JID:FSS AID:6731 /FLA [m3SC+; v1.201; Prn:23/01/2015; 15:31] P.

1 (1-19)
Available online at www.sciencedirect.com

1
ScienceDirect 1
2 2
3 3
Fuzzy Sets and Systems ••• (••••) •••–•••
4 www.elsevier.com/locate/fss 4
5 5
6 6
7 7
8 A practical approach for estimating weights of interacting criteria 8
9 9
10 from profile sets 10
11 11
12 a,∗ b b 12
Hazel V. Rowley , Arne Geschke , Manfred Lenzen
13 13
14 a UNSW Water Research Centre, The University of New South Wales, UNSW, Sydney, NSW 2052, Australia 14
15 b ISA, School of Physics, A28, The University of Sydney, NSW 2006, Australia 15
16 Received 11 November 2013; received in revised form 1 October 2014; accepted 12 January 2015 16
17 17
18 18
19 19
20 20
21 Abstract 21
22 Aggregating multi-criteria data is an important problem with many applications. Commonly-used additive aggregation methods, 22
23 such as the weighted arithmetic mean, cannot account for the criteria interactions encountered in many practical multi-criteria deci- 23
24 sion problems. The Choquet integral is a suitable aggregation operator in the presence of interacting criteria. It replaces the weight 24
25 vector with a fuzzy measure that models the importance of each subset or coalition of criteria, rather than just the importance of 25
26 individual criteria. However, estimating the fuzzy measures in practice has been problematic. Conventional approaches are cogni- 26
27 tively challenging for decision makers, while more recent approaches suffer from prohibitive data requirements. In this paper, we 27
28
present a formulation for the weights of the Choquet integral that uses principal component analysis to account for criteria interac- 28
29
tion. This novel unsupervised approach to estimating the required fuzzy measures overcomes the limitations of other methods. The 29
approach is applied to two case studies in environmental and sustainability analysis, and the results are compared with those of
30 30
the weighted arithmetic mean. The first case is a triple bottom line analysis of 135 Australian industry sectors evaluated against 11
31 31
criteria, while the second case is an environmental life cycle assessment of 8 alternative biosolids management options evaluated
32 32
against 5 criteria. These examples demonstrate the ability of the proposed approach to account for criteria interaction in these and
33 other decision contexts. 33
34 © 2015 Published by Elsevier B.V. 34
35 35
36 36
Keywords: Multiple criteria evaluation; Non-additive measures; Decision analysis; Multivariate statistics; Environment; Sustainability; Interacting
37 37
criteria; Choquet integral
38 38
39 39
40 40
41
1. Introduction 41
42 42
43
The choice of a set of criteria against which alternatives are evaluated is itself a value judgment in all multi-criteria 43
44
decision methods, not only with regard to which criteria are selected, but also with regard to how the criteria are 44
45
defined, and in how much detail they are represented. In particular, in practical applications of multi-criteria deci- 45
46 46
47 * Corresponding author. Tel.: +61 2 9385 4151; fax: +61 2 9313 8624. 47
48 E-mail addresses: h.rowley@unsw.edu.au (H.V. Rowley), geschke@physics.usyd.edu.au (A. Geschke), m.lenzen@physics.usyd.edu.au 48
49 (M. Lenzen). 49
50 50
http://dx.doi.org/10.1016/j.fss.2015.01.011
51 51
0165-0114/© 2015 Published by Elsevier B.V.
52 52
 JID:FSS AID:6731 /FLA [m3SC+; v1.201; Prn:23/01/2015; 15:31] P.2 (1-19)
2 H.V. Rowley et al. / Fuzzy Sets and Systems ••• (••••) •••–•••

1 sion methods, variables can in general not be constructed without interaction (i.e. dependant relationships among 1
2 criteria). For example, in a multi-criteria environmental systems analysis, the criteria ‘eutrophication potential’ and 2
3 ‘NOx emissions’ convey some common information because eutrophication by NOx is represented in both criteria. 3
4 However, each criterion also conveys independent information: eutrophication by other substances, and the other en- 4
5 vironmental impacts of NOx (e.g. human health impacts), respectively. Careful selection of criteria can often improve 5
6 this situation, but some interaction will generally remain. 6
7 Commonly-employed aggregation procedures such as the arithmetic mean and simple additive weighting are un- 7
8 able to account for criteria interaction. The results of decision analyses with interacting criteria may therefore be 8
9 contended, because issues represented by more than one criterion are likely to have a disproportionately high influence 9
10 on the final outcome [1,2]. If, for example, a triple bottom line (economic, environmental and social sustainability) 10
11 analysis considers more economic criteria than environmental criteria, then those alternatives with strong economic 11
12 performance will be given an automatic advantage under a conventional aggregation scheme. 12
13 During the past three decades, a method has been developed that provides a way of aggregating interacting criteria. 13
14 The Choquet integral [3] is a fuzzy integral which is based on a non-additive measure and can be used as an aggrega- 14
15 tion operator when criteria interact. However, determining weights in practice has been an often revisited topic. The 15
16 conventional approaches are cognitively challenging for decision makers, while more recent approaches suffer from 16
17 a so-called ‘curse of dimensionality’ in terms of their data requirements (Section 2.3). The aim of this paper is to 17
18 provide a novel solution to these problems. It presents a formulation for the weights of the Choquet integral that uses 18
19 principal component analysis (PCA) to account for criteria interaction. 19
20 The following section briefly reviews the mathematics of fuzzy integrals,1 and describes existing approaches 20
21 to estimating the weights of the Choquet integral. After this, the equivalent number of non-interacting criteria is 21
22 introduced, and an intuitive interpretation is given with the help of simple examples. The new measure is then ap- 22
23 plied to two case studies. Finally, conclusions are drawn about the applicability and limitations of the proposed 23
24 method. 24
25 25
26 26
2. Background
27 27
28 28
2.1. The Choquet integral
29 29
30 30
Consider a finite set of decision alternatives X = {a1 , . . . , am } and a finite set of criteria C = {c1 , . . . , cn } against
31 31
which the alternatives are evaluated.
32 32
Each alternative ak ∈ X is associated with a profile of partial scores pk = (p1k , . . . , pnk ) ∈ Rn where, for all i =
33 33
34
1, . . . , n, pik is the evaluation of alternative ak with respect to criterion ci , with pik ∈ Pi ⊆ R [4,5]. 34
35 Assume further that the evaluations are defined according to the same interval scale (Pi = P ∀i) [5].2 For example, 35
36 it is common to scale the evaluations to the range [0, 1] by a linear transformation. This assumption implies that all 36
37 the attributes are commensurate [8]. 37
38 Within many multi-criteria decision methods that allow criteria interaction and interdependence, the problem is 38
39 to compute a global score M(p k ) for each alternative ak ∈ X, which takes into account all of the partial scores 39
40 pk = (p1k , . . . , pnk ) and the weights representing the importance of the criteria. An aggregation operator M : Rn → R 40
41 is used for this purpose [8–10]. 41
42 One of the most commonly used aggregation operators is the weighted arithmetic mean [4,8,9]. This operator gives 42
43 the global score M(p) associated with the profile p = (p1 , . . . , pn ) as 43
44 44

n
45
Mω (p) = ω i pi (1) 45
46 46
i=1
47 n 47
48 where, for all i = 1, . . . , n, ωi is the weight of criterion ci with ωi ≥ 0 and i=1 ωi = 1 [8]. 48
49 49
50 1 The following review is non-exhaustive and deals only with the aspects relevant to this problem. 50
51 2 The scale transformation mentioned here is not trivial. However, this topic is not within the scope of this research article. The interested reader 51
52 may refer to [6] or [7]. 52
 JID:FSS AID:6731 /FLA [m3SC+; v1.201; Prn:23/01/2015; 15:31] P.3 (1-19)
H.V. Rowley et al. / Fuzzy Sets and Systems ••• (••••) •••–••• 3

