure model to
46
47 deal with positive and negative values. Management Decision. 57 (2019) 2520-2540.
48
49 E. Zeinodin and S. Ghobadi, Merging decision-making units under inter-temporal dependence. IMA
50 Journal of Management Mathematics. 31 (2020) 139-166.
51
52 B. Zhang, J. Guo, Z. Wen, Z. Li and N. Wang, Ecological evaluation of industrial parks using a comprehensive
53 DEA and inverted-DEA model. Mathematical Problems in Engineering. (2020) 11 pages.
54
55 M. Zhang and J.C. Cui, The extension and integration of the inverse DEA method. Journal of the
56
Operational Research Society. 67 (2016) 1212-1220.
57
58
59
60
61
62
63
64
65
1
2
3
4 X.S. Zhang and J.C. Cui, A project evaluation system in the state economic information system of china an
5
6 operations research practice in public sectors. International Transactions in Operational Research. 6,
7 (1999) 441-452.
8
9 J. Zhang, W. Jin, G.L. Yang, H. Li, Y. Ke and S.P. Philbin, Optimizing regional allocation of CO2 emissions
10 considering output under overall efficiency. Socio-Economic Planning Sciences. (2021) 101012.
11
12 C.H. Wang, Gopal, R. Zionts. S. Use of data envelopment analysis in assess in information technology
13 impact on rm performance. Annals of Operation Research, 73, (1997) 191-213.
14
15
16 M. Kalantary, R. Farzipoor,.Assessing sustainability of supply chains: An inverse network dynamic DEA
17 model. Computer and industrial engineering, 135, (2019), 1224-1238.
18
19
20 S.S. Khiavi, S. Eskandari. The Design of Inverse Network DEA Model for Measuring the Bullwhip Effect in
21
22
Supply Chains with Uncertain Demands. (2021), 14, 93-104.
23
24 Farzipoor Saen, R. and Seyedi Hosseini Nia, S.S., "Evaluating after-sales service units by developing inverse
25 network data envelopment analysis model", Benchmarking: An International Journal, (2020), Vol. 27 No.
26 2, pp. 695-707. https://doi.org/10.1108/BIJ-01-2019-0017.
27
28
29 Majid Kalantary, Reza Farzipoor Saen, Abbas Toloie Eshlaghy. Sustainability Assessment of Supply Chains
30 by Inverse Network Dynamic Data Envelopment Analysis, Scientia Iranica, (2018), 25, 3723-3743.
31
32
Modhej, D., M., Sanei, N., Shoja, F., Hosseinzadeh Lotfi. Integrating inverse data envelopment analysis and
33
34 neural network to preserve relative efficiency values, Journal of Intelligent & Fuzzy Systems, (2017). vol.
35 32, no. 6, pp. 4047-4058.
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65

You might also like