Вы находитесь на странице: 1из 15

Expert Systems with Applications 41 (2014) 2083–2097

Contents lists available at ScienceDirect

Expert Systems with Applications


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

Integration of semi-fuzzy SVDD and CC-Rule method for supplier


selection
Xuesong Guo ⇑, Zhengwei Zhu, Jia Shi
School of Public Policy and Administration, Xi’an Jiaotong University, 710049, China

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

Keywords: A model based on semi-fuzzy support vector domain description (semi-fuzzy SVDD) is put forward to
Supplier selection address multi-classification problem involved in supplier selection. By preprocessing using semi-fuzzy
Semi-fuzzy kernel clustering algorithm kernel clustering algorithm, original samples are divided into two subsets: deterministic samples and
Support vector domain description fuzzy samples. Only the fuzzy samples, rather than all original ones, require expert judgment to decide
Cooperative coevolution algorithm
their categories and are selected as training samples to accomplish SVDD specification. Therefore, the
samples preprocessing method can not only decrease experts working strength, but also achieve less
computational consumption and better performance of the classifier. Nevertheless, in order to accom-
plish practical decision