1 However, aggregation operators of this form are not able to represent interaction among criteria [9]. They are 1
2 characterised by an independence axiom which is rarely verified in practice [8]. 2
3 An alternative approach has recently been proposed which allows the modelling of interacting criteria [9,11]. It is 3
4 well-known that a set of n elements (in this case, criteria) has 2n subsets. A fuzzy measure [12] or capacity [3] on the 4
5 set C = {c1 , . . . , cn } is then a monotone set function μ : 2n → R which satisfies μ(∅) = 0, μ(C) = 1 where ∅ denotes 5
6 the empty set; and μ(S) ≤ μ(T ) for all S ⊆ T ⊆ C [8,13]. Replacing the weight vector ω with the fuzzy measure 6
7 μ enables the modelling of the importance of each subset or coalition of criteria, rather than just the importance of 7
8 individual criteria [8,9,11,13]. 8
9 One suitable aggregation operator that generalises the weighted arithmetic mean when criteria interact is the dis- 9
10 crete Choquet integral with respect to the fuzzy measure μ [3,4,9]. The Choquet integral has been well-researched 10
11 in the context of MCDA due to its ability to account for any form of criteria interaction [5], of which there are 11
12 three distinct types: correlation, substitutiveness/complementarity, and preferential dependence [9]. Other properties 12
13 and characterisations of the Choquet integral can be found in one of the many comprehensive references on the 13
14 subject, such as those by Grabisch [11], Marichal [9], Grabisch & colleagues [10], Grabisch & Labreuche [13], Den- 14
15 neberg [14], or Murofushi & Sugeno [15]. 15
16 The Choquet integral of p k ∈ Rn with respect to μ is defined by 16
17 17
18   
n
  18
19 C μ
p k
= k
p(i) − p(i−1)
k
μ(A(i) ) (2) 19
20 i=1 20
21 21
where we recall that =
pk (p1k , . . . , pnk )
is the profile of normalised partial scores of alternative ak against the n
22 22
criteria; and where the subscript (i) indicates a permutation of the indices i = 1, . . . , n so that 0 ≤ p(1) k ≤ · · · ≤ pk
(n)
23 23
and p(0) k = 0, i.e. p k is sorted in ascending order; and where A = {c , . . . , c } ⊆ C denotes the subset of criteria
24 (i) (i) (n) 24
25 (i), . . . , (n), i.e. A(i) is the subset of C that includes only those criteria for which the normalised score of alternative ak 25
26 is better than or equal to its normalised score on ci [9]. 26
27 Note that if the fuzzy measure μ is additive; that is, μ(S ∪ T ) = μ(S) + μ(T ) whenever S ∩ T = ∅ (i.e. the 27
28 criteria are non-interacting), then the Choquet integral coincides with the weighted arithmetic mean (discrete Lebesgue 28
29 integral) [9]. 29
30 30
31 2.2. Criteria importance and interaction 31
32 32
33 The Choquet integral formulation naturally leads to definitions of criteria importance and interaction [16] which 33
34 allow “a better comprehension of the interaction phenomena modelled by the underlying capacity” [8, p. 132]. Such 34
35 questions arise because the importance of a criterion ci ∈ C is not only determined by μ(ci ), but also by all μ(S) 35
36 such that ci ∈ S [5]. The coefficient of importance proposed by Shapley [17], which is a fundamental concept in game 36
37 theory, takes this aspect into account [5]. The Shapley importance index (Shapley value) S μ of criterion ci with respect 37
38 to μ is defined by 38
39 39
40
 (|C| − |A| − 1)!|A|!     40
S (ci ) =
μ
μ A ∪ {ci } − μ(A) (3)
41 |C|! 41
A⊆C\{ci }
42 42

43 where A⊆C\{ci } means the sum over the subset A of C, where A is any set of criteria which does not contain ci (i.e. 43
44 the sum extends over all subsets A of C not containing criterion ci ); |C| is the cardinal of C (i.e. the total number of 44
45 criteria, equal to n); and |A| is the cardinal
 of A (i.e. the number of criteria in the subset A) [5]. The combinatoric 45
46 factor normalises the function so that ni=1 S μ (ci ) = 1. The Shapley value “can be interpreted as a weighted average 46
47 value of the marginal contribution of [criterion ci ] alone in all coalitions” [5, pp. 644]. 47
48 As well as criteria importance, it is also interesting to consider the magnitude of the interaction between criteria. 48
49 For any pair of criteria {ci , cj } ∈ C, the difference between μ(ci , cj ) and μ(ci ) + μ(cj ) reflects the degree of inter- 49
50 action between ci and cj . If μ(ci , cj ) > μ(ci ) + μ(cj ), then there is positive (synergistic) interaction. If μ(ci , cj ) < 50
51 μ(ci ) + μ(cj ), then there is negative interaction (redundancy/criteria overlap). If μ(ci , cj ) = μ(ci ) + μ(cj ), there is 51
52 no interaction [5]. 52
 JID:FSS AID:6731 /FLA [m3SC+; v1.201; Prn:23/01/2015; 15:31] P.4 (1-19)
4 H.V. Rowley et al. / Fuzzy Sets and Systems ••• (••••) •••–•••

1 Murofushi & Soneda [18] have proposed a definition of a coefficient of interaction which draws on concepts similar 1
2 to those underlying the Shapley value [5]. Let P(C) be the power set of C; that is, the set of all subsets D ⊆ C. The 2
3 Shapley interaction index I μ : P(C) → R between two criteria ci and cj is defined by [9] 3
4 4
5  (|C| − |A| − 2)!|A|!         5
I μ (ci , cj ) = μ A ∪ {ci , cj } − μ A ∪ {ci } − μ A ∪ {cj } + μ(A) (4)
6
(|C| − 1)! 6
7 A⊆C\{ci ,cj } 7
8 8
and can be interpreted as a measure of the average marginal interaction between two criteria [4]. The index I μ is
9 9
negative in case of redundancy (substitutive criteria), positive in case of synergy (complementary criteria) [9], and is
10 10
not defined for the case i = j . This index has also been further generalised [19,20] for coalitions of more than two
11 11
12
criteria [4], but this generalisation is not explored here. 12
13 13
14 2.3. Estimating the fuzzy measures μ 14
15 15
16 Just as the use of the weighted arithmetic mean as an aggregation operator first requires the definition of the weight 16
17 vector ω, so too the use of the Choquet integral as an aggregation operator first requires the definition of the fuzzy 17
18 measure μ [8]. In the context of MCDA, ω or μ can, in principle, be determined based on the initial preferences [21] 18
19 expressed by the decision maker – i.e. ideas about the importance of the attributes, relationships among them, etc. [8]. 19
20 20
Such an approach, in which the result of the aggregation takes into account a decision maker’s point of view, may be
21 21
referred to as a supervised approach [8].
22 22
While the weighted arithmetic mean requires only n criteria weights, the evaluation of the fuzzy measure requires
23
2n − 2 coefficients [4]. This is because the alternatives ak ∈ X generally exhibit different rank orders with regard to 23
24
their criteria scores, requiring different subsets A(i) each time, and there are 2n possible subsets of C (including C 24
25
and ∅). 25
26
In principle, these measures have to be elicited from the decision-maker [5,22], but for n  6 this daunting task 26
27 27
is likely to exceed human cognitive abilities. In some cases symmetry considerations, veto and pass relationships can
28 28
reduce this number, but for sufficiently large n, the number of coefficients will still be prohibitively high. In general,
29 29
“unlike for the weighted arithmetic means, it is rather unrealistic to assume that [these coefficients] can be provided
30 30
by the decision maker especially when the number of criteria is large” [4, p. 742].
31 31
In view of this, Kojadinovic [4,8] has proposed an unsupervised approach as an alternative to the supervised
32 32
approaches “when the prior knowledge they rely on cannot be provided” [4,8, p. 129]. In an unsupervised approach,
33 33
the concept of criteria importance is replaced by the alternative concept of how much unique information is measured
34 34
by each criterion – i.e. their information content – which can be estimated from the evaluations of the alternatives
35 35
against the criteria [8]. It may therefore be useful in the absence of initial preferences.
36 36
Kojadinovic [4,8] takes a probabilistic view of identifying fuzzy measures in an unsupervised way: instead of
37 37
38
eliciting information from a decision-maker, a sample of m profiles p Ai = {pjAki } j = i, . . . , n on subsets Ai of the 38
k = 1, . . . , m
39
criteria set C = {c1 , . . . , cn } is used to estimate the coefficients μ(Ai ) of the Choquet integral. Kojadinovic suggests 39
40
using the concept of mutual information contained in a subset Ai ⊆ C, and defines 40
41 41
42 h(Ai ) 42
43 μ(Ai ) = (5) 43
h(C)
44 44
45 where h(Ai ) is the entropy of any of the (n − i + 1)! vectors composed of the probability distributions of the obser- 45
46 vations p ∈ Ai . However, Kojadinovic [4, p. 751] notes the method’s “curse of dimensionality” because “to obtain 46
47 estimates of the global scores of a given accuracy, the number of profiles has to grow almost exponentially with the 47
48 number of criteria” due to its reliance on probability distributions. In many practical applications, a suitably large set 48
49 of profiles would not be available. 49
50 In view of the cognitive difficulties of the supervised approaches, and the very large data requirements of Kojadi- 50
51 novic’s unsupervised approach, the following section describes an alternative unsupervised approach to the estimation 51
52 of the fuzzy measures without the prohibitive requirements of previous approaches. 52
 JID:FSS AID:6731 /FLA [m3SC+; v1.201; Prn:23/01/2015; 15:31] P.5 (1-19)
H.V. Rowley et al. / Fuzzy Sets and Systems ••• (••••) •••–••• 5

1 3. A novel approach to estimating the fuzzy measures μ 1

2 2
3 3.1. Equivalent number of non-interacting criteria 3
4 4
5
Let rij be an element of the absolute |R| of a correlation matrix R describing the correlation between criteria ci 5
6
and cj with i, j ∈ {1, . . . , n}, calculated from a sufficient number of profiles {pik }k=1,...,m and {pjk }k=1,...,m . If the 6
7 7
8
alternatives are known a priori to be unequally important,3 weighted correlation coefficients rij can be computed 8
9
[23,24] which take into account the importance yk of each alternative within the total importance Σl yl : 9
10 m myk k 10
k=1 l yl (pi − pi )(pj − pj )
 k
11
rij =   (6) 11
m  myk myk
12
k=1 y
(pik − pi )2 m k=1

y
(pjk − pj )2 12
l l l l
13 13
14 where 14
15 15

m
yk
16
pi =  pik (7) 16
17 l yl
k=1
17
18 18
19 is a weighted average over scores pik . Otherwise, unweighted Pearson correlation coefficients can be used. 19
20 In the following, the influence of criteria interaction will be illustrated for n = 5. Absolutes are taken so that criteria 20
21 are also redundant if they are negatively correlated: for example ‘employment’ and ‘unemployment’ are really only 21
22 one criterion, and not two. Accordingly, all correlation matrices will be symmetrical, with rij = rj i > 0. If all five 22
23 criteria are completely uncorrelated, R1 is equal to the 5-by-5 identity matrix I . Let J ∗ denote the equivalent number 23
24 of non-interacting criteria, as a measure of independence amongst the pik . Then in the completely uncorrelated case 24
25 R1 = I, we know n = J ∗ = 5. 25
26 Assume now that criteria 4 and 5 are almost completely correlated, that is 26
27 ⎛ ⎞ 27
28 1 0 0 0 0 28
⎜0 1 0 0 0 ⎟
29 ⎜ ⎟ 29
⎜
R2 = ⎜ 0 0 1 0 0 ⎟ (8)
30 ⎟ 30
⎝0 0 0 1 0.99 ⎠
31 31
32 0 0 0 0.99 1 32
33
In this case, n = 5, but we expect J∗ to be in the vicinity of 4, because c4 and c5 are really only one criterion. In the 33
34 34
extreme case
35
⎛ ⎞ 35
36 1 0.99 0.99 0.99 0.99 36
⎜ 0.99 1 0.99 0.99 0.99 ⎟
37 ⎜ ⎟ 37
38 R3 = ⎜⎜ 0.99 0.99 1 0.99 0.99 ⎟
⎟ (9) 38
39 ⎝ 0.99 0.99 0.99 1 0.99 ⎠ 39
40 0.99 0.99 0.99 0.99 1 40
41 41
all criteria are almost completely correlated; in other words there is really only one criterion and we expect ≈ 1. J∗
42 42
Using the non-interacting equivalent, and considering the definitions in the previous section, a fuzzy measure μ(Ai )
43 43
for any criteria subset Ai ⊆ C can be defined as
44 44
45 J ∗ (Ai ) 45
46 μ(Ai ) = ∗ (10) 46
J (C)
47 47
48 A novel way to compute J ∗ for use in the Choquet integral formulation, using principal component analysis, is now 48
49 described. 49
50 50
51 3 For example, as indicated by their relative market share or, as occurs in the first case study below, their respective contributions to the national 51
52 economy. 52
 JID:FSS AID:6731 /FLA [m3SC+; v1.201; Prn:23/01/2015; 15:31] P.6 (1-19)
6 H.V. Rowley et al. / Fuzzy Sets and Systems ••• (••••) •••–•••

1 3.2. Principal component analysis 1

2 2
3 Principal component analysis (PCA) is a multivariate data analysis method that can be used to interpret multivariate 3
4 datasets by reducing their dimensionality. Multivariate data analysis is increasingly used in many varied fields for 4
5 handling large amounts of data – from survey responses in the social sciences, to spectral data in analytical chemistry. 5
6 However, according to our knowledge, it has not been used in the way proposed here. 6
7 The principal components (PCs) of a set of evaluations on any criteria subset A ⊆ C (including C itself) are 7
8 calculated by finding an orthonormal basis such that each successive PC captures maximal remaining variance present 8
9 in the evaluations, while being independent of the previously-determined PCs. The vector z of PCs, where zk is the 9
10 PC related to the k th element of the basis, is thus defined by: 10
11 11
12 z = Φ TA (11) 12
13 13
14
where Φ is a square orthogonal matrix with its k th column being the eigenvector corresponding to the k th -largest 14
15
eigenvalue λk of the correlation matrix R (or covariance matrix) of the evaluations [25]. 15
16
Having derived these PCs, one can determine how much variance is captured separately by each one [25]. 16
17
PCA is in some ways similar to factor analysis, although the latter is founded on the existence of a model relating 17
18
the observed and latent variables (PCs) [25]. PCA is therefore preferred over factor analysis for the present application, 18
19
which is concerned only with PC identification. 19
20
Interested readers are referred to Shlens [26] for a conceptually clear introduction to PCA, or to Smith [27] for the 20
21
introductory mathematics. On the details of the method’s derivation and practical application, Jolliffe [25,28] is an 21
22
authoritative voice. 22
23 23
24 3.3. Proposed method 24
25 25
26 The principles of PCA can be used to compute the non-interacting equivalent, J ∗ , for each case of x where x is 26
27 either the whole set C or a relevant subset A. 27
28 Before implementing PCA, some data manipulations are required. First, to avoid undue influence of criteria with 28
29 large variances,4 the profiles are scaled to unit variance by dividing each score by the standard deviation of the 29
30 criterion it expresses. Second, to prevent the first PC from simply representing the position of the centroid of the 30
31 profiles relative to the origin, the data are also mean-centred by subtracting the mean of each scaled criterion from 31
32 each score on that criterion [25,29].5,6 32
33 A correlation matrix for the criteria is then computed based on the profiles. The eigenvectors of the correlation 33
34 matrix R are computed and sorted in descending order. The proportion of variance Vk captured by the k th PC can be 34
35 computed by dividing the k th -largest eigenvalue λk by the sum of the eigenvalues [31] 35
36 36
λk
37 Vk =  (12) 37
38 k λk 38
39 The next step is to compute J ∗ , the equivalent number of non-interacting criteria, based on this PCA. The various 39
40 rules [25] for choosing a sufficient number of PCs all return J ∗ ∈ N, to suit conventional PCA applications. For 40
41 example, it is common to choose all PCs corresponding to eigenvalues λk ≥ 1. The application described here allows 41
42 the flexibility to choose J ∗ ∈ R and therefore a novel method of ‘selection’ uniquely appropriate to this situation can 42
43 be used. J ∗ can be defined as 43
44
   44
45 J ∗ (Ai ) = λk + {λk |λk ≥ 1} (13) 45
46 k;λk <1 46
47 47
48 48
4 For example, a criterion measured on a scale of thousands of kilogrammes will arbitrarily have a much greater influence on the model than one
49 49
represented by tens of kilometres, while expressing the same data in tonnes and metres respectively would have the opposite effect.
50 5 An excellent graphical representation of these procedures is Fig. 3.7 of Eriksson & colleagues [29]. 50
51 6 The combination of unit variance scaling and mean-centring is sometimes known as auto-scaling, although this term is avoided here as it is also 51
52 sometimes used to describe unit variance scaling without mean-centring [30] and is therefore ambiguous. 52
 JID:FSS AID:6731 /FLA [m3SC+; v1.201; Prn:23/01/2015; 15:31] P.7 (1-19)
H.V. Rowley et al. / Fuzzy Sets and Systems ••• (••••) •••–••• 7

1 where | | denotes the total number of elements in a finite set. The λk are the eigenvalues of the matrix that is made up 1
2 by rows and columns of the covariance matrix that are associated with the criteria in Ai . 2
3 This overcomes the need to impose an arbitrary cut-off and thereby extracts the maximum possible information 3
4 from each dataset. 4
5 In the following it will be shown that 5
6
J ∗ (Ai ) 6
7 μ(Ai ) = 7
8
J ∗ (C) 8
9 is in fact a fuzzy measure for the sets Ai . 9
10 10
J ∗ (∅)
11 1. μ(∅) =J ∗ (C) = 0 11
∗
12 2. μ(C) = JJ ∗ (C)
(C) = 1
12
13 3. The last property that needs to be shown is the monotony of μ, which is equivalent to Ai ⊆ Aj ⇒ μ(Ai ) ≤ μ(Aj ). 13
14 It is sufficient to show that J ∗ (Ai ) ≤ J ∗ (Aj ) since J ∗ (C) is a positive constant. Without loss of generality, assume 14
15 that 15
16 16
17 AI = {c1 , c2 , . . . , ci } 17
18 18
and
19 19
20 AJ = {c1 , c2 , . . . , ci , ci+1 } 20
21 21
22 Let RI be the matrix consisting of the rows and columns of the covariance matrix that are associated with the cri- 22
23 teria in AI , and let λIl,l=1,...,i be the eigenvalues of the matrix of RI . RI is positive semi-definite, and therefore all 23
24 eigenvalues are real and positive. Sort the eigenvalues in increasing order such that 24
25 25
λI1 ≤ λI2 ≤ · · · ≤ λIi ,
26 26
27 and assume that the first M eigenvalues are smaller than one. Then 27
28 28

M
29 J ∗ (AI ) = λIl + (i − M + 1) (14) 29
30 30
l=1
31 31
Let RJ be the matrix consisting of the rows and columns of the covariance matrix that are associated with the criteria
32 32
in AJ , and let λIk,k=1,...,i,i+1 be the eigenvalues of the matrix RJ . Then RI can be obtained from RJ from by removing
33 33
the row and column associated with criteria ci+1 . RJ is also positive semi-definite, and according to the Cauchy
34 34
Interlace theorem (for example see [32]), the eigenvalues of RI and RJ can be arranged as follows.
35 35
36 λJ1 ≤ λI1 ≤ λJ2 ≤ λI2 ≤ · · · ≤ λIM ≤ λJM+1 ≤ λIM+1 · · · ≤ λIi ≤ λJi+1 (15) 36
37 37
Hence, the largest (i − M + 1) eigenvalues of RJ (indices k = M + 2, . . . , i +
1) areI larger than one. The sum of the
38 38
remaining eigenvalues of RJ (indices k = 1, . . . , M + 1) larger than the sum M l=1 λl . Hence
39 39
40 
M
  
M+1 
M 40
41 J ∗ (AJ ) = λJk + min λJM , 1 + (i − M + 1) ≥ λJk + (i − M + 1) ≥ λIl + (i − M + 1) = J ∗ (AI ) 41
42 k=1 k=1 l=1 42
43 Therefore, μ is monotonous. 43
44 To illustrate (Table 1), these operations are performed on the example correlation matrices presented above. Anal- 44
45 ysis of R2 introduced in Eq. (8) gives the equivalent number of non-interacting criteria to be J ∗ = 4.01 (i.e. J ∗ ≈ 4 as 45
46 expected). Analysis of R1 using the same method gives J ∗ = 5 (as expected). For R3 it gives J ∗ = 1.04 (i.e. J ∗ ≈ 1 46
47 as expected). 47
48 In the following section, these fuzzy measures and Choquet integrals are applied to a recent Australian case study. 48
49 In this application we interpret and validate our modified Choquet integral as a preference degree. This is because 49
50 the only difference between a conventional Choquet integral using entropy metrics and our eigenvalue-based Choquet 50
51 formulation is the modified form of the aggregation method, i.e. the fuzzy measure. A change in the aggregation 51
52 method does not affect the interpretation and validation of the Choquet integral as a preference degree. 52
 JID:FSS AID:6731 /FLA [m3SC+; v1.201; Prn:23/01/2015; 15:31] P.8 (1-19)
8 H.V. Rowley et al. / Fuzzy Sets and Systems ••• (••••) •••–•••

1 Table 1 1
2 PCA models for the example correlation matrices R1 , equal to the identity matrix; R2 introduced in Eq. (8); and R3 introduced in Eq. (9). 2
3 PC # R1 , the identity matrix R2 , Eq. (8) R3 , Eq. (9) 3
4 Eig VE CVE Eig VE CVE Eig VE CVE 4
5 5
1 1 20% 20% 1.99 39.8% 39.8% 4.96 99.2% 99.2%
6 2 1 20% 40% 1 20% 59.8% 0.01 0.2% 99.4% 6
7 3 1 20% 60% 1 20% 79.8% 0.01 0.2% 99.6% 7
8 4 1 20% 80% 1 20% 99.8% 0.01 0.2% 99.8% 8
9
5 1 20% 100% 0.01 0.2% 100% 0.01 0.2% 100% 9
10 Eig = Eigenvalue; VE = Variance Explained; CVE = Cumulative Variance Explained. 10
11 11
12 12
13 13
4. Application to a triple bottom line assessment of industry sectors
14 14
15 15
16 4.1. The original study 16
17 17
18 18
The report Balancing Act [33] presents a triple bottom line account for the Australian economy. It presents the
19 19
performance, against eleven criteria, of each of the 135 industry sectors defined by the Australian Bureau of Statistics’
20 20
Input–Output Product Classification. Together, these sectors constitute the entire Australian economy.
21 21
The eleven criteria include positive ones (+, ‘more is better’) and negative ones (–, ‘less is better’); they are [33]:
22 22
23 23
1. Primary energy consumption: the combustion of non-renewable fossil fuels (MJ, –);
24 24
2. Greenhouse gas emissions: the combined effect of CO2 , CH4 , N2 O, SF6 , hydrofluorocarbons, perfluorocar-
25 25
bons and other gases in the atmosphere, expressed in kilogrammes of an equivalent amount of carbon dioxide
26 26
(kg CO2 -e, –);
27 27
28
3. Water use: the consumption of self-extracted water (water from rivers, lakes and aquifers, mainly extracted by 28
29
farmers for irrigation) and mains water (L, –); 29
30
4. Land disturbance: land use in units of hectares (ha, –) weighted by a factor reflecting land condition or intensity 30
31
of transformation; 31
32
5. Gross operating surplus: consists of operating profits and consumption of fixed capital (A$, +), calculated as the 32
33
residual of an industry’s total inputs, after subtracting all intermediate inputs, compensation of employees, and 33
34
net taxes and subsidies; 34
35 6. Exports: Australian production of primary commodities that are destined for demand outside Australia (A$, +); 35
36 7. Imports: the value of goods and services purchased from foreign economies (A$, –); 36
37 8. Employment: full-time-equivalent employment-years (emp-y, +) of employees, including employers, own ac- 37
38 count workers and contributing family workers; 38
39 9. Income: compensation of employees (A$, +), comprising wages and salaries, and employers’ social contributions; 39
40 10. Government revenue: consists of taxes less subsidies on products for intermediate demand, net taxes on produc- 40
41 tion, and net taxes on products for final demand (A$, +); 41
42 11. Upstream linkage: power of dispersion of a sector’s purchases, calculated as the increase in the activity of the 42
43 whole economy as a result of a unit change in the final demand of the respective sector (dimensionless, +). 43
44 44
45 These 11 criteria were chosen by the Australian Government to reflect nation-wide issues in the context of industry 45
46 sector evaluations. However, the way the criteria were constructed was influenced by what data were available. At 46
47 the beginning of the study, no attempt was made to construct non-interacting criteria. Since the study was launched, 47
48 national bodies have become interested in distilling the data even further to allow benchmarking of the sectors. Using 48
49 a simple arithmetic mean in this distillation would certainly lead to objections by representatives of some sectors. 49
50 For example, sectors that do not perform well on income and employment could object that both criteria were largely 50
51 telling the same story and their sector was disadvantaged by an aspect of the study’s architecture that they might regard 51
52 as arbitrary. 52
 JID:FSS AID:6731 /FLA [m3SC+; v1.201; Prn:23/01/2015; 15:31] P.9 (1-19)
H.V. Rowley et al. / Fuzzy Sets and Systems ••• (••••) •••–••• 9

1 Table 2 1
2 Criteria correlation in the Australian economy, evaluated according to Eq. (6). Policy objectives are compatible if values are positive, and vice-versa. 2
3 Gross Exports Imports Employment Income Government GHG Water Land Primary Upstream 3
4 operating revenue emissions use disturbance energy linkage 4
5 surplus 5
6 Gross operating 1 0.04 0.33 −0.68 −0.88 −0.15 −0.05 −0.08 −0.09 −0.07 0.19 6
7 surplus 7
Exports 1 −0.28 −0.20 −0.25 0.01 −0.13 −0.05 −0.24 −0.27 −0.37
8 8
Imports 1 0.05 0.12 −0.38 0.05 −0.01 −0.03 0.13 0.40
9 9
Employment 1 0.78 −0.05 −0.03 −0.08 −0.16 0.16 −0.28
10 Income 1 −0.02 0.08 0.09 0.10 0.14 −0.01 10
11 Government 1 0.03 0.02 0.04 0.00 −0.06 11
12 revenue 12
GHG emissions 1 0.09 0.40 0.23 0.14
13 13
Water use 1 0.15 0.01 0.12
14 14
Land disturbance 1 −0.02 0.20
15 Primary energy 1 0.16 15
16 Upstream linkage 1 16
17 17
18 18
4.2. Method
19 19
20 20
Let p̃ik
be the partial score of sector ak expressed in physical units of criterion ci . The normalised score pik of each
21 21
sector against each criterion was calculated as
22 22
23 αi p̃ik − minm {αi p̃im } 23
pik = (16)
24
maxm {αi p̃im } − minm {αi p̃im } 24
25 25
26 where αi is the quality of the criterion ci (positive: αi = +1, negative: αi = −1), so that a value of pik = 0 represents 26
27 economy-wide worst practice, and pik = 1 represents economy-wide best practice. 27
28 For each computation of J ∗ , the data were pre-treated using unit variance scaling and mean-centring as described in 28
29 the previous section. Throughout the computation, weighted correlation matrices as per Eq. (6) were used to account 29
30 for each sector’s final demand yk within the total final demand of the economy. 30
31 31
32 4.3. Results and discussion 32
33 33
34 4.3.1. Correlations 34
35 The multi-criteria data on Australia’s industry sectors can be correlated, not only to find out (partly) redundant 35
36 criteria, but also to identify which policy objectives are compatible. The following correlations and trade-offs can be 36
37 established (Table 2): 37
38 38
39 • Employment and income are highly positively correlated, as expected, since the criteria differ only through vary- 39
40 ing average salaries across the sectors; both policy objectives are highly compatible; 40
41 • The objectives of increasing surplus and employment/income are incompatible: productive income can benefit 41
42 either employees or investors, but not both; 42
43 • Sectors with high export earnings perform poorly on other criteria except government revenue: this is clear as 43
44 Australia’s exports are mainly primary commodities such as minerals, metals and livestock which rely on natural 44
45 resources and fuels rather than on people or inputs from the rest of the economy; at the same time, export taxes 45
46 flow into government funds; 46
47 • Greenhouse gas emissions are more correlated with land disturbance than with energy use; this reflects that Aus- 47
48 tralia’s greenhouse gas inventory exhibits a high portion of non-combustion greenhouse gas sources such as CH4 48
49 from enteric fermentation in sheep and cattle, or CO2 from land use changes; nevertheless, simultaneous improve- 49
50 ments to energy use, greenhouse gas emissions and land disturbance are viable due to their correlations; 50
51 • Increasing employment and reducing energy use are compatible objectives, as shown by the correlation in those 51
52 criteria (compare [34]); 52
 JID:FSS AID:6731 /FLA [m3SC+; v1.201; Prn:23/01/2015; 15:31] P.10 (1-19)
10 H.V. Rowley et al. / Fuzzy Sets and Systems ••• (••••) •••–•••

1 Table 3 1
2 PCA model for the Balancing Act dataset. 2
3 PC # Eigenvalue Variance explained Cumulative variance 3
4 explained 4
5 1 2.69 24.4% 24.4% 5
6 2 2.13 19.3% 43.7% 6
7 3 1.54 14.0% 57.7% 7
4 1.05 9.6% 67.3%
8 8
5 0.96 8.7% 76.0%
9 9
6 0.88 8.0% 84.0%
10 7 0.61 5.6% 89.6% 10
11 8 0.55 5.0% 94.6% 11
12 9 0.42 3.8% 98.3% 12
10 0.17 1.6% 99.9%
13 13
11 0.01 0.1% 100%
14 14
15 15
16 • Sectors with low imports also have strong upstream linkages, since their operating inputs tend to be sourced from 16
17 the domestic economy. 17
18 18
19 Further insight into the structure of the data can be gained from the PCA model (Table 3) which shows that only 8 19
20 PCs are required to capture 90% of the variance in the dataset. 20
21 21
22 4.3.2. Importance indices (Shapley values) 22
23 The Shapley values (Table 4) represent the relative importance of each individual criterion in relation to the criteria 23
24 set. Because of their multiple and high mutual correlation, gross operating surplus, income and employment are of 24
25 low importance, compared with government revenue, greenhouse gas emissions and water use, which are relatively 25
26 independent and therefore important criteria. These important criteria would be underemphasised in an aggregation 26
27 scheme that does not account for criteria interaction. 27
28 28
29 4.3.3. Interaction indices 29
30 Some interaction indices (Table 4) are negative, indicating mutual redundancy, while others are positive, indicat- 30
31 ing synergism. Once again, the high redundancy between gross operating surplus, income and employment can be 31
32 observed. Some redundancy between imports and government revenue can also be seen. 32
33 33
34 4.3.4. Choquet integrals 34
35 We apply Choquet integrals according to Eq. (2) to profiles as in Eq. (13) in order to calculate global scores, and 35
36 interpret these global scores as a preference degree as justified in Section 3.3. In other words, the global scores show 36
37 whether on average over all (effectively non-interacting) criteria, a sector is performing better or worse (Table 5). 37
38 Some sectors feature low p values in terms of some of the criteria (rice on water; forestry on greenhouse gas 38
39 emissions; electricity supply on energy; beef cattle on land). These sectors are located in the lower tercile of Table 5. 39
40 Sectors in the highest tercile either show balanced outcomes on most or all criteria (e.g. beer and malt), or compensate 40
41 for a low p on one criterion with a high p on another (e.g. education, which naturally performs below-average on gross 41
42 operating surplus, government revenue and exports, but performs near best-practice on all environmental criteria and 42
43 on income). 43
44 Comparing the Choquet integrals with the arithmetic means (Fig. 1) shows that the sectors gaining the most ad- 44
45 vantage from the Choquet integral (i.e. those that were most disadvantaged by the arithmetic mean) are the sectors 45
46 with their best normalised scores against criteria that are relatively independent, with those criteria having proportion- 46
47 ally more influence on the Choquet aggregation than on the arithmetic mean. The sectors that had an arguably unfair 47
48 advantage under the arithmetic mean are those that perform relatively well against criteria that were telling related 48
49 stories. 49
50 Fig. 1 shows that both situations occurred regardless of sectors’ overall ranking: arithmetic mean equally favours 50
51 the personal services (ranked 11) and forestry (35) sectors while the Choquet integral equally favours the education (1) 51
52 and other petroleum & coal products (132) sectors (measured as an absolute change in aggregated score). The differ- 52
52
51
50
49
48
47
46
45
44
43
42
41
40
39
38
37
36
35
34
33
32
31
30
29
28
27
26
25
24
23
22
21
20
19
18
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1

JID:FSS AID:6731 /FLA

Table 4

H.V. Rowley et al. / Fuzzy Sets and Systems ••• (••••) •••–•••
Shapley values (Eq. (3)) and interaction indices (Eq. (4)) for the eleven criteria in the Balancing Act.
Shapley Interaction indices
values Gross Exports Imports Employment Income Government GHG Water Land Primary Upstream
operating revenue emissions use disturbance energy linkage
surplus
Gross operating 0.057 – 1.98 × 10−3 −1.37 × 10−2 −2.96 × 10−2 −6.11 × 10−2 −3.42 × 10−3 −5.67 × 10−4 −1.74 × 10−3 −1.48 × 10−3 1.58 × 10−5 −3.42 × 10−3
surplus
Exports 0.095 – −9.76 × 10−3 −4.77 × 10−3 −5.87 × 10−3 2.53 × 10−3 2.93 × 10−3 9.03 × 10−4 −7.99 × 10−3 −1.11 × 10−2 −1.70 × 10−2
Imports 0.085 – 1.52 × 10−3 −1.79 × 10−3 −2.99 × 10−2 −4.60 × 10−4 7.69 × 10−5 −1.19 × 10−4 −2.02 × 10−3 −1.88 × 10−2
Employment 0.062 – −4.85 × 10−2 −1.13 × 10−3 4.96 × 10−4 −1.78 × 10−3 −4.92 × 10−3 −4.35 × 10−3 −1.09 × 10−2
Income 0.057 – −2.75 × 10−4 −4.85 × 10−4 −1.88 × 10−3 −1.24 × 10−3 −3.14 × 10−4 3.37 × 10−3
Government 0.113 – −3.72 × 10−4 −4.48 × 10−4 8.12 × 10−4 6.30 × 10−4 1.91 × 10−3
revenue
GHG emissions 0.106 – 1.4 × 10−4 −2.58 × 10−2 −9.84 × 10−3 7.52 × 10−4
Water use 0.122 – −7.01 × 10−3 1.36 × 10−4 −2.78 × 10−3

[m3SC+; v1.201; Prn:23/01/2015; 15:31] P.11 (1-19)

Land disturbance 0.100 – 4.73 × 10−3 −4.17 × 10−3
Primary energy 0.112 – −2.82 × 10−3
Upstream linkage 0.091 –

11
52
51
50
49
48
47
46
45
44
43
42
41
40
39
38
37
36
35
34
33
32
31
30
29
28
27
26
25
24
23
22
21
20
19
18
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1
 JID:FSS AID:6731 /FLA [m3SC+; v1.201; Prn:23/01/2015; 15:31] P.12 (1-19)
12 H.V. Rowley et al. / Fuzzy Sets and Systems ••• (••••) •••–•••

1 Table 5 1
2 Ranked Choquet integrals, compared with arithmetic means, for 135 Australian industry sectors. Difference 2
3 in rank is also shown (e.g. ‘other services’ is ranked 6 by arithmetic mean and 4 by Choquet integral, with 3
4 the difference shown as ↑ 2). 4
5 Rank Alternative Choquet inteqral Arithmetic mean  Rank 5
6 1 Education 0.791 0.704 0 6
7 2 Health services 0.764 0.682 0 7
8 3 Black coal 0.755 0.669 0 8
4 Other services 0.751 0.663 ↑2
9 9
5 Services to finance 0.748 0.662 ↑2
10 6 Sanitary and garbage disposal 0.748 0.659 ↑4 10
11 7 Iron ores 0.745 0.666 ↓3 11
12 8 LNG, LPG 0.744 0.659 ↑1 12
13 9 Ownership of dwellings 0.743 0.629 ↑ 18 13
10 Retail trade 0.730 0.663 ↓5
14 14
11 Personal services 0.726 0.662 ↓3
15 12 Insurance 0.726 0.633 ↑ 12 15
16 13 Tobacco products 0.724 0.629 ↑ 16 16
17 14 Gold and lead 0.723 0.643 0 17
18 15 Libraries, museums, arts 0.722 0.658 ↓4 18
16 Government administration 0.720 0.641 ↑1
19 19
17 Crude oil 0.719 0.632 ↑8
20 18 Wheat & other grains 0.716 0.649 ↓6 20
21 19 Beer and malt 0.712 0.637 ↑3 21
22 20 Barley 0.711 0.648 ↓7 22
23 21 Other services to transport 0.709 0.638 0 23
22 Banking 0.709 0.642 ↓6
24 24
23 Silver and zinc 0.708 0.631 ↑3
25 24 Community care 0.706 0.639 ↓5 25
26 25 Other manufacturing 0.704 0.640 ↓7 26
27 26 Other business services 0.703 0.642 ↓ 11 27
28 27 Railway freight 0.700 0.625 ↑6 28
28 Mechanical repairs 0.700 0.636 ↓5
29 29
29 Non-bank finance 0.699 0.639 ↓9
30 30 Brown coal 0.699 0.611 ↑ 22 30
31 31 Scientific and technical services 0.696 0.629 ↓3 31
32 32 Gas production and distribution 0.694 0.623 ↑4 32
33 33 Leather products 0.694 0.624 ↑1 33
34 Natural gas 0.693 0.606 ↑ 25
34 34
35 Other non-ferrous metal ores 0.693 0.616 ↑8
35 36 Other transport 0.691 0.613 ↑ 13 35
36 37 Pipeline transport 0.690 0.612 ↑ 13 36
37 38 Railway passenger transport 0.690 0.615 ↑8 37
38 39 Non-residential construction 0.689 0.628 ↓9 38
40 Business management services 0.689 0.623 ↓3
39 39
41 Bakery products 0.688 0.628 ↓ 10
40 42 Other repairs 0.688 0.628 ↓ 10 40
41 43 Wholesale trade 0.688 0.619 ↓4 41
42 44 Water supply, sewerage and drainage 0.687 0.611 ↑ 10 42
43 45 Accommodation, cafes and restaurants 0.685 0.622 ↓7 43
46 Communication 0.685 0.624 ↓ 11
44 44
47 Copper 0.680 0.606 ↑ 14
45 48 Fabricated metal products 0.678 0.619 ↓8 45
46 49 Sport and recreation 0.677 0.618 ↓8 46
47 50 Defence 0.677 0.601 ↑ 15 47
48 51 Other machinery and equipment 0.676 0.615 ↓6 48
52 Aluminium 0.676 0.605 ↑ 10
49 49
53 Vegetable and fruit 0.674 0.617 ↓ 11
50 54 Services to agriculture 0.673 0.616 ↓ 10 50
51 55 Sheep and shorn wool 0.673 0.607 ↑3 51
52 (continued on next page) 52
 JID:FSS AID:6731 /FLA [m3SC+; v1.201; Prn:23/01/2015; 15:31] P.13 (1-19)
H.V. Rowley et al. / Fuzzy Sets and Systems ••• (••••) •••–••• 13

1 1
2
Table 5 (continued) 2
3 Rank Alternative Choquet inteqral Arithmetic mean  Rank 3
4 56 Furniture 0.672 0.608 ↑1 4
5 57 Wine and spirits 0.672 0.615 ↓ 10 5
58 Road freight 0.672 0.609 ↓2
6 6
59 Sawmill products 0.671 0.614 ↓ 11
7 7
60 Bus and tramway transport 0.671 0.611 ↓7
8 61 Taxi and hired car with driver 0.671 0.610 ↓6 8
9 62 Footwear 0.670 0.606 ↓2 9
10 63 Scientific equipment 0.668 0.598 ↑6 10
64 Ceramic products 0.665 0.612 ↓ 13
11 11
65 Other basic non-ferrous metal product 0.665 0.585 ↑ 20
12 12
66 Publishing 0.663 0.599 ↑2
13 67 Confectionery 0.662 0.595 ↑6 13
14 68 Other wood products 0.661 0.594 ↑7 14
15 69 Property services 0.661 0.585 ↑ 15 15
70 Services to mining 0.660 0.591 ↑ 10
16 16
71 Glass products 0.659 0.602 ↓8
17 17
72 Other food products 0.659 0.601 ↓8
18 73 Alumina 0.657 0.595 ↑1 18
19 74 Fibres, yars, fabrics 0.657 0.594 ↑3 19
20 75 Plaster and other concrete products 0.657 0.594 ↑1 20
76 Bauxite 0.657 0.582 ↑ 15
21 21
77 Motion picture, radio and TV 0.656 0.597 ↓7
22 22
78 Dairy cattle & milk 0.655 0.600 ↓ 12
23 79 Other mining 0.654 0.585 ↑7 23
24 80 Poultry and eggs 0.654 0.596 ↓9 24
25 81 Sugar cane 0.654 0.599 ↓ 14 25
82 Residential building 0.652 0.591 ↓3
26 26
83 Electrical equipment 0.650 0.590 ↓2
27 27
84 Aircraft 0.650 0.539 ↑ 38
28 85 Flour and cereal foods 0.648 0.589 ↓3 28
29 86 Other non-metallic mineral products 0.647 0.596 ↓ 14 29
30 87 Cotton 0.647 0.594 ↓9 30
88 Printing 0.646 0.575 ↑ 10
31 31
89 Fruit and vegetable products 0.646 0.586 ↓6
32 32
90 Structural metal products 0.645 0.584 ↓2
33 91 Pigs 0.644 0.583 ↓2 33
34 92 Air and space transport 0.644 0.570 ↑ 11 34
35 93 Dairy products 0.644 0.584 ↓6 35
94 Agricultural and other machinery 0.643 0.579 0
36 36
95 Sheet metal products 0.642 0.580 ↓2
37 37
96 Hardwoods 0.641 0.582 ↓6
38 97 Meat products 0.641 0.580 ↓5 38
39 98 Soft drinks, cordials and syrups 0.638 0.575 ↓1 39
40 99 Cosmetics and toiletry preparations 0.638 0.575 ↓3 40
100 Rubber products 0.638 0.563 ↑6
41 41
101 Clothing 0.636 0.569 ↑3
42 42
102 Prefabricated buildings 0.636 0.577 ↓7
43 103 Textile products 0.635 0.572 ↓2 43
44 104 Water transport 0.632 0.561 ↑4 44
45 105 Commercial fishing 0.631 0.572 ↓6 45
106 Basic iron and steel 0.630 0.570 ↓4
46 46
107 Softwoods 0.630 0.572 ↓7
47 47
108 Pulp, paper and paperboard 0.630 0.563 ↓3
48 109 Knitting mill products 0.629 0.560 ↑1 48
49 110 Concrete and mortar 0.626 0.561 ↓1 49
50 111 Plastic products 0.624 0.563 ↓4 50
112 Cement 0.624 0.559 ↓1
51 51
(continued on next page)
52 52
 JID:FSS AID:6731 /FLA [m3SC+; v1.201; Prn:23/01/2015; 15:31] P.14 (1-19)
14 H.V. Rowley et al. / Fuzzy Sets and Systems ••• (••••) •••–•••

1 Table 5 (continued) 1
2 2
Rank Alternative Choquet inteqral Arithmetic mean  Rank
3 3
113 Electronic equipment 0.623 0.547 ↑4
4 4
114 Household appliances 0.622 0.555 ↓1
5 115 Railway equipment 0.622 0.535 ↑ 10 5
6 116 Automotive petrol 0.621 0.513 ↑ 14 6
7 117 Ships and boats 0.621 0.544 ↑3 7
118 Motor vehicles and parts 0.619 0.557 ↓6
8 8
119 Paper containers and products 0.616 0.547 ↓1
9 9
120 Lime 0.615 0.549 ↓5
10 121 Other chemical products 0.614 0.549 ↓5 10
11 122 Pharmaceuticals 0.614 0.550 ↓8 11
12 123 Agricultural chemicals 0.610 0.544 ↓4 12
124 Paints 0.610 0.536 0
13 13
125 Gas oil, fuel oil 0.607 0.503 ↑8
14 14
126 Kerosene and aviation jet fuel 0.606 0.507 ↑5
15 127 Oils and fats 0.600 0.536 ↓4 15
16 128 Soap and other detergents 0.597 0.532 ↓2 16
17 129 Beef cattle 0.589 0.527 ↓2 17
130 Electricity supply 0.587 0.525 ↓2
18 18
131 Rice 0.579 0.541 ↓ 10
19 19
132 Other petroleum and coal products 0.579 0.486 ↑3
20 133 Basic chemicals 0.577 0.523 ↓4 20
21 134 Mixed fertilisers 0.563 0.507 ↓2 21
22 135 Forestry 0.552 0.493 ↓1 22
23 23
24 24
25 25
26 26
27 27
28 28
29 29
30 30
31 31
32 32
33 33
34 34
35 35
36 36
37 37
38 38
39 39
Fig. 1. Comparison between the arithmetic means and Choquet integrals for the 135 Australian industry sectors. The arithmetic mean favours some
40 sectors (those close to the diagonal, e.g. personal services, rice, forestry) and the Choquet integral favours others (those far from the diagonal, e.g. 40
41 education, aircraft, other petroleum & coal products), regardless of those sectors’ overall rankings. 41
42 42
43 ence between the first and last-ranked sectors under each aggregation scheme is approximately equal; but the position 43
44 of sectors within that range has changed. In other words, this was not a case of the ‘rich getting richer’, but rather that 44
45 the effects of criteria interaction were independent of the result of the aggregation itself. 45
46 The original Balancing Act study found that some sectors were disadvantaged by accounting anomalies or tech- 46
47 nological change. In forestry, for example, the use of gross versus net greenhouse gas emissions did not adequately 47
48 reflect the overall benefit of forest and plantation sectors’ carbon sequestration [35]. Similarly, due to primary data 48
49 limitations, the beef cattle sector’s profile reflects the high greenhouse gas emissions associated with land clearing in 49
50 the early 1990s which has since declined. Furthermore, academic discussion still revolves about the net carbon effect 50
51 of vegetation thickening in rangeland areas. Finally, the high water use intensity of rice growing has been improved 51
52 by the industry over the past decade through advances in irrigation practices. In general, there are many technologi- 52
 JID:FSS AID:6731 /FLA [m3SC+; v1.201; Prn:23/01/2015; 15:31] P.15 (1-19)
H.V. Rowley et al. / Fuzzy Sets and Systems ••• (••••) •••–••• 15

1 1
2 2
3 3
4 4
5 5
6 6
7 7
8 8
9 9
10 10
11 11
12 12
13 13
14 14
15 15
Fig. 2. Frequency distribution of the normalised scores pki , arithmetic means, and Choquet integrals (Eq. (2)) for the 135 Australian industry sectors.
16 16
By definition (Section 4.2), a normalised score of 1 represents best practice among the industry sectors, and a normalised score of 0 represents
17 worst practice. 17
18 18
19 19
cal solutions that might assist the below-best-practice sectors and may already be doing so, but in reality, significant
20 20
transformation of operating regimes usually takes place over decades rather than years.
21 21
The use of only 11 macro-economic criteria in Balancing Act to define ‘best practice’ has also received some valid
22 22
comment, and various anomalies were criticised. For example, the presence of the beer and malt and tobacco sectors in
23 23
the top tercile in Table 5 is due mainly to their high contribution to government revenue through excise taxes. Each of
24 24
these sectors has considerable downstream impacts on issues such as human health and social dislocation. These have
25 25
not been entered into the analysis in a ‘net-cost’ sense. For example, tobacco excise contributes A$5 billion annually
26 26
to government revenue, but associated health costs are estimated at approximately A$10 billion [36]. Applying an
27 27
extended set of criteria would improve the balance and effectiveness of the ranking in Table 5, although increasing the
28 28
number of criteria will not necessarily increase their utility.
29 29
30 4.3.5. Distribution of partial and aggregated scores 30
31 A frequency distribution (Fig. 2) of the normalised profiles p, their arithmetic mean, and their Choquet integral 31
32 shows that even though the single criteria profiles vary widely (dashed line), on average sectors are located within quite 32
33 a narrow band spanning 45% to 70% of best practice in terms of the arithmetic mean. The Choquet integral spans a 33
34 similar range, but is shifted to higher values between about 55% and 80% best practice. The peak at the upper end of 34
35 the scale for the profiles is due to the fact that a few primary sectors have very low absolute performances on some 35
36 environmental criteria, pushing many tertiary sectors very near best-practice on these criteria. Because of the nature 36
37 of the criteria interaction, Choquet integrals are not linearly related to arithmetic means, but are systematically higher 37
38 (Fig. 1). Inspection of the underlying data suggests that the reason for this is that, due to the distribution of the raw 38
39 data, the average of all sectors’ normalised scores against criteria that are relatively independent is typically higher 39
40 than the average of all sectors’ scores against criteria that are relatively dependent (the latter having less influence on 40
41 the Choquet aggregation than on the arithmetic mean). 41
42 42
43 5. Application to an LCA of biosolids management options 43
44 44
45 The method proposed here can be applied to any multi-criteria decision problem where criteria interact. A typical 45
46 application is in environmental life cycle assessment (LCA) where the whole-of-life burdens of a set of technical 46
47 options are assessed against a set of environmental criteria [37]. While the purpose of aggregating the Balancing Act 47
48 data was to rank the industry sectors to allow benchmarking, the purpose of a comparative LCA is often to recommend 48
49 the best alternative. Another difference is that, unlike in the previous case study, LCA criteria are often considered in 49
50 a net-cost sense (e.g. in the example below, the energy produced by some alternatives was taken into account when 50
51 calculating total energy use). As a demonstration, the algorithm is now applied to a process-based LCA of biosolids 51
52 management systems [38]. 52
 JID:FSS AID:6731 /FLA [m3SC+; v1.201; Prn:23/01/2015; 15:31] P.16 (1-19)
16 H.V. Rowley et al. / Fuzzy Sets and Systems ••• (••••) •••–•••

1 Table 6 1
2 PCA model for the biosolids dataset. 2
3 PC # Eigenvalue Variance explained Cumulative variance 3
4 explained 4
5 1 3.17 63.4% 63.4% 5
6 2 1.20 24.0% 87.4% 6
7 3 0.48 9.5% 96.9% 7
4 0.14 2.8% 99.8%
8 8
5 0.01 0.2% 100%
9 9
10 10
11 Table 7 11
12 Criteria correlation for the biosolids management options. 12
13 13
WU TEC CF TETP HTP
14 14
Water use (WU) 1 0.17 −0.25 −0.35 −0.54
15 Total energy consumption (TEC) 1 0.51 0.78 0.55 15
16 Carbon footprint (CF) 1 0.63 0.66 16
17 Terrestrial ecotoxicity potential (TETP) 1 0.94 17
18
Human toxicity potential (HTP) 1 18
19 19
20 20
5.1. The original study
21 21
22 22
23
The original study examined the environmental burdens of eight technical options for managing one day’s sewage 23
24
sludge generated by an equivalent population of 40 000 people located in rural Australia. The criteria assessed were: 24
25
water use (WU), total energy consumption (TEC), and carbon footprint (CF), all as defined by Foran et al. [33]; and 25
26
human toxicity potential (HTP, the human health burden of carcinogenic and non-carcinogenic toxins) and terrestrial 26
27
ecotoxicity potential (TETP, the burden of toxins on terrestrial ecosystem health), both as defined by Lundie et al. [39]. 27
28
These indicators were chosen by an industry panel and the research team to reflect the most significant issues 28
29
in relation to biosolids processing in Australia. The alternatives (technical options) involved various combinations of 29
30
treatment processes and end-uses and were called ‘Landfill’, ‘DigAg’ (digestion; agricultural use), ‘DigComp’ (diges- 30
31
tion; composting), ‘LimeAg’ (adding lime; agricultural use), ‘DryAg’ (drying; agricultural use), ‘DryKiln’ (drying; 31
32
energy recovery), ‘DigDryAg’ (digestion; drying; agricultural use), and ‘DigDryKiln’ (digestion; drying; energy re- 32
33
covery). 33
34 34
35 5.2. Method 35
36 36
37 The algorithm was implemented as described above, inverting the consistently negative indicators and using (un- 37
38 weighted) Pearson correlation coefficients in the absence of any a priori information about the relative importance of 38
39 the alternatives. The overall PCA model for the biosolids dataset is shown in Table 6. 39
40 40
41 5.3. Results and discussion 41
42 42
43 TETP and HTP are highly correlated with each other, and also with TEC and CF (Table 7). This is because both 43
44 TETP and HTP are primarily influence by the same coal combustion emissions, and Australian energy supplies are 44
45 dominated by the coal industry. It also reflects the important assumption that trace contaminants in the biosolids are 45
46 ‘locked’ into the soil matrix. These calculations could also be performed with the results of the alternative, conserva- 46
47 tive assumption (results shown in [38]), that all trace contaminants become bioavailable. 47
48 Shapley values and interaction indices are shown in Table 8. Water use is the most discriminating criterion, having 48
49 the largest Shapley value and showing the smallest interactions with the other criteria, as well as being negatively 49
50 correlated with CF, TETP and HTP (Table 7). One implication of this relates to the increasingly popular use of CF 50
51 as a proxy indicator for all environmental damage: although our results suggest this may be appropriate for the other 51
52 environmental indicators (TEC, TETP, HTP), water use would not be appropriately represented by such an approach. 52
 JID:FSS AID:6731 /FLA [m3SC+; v1.201; Prn:23/01/2015; 15:31] P.17 (1-19)
H.V. Rowley et al. / Fuzzy Sets and Systems ••• (••••) •••–••• 17

1 Table 8 1
2 Shapley values (Eq. (3)) and interaction indices (Eq. (4)) for the five criteria in the biosolids case study. 2
3 Shapley values Interaction indices 3
4 WU TEC CF TETP HTP 4
5 5
Water use (WU) 0.292 – −0.005 −0.011 −0.025 −0.082
6 Total energy consumption (TEC) 0.213 – −0.059 −0.123 −0.050 6
7 Carbon footprint (CF) 0.221 – −0.059 −0.072 7
8 Terrestrial ecotoxicity potential (TETP) 0.135 – −0.139 8
9
Human toxicity potential (HTP) 0.138 – 9
10 10
11 11
12 12
13 13
14 14
15 15
16 16
17 17
18 18
19 19
20 20
21 21
22 22
23 23
24 24
Fig. 3. Choquet integrals and arithmetic means for the biosolids processing alternatives. Differences in rank order are noted in brackets.
25 25
26 26
27 Using the Choquet integral instead of the arithmetic mean to aggregate the normalised results again changes the 27
28 rank order of the alternatives’ global scores (Fig. 3). LimeAg shows the largest positive difference, followed by DigAg 28
29 and DigComp; while DryKiln and DigDryKiln move down the rankings. Inspection of the performance table (Table 3 29
30 in [38]) and the correlation matrix (Table 7) suggests that this is because LimeAg’s best normalised scores are against 30
31 criteria that are relatively independent, so that those criteria have more influence on the Choquet aggregation than on 31
32 the arithmetic mean. The opposite is true for the scenarios that move down the rankings. 32
33 33
34 6. Conclusions 34
35 35
36 Aggregating multi-criteria data is an important problem with many applications. Commonly-used additive aggrega- 36
37 tion methods, including the weighted arithmetic mean and sum, assume criteria independence. This is rarely achieved 37
38 in practice, particularly when the criteria are chosen based on data availability. 38
39 The Choquet integral is a suitable aggregation operator in the presence of interacting criteria. It replaces the weight 39
40 vector with a fuzzy measure, which models the importance of each subset or coalition of criteria, rather than just the 40
41 importance of individual criteria. The fuzzy measures can also be used to compute other informative parameters such 41
42 as the Shapley importance index, which represents the marginal contribution of each individual criterion alone in all 42
43 coalitions, and the Shapley interaction index, which represents the average marginal interaction between criteria. 43
44 However, estimating the fuzzy measures has been problematic. Conventional supervised approaches such as that 44
45 described by Marichal [40] are cognitively challenging because they require the decision maker to assign weights not 45
46 only to each individual criterion but also to each subset (coalition) of multiple criteria. 46
47 In response to this, Kojadinovic [4,8] has proposed an unsupervised approach to fuzzy measure estimation, based 47
48 on the use of entropy functionals. However, this method requires a very large number of profiles to sufficiently densely 48
49 populate the required probability distributions. This so-called ‘curse of dimensionality’ reduces its accuracy for prob- 49
50 lems with a small number of alternatives, such as in typical environmental and sustainability assessment problems. 50
51 In this paper, the authors propose a novel approach to estimating the required fuzzy measures which overcomes 51
52 these limitations. First, the notion of an equivalent number of non-interacting criteria was introduced, via a series of 52
 JID:FSS AID:6731 /FLA [m3SC+; v1.201; Prn:23/01/2015; 15:31] P.18 (1-19)
18 H.V. Rowley et al. / Fuzzy Sets and Systems ••• (••••) •••–•••

1 intuitive examples, as a suitable basis for estimating the fuzzy measures. Second, it was proposed and demonstrated 1
2 that the equivalent number of non-interacting criteria could be determined using principal component analysis. 2
3 The authors then successfully applied the approach to two case studies, by comparing the results of the Choquet 3
4 integral aggregation with the results of the weighted arithmetic mean. 4
5 The first case study was an input-output analysis of 135 Australian industry sectors evaluated against 11 sustain- 5
6 ability criteria. Compared to the arithmetic mean, the Choquet integral favoured sectors that had their best normalised 6
7 scores against the relatively independent criteria. The arithmetic mean gave an arguably unfair advantage to those 7
8 sectors that performed well against criteria that interact and tell largely related stories. For example, gross operat- 8
9 ing surplus, income and employment were shown to be highly correlated and therefore had lower importance in the 9
10 Choquet aggregation. 10
11 The second case study was a typical LCA study of 8 alternative biosolids management process streams evaluated 11
12 against 5 environmental criteria. Many of the criteria were found to be correlated, with 87% of the total variance in 12
13 the dataset being explained by the first two PCs. Because it was able to account for these interactions, the Choquet 13
14 integral ranked the alternatives in a different order than they were ranked by the arithmetic mean. 14
15 These examples demonstrated the ability of the proposed approach to account for criteria interaction. However, 15
16 it also has limitations. First, correlation coefficients can vary with the number of profiles, although towards large 16
17 numbers of observations, stability can be expected. Second, and probably more importantly, the (weighted and un- 17
18 weighted) Pearson correlation coefficient used here only accounts for linear relationships between variables. Third, 18
19 by removing unintentional criteria interaction from the aggregation scheme, the ability of analysts to intentionally 19
20 include interacting criteria has also been removed (e.g. a cost-benefit approach in which including both energy costs 20
21 and GHG emission tax costs may be desired). 21
22 This novel approach to the estimation of Choquet integral weights represents a significant advancement in the 22
23 applicability of this aggregation model, particularly for environmental and sustainability assessment problems. Ap- 23
24 plying the Choquet integral in situations where criteria interact can in principle defuse debates over criteria choice and 24
25 construction, and thus lead to more acceptable outcomes. 25
26 26
27 Acknowledgements 27
28 28
29 We thank Gregory Peters, Barney Foran, and Sven Lundie for providing comments on an early version of this 29
30 paper. Julien Ugon gave valuable input for the proof of the monotony of J ∗ . We also thank Ivan Kojadinovic, Fred 30
31 Rowley, and other anonymous reviewers for their valuable comments on drafts of this paper. 31
32 32
33 References 33
34 34
35 [1] J.R. Doyle, Multiattribute choice for the lazy decision maker: let the alternatives decide!, Organ. Behav. Hum. Decis. Process. 62 (1995) 35
87–100.
36 36
[2] M. Lenzen, R. Schaeffer, R. Matsuhashi, Selecting and assessing sustainable CDM projects using multi-criteria methods, Climate Policy 7
37 37
(2007) 121–138.
38 [3] G. Choquet, Theory of capacities, Ann. Inst. Fourier 5 (1953) 131–295. 38
39 [4] I. Kojadinovic, Estimation of the weights of interacting criteria from the set of profiles by means of information-theoretic functionals, Eur. J. 39
40 Oper. Res. 155 (2004) 741–751. 40
[5] J.L. Marichal, M. Roubens, Determination of weights of interacting criteria from a reference set, Eur. J. Oper. Res. 124 (2000) 641–650.
41 41
[6] U. Ebert, H. Welsch, Meaningful environmental indices: a social choice approach, J. Environ. Econ. Manag. 47 (2004) 270–283.
42 [7] S. Angilella, S. Greco, B. Matarazzo, Non-additive robust ordinal regression: a multiple criteria decision model based on the Choquet integral, 42
43 Eur. J. Oper. Res. 201 (2010) 277–288. 43
44 [8] I. Kojadinovic, Unsupervized aggregation of commensurate correlated attributes by means of the Choquet integral and entropy functionals, 44
45
Int. J. Intell. Syst. 23 (2008) 128–154. 45
[9] J.L. Marichal, An axiomatic approach of the discrete Choquet integral as a tool to aggregate interacting criteria, IEEE Trans. Fuzzy Syst. 8
46 46
(2000) 800–807.
47 [10] M. Grabisch, J.-L. Marichal, R. Mesiar, E. Pap, Aggregation Functions, Cambridge University Press, Cambridge, 2009. 47
48 [11] M. Grabisch, The application of fuzzy integrals in multicriteria decision making, Eur. J. Oper. Res. 89 (1996) 445–456. 48
49 [12] M. Sugeno, Theory of Fuzzy Integrals and Its Applications, Tokyo Institute of Technology, Tokyo, 1974. 49
[13] M. Grabisch, C. Labreuche, Fuzzy measures and integrals in MCDA, in: J. Figueira, S. Greco, M. Ehrgott (Eds.), Multiple Criteria Decision
50 50
Analysis: State of the Art Surveys, Springer, New York, 2005, pp. 563–608.
51 51
[14] D. Denneberg, Non-additive measure and integral, basic concepts and their role for applications, in: M. Grabisch, T. Murofushi, M. Sugeno
52 (Eds.), Fuzzy Measures and Integrals – Theory and Applications, Physica Verlag, Heidelberg, 2000, pp. 42–69. 52
 JID:FSS AID:6731 /FLA [m3SC+; v1.201; Prn:23/01/2015; 15:31] P.19 (1-19)
H.V. Rowley et al. / Fuzzy Sets and Systems ••• (••••) •••–••• 19

1 [15] T. Murofushi, M. Sugeno, Fuzzy measures and fuzzy integrals, in: M. Grabisch, T. Murofushi, M. Sugeno (Eds.), Fuzzy Measures and Integrals 1
2 – Theory and Applications, Physica Verlag, Heidelberg, 2000, pp. 3–41. 2
3 [16] M. Grabisch, The representation of importance and interaction of features by fuzzy measures, Pattern Recognit. Lett. 17 (1996) 567–575. 3
[17] L.S. Shapley, A value for n-person games, in: H.W. Kuhn, A.W. Tucker (Eds.), Contributions to the Theory of Games, Princeton University
4 4
Press, Princeton, NJ, 1953, pp. 307–317.
5 5
[18] T. Murofushi, S. Soneda, Techniques for reading fuzzy measures (III): interaction index, in: 9th Fuzzy System Symposium, Sapporo, Japan,
6 1993, pp. 693–696. 6
7 [19] M. Grabisch, k-order additive discrete fuzzy measures and their representation, Fuzzy Sets Syst. 92 (1997) 167–189. 7
8 [20] M. Roubens, Interaction between criteria and definition of weights in MCDA problems, in: 44th Meeting of the European Working Group 8
“Multicriteria Aid for Decisions”, Brussels, Belgium, 1996.
9 9
[21] T. Marchant, Towards a theory of MCDM: stepping away from social choice theory, Math. Soc. Sci. 45 (2003) 343–363.
10 10
[22] S. Angilella, S. Greco, F. Lamantia, B. Matarazzo, Assessing non-additive utility for multicriteria decision aid, Eur. J. Oper. Res. 158 (2004)
11 734–744. 11
12 [23] J.M. Bland, D.G. Altman, Statistics notes: calculating correlation coefficients with repeated observations: part 2–correlation between subjects, 12
13 Br. Med. J. 310 (1995) 633. 13
[24] J.M. Bland, D.G. Altman, Correction, Br. Med. J. 312 (1996) 572a.
14 14
[25] I.T. Jolliffe, Principal Component Analysis, 2nd ed., Springer-Verlag, New York, 2002.
15 15
[26] J. Shlens, A tutorial on principal component analysis: derivation, discussion and singular value decomposition (Version 1), 2003.
16 [27] L.I. Smith, A tutorial on principal components analysis, 2002. 16
17 [28] I.T. Jolliffe, Principal Component Analysis, Springer-Verlag, New York, 1986. 17
18 [29] L. Eriksson, E. Johansson, N. Kettaneh-Wold, S. Wold, Multi- and Megavariate Data Analysis: Principles and Applications, Umetrics 18
Academy, Umeå, Sweden, 2001.
19 19
[30] R. Kramer, Chemometric Techniques for Quantitative Analysis, Marcel Dekker, New York, 1998.
20 20
[31] B.M. Wise, N.B. Gallagher, R. Bro, J.M. Shaver, W. Windig, R.S. Koch, PLS_Toolbox Version 4.0 for use with MATLAB™, Eigenvector
21 Research, Inc., Wenatchee, WA, 2006. 21
22 [32] S.-G. Hwang, Cauchy’s interlace theorem for eigenvalues of Hermitian matrices, Am. Math. Mon. 111 (2004) 157–159. 22
23 [33] B. Foran, M. Lenzen, C. Dey, Balancing Act: A Triple Bottom Line Analysis of the Australian Economy, CSIRO and University of Sydney, 23
Australia, 2005.
24 24
[34] M. Lenzen, C.J. Dey, Economic, energy and greenhouse emissions impacts of some consumer choice, technology and government outlay
25 25
options, Energy Econ. 24 (2002) 377–403.
26 [35] ISA, Clarification of the Derivation and Meaning of Figures Relating to Greenhouse Gas Emissions in the Balancing Act Report on Three 26
27 Forestry Sectors, Centre for Integrated Sustainability Analysis, The University of Sydney, Sydney, 2005. 27
28 [36] J. Gordon, E. Duff, Smokers priced out of the habit, in: The Sydney Morning Herald, Fairfax, Sydney, 2009. 28
[37] ISO, ISO 14040:2006(E) Environmental management—Life cycle assessment—Principles and framework, International Organisation for
29 29
Standardisation, Geneva, 2006.
30 30
[38] G.M. Peters, H.V. Rowley, Environmental comparison of biosolids management systems using life cycle assessment, Environ. Sci. Technol.
31 43 (2009) 2674–2679. 31
32 [39] S. Lundie, M.A.J. Huijbregts, H.V. Rowley, N.J. Mohr, A.J. Feitz, Australian characterisation factors and normalisation figures for human 32
33 toxicity and ecotoxicity, J. Clean. Prod. 15 (2007) 819–832. 33
[40] J.-L. Marichal, Behavioral analysis of aggregation in multicriteria decision aid, in: J. Fodor, B. de Baets, P. Perny (Eds.), Preferences and
34 34
Decisions under Incomplete Knowledge, Physica Verlag, Heidelberg, 2000, pp. 53–178.
35 35
36 36
37 37
38 38
39 39
40 40
41 41
42 42
43 43
44 44
45 45
46 46
47 47
48 48
49 49
50 50
51 51
52 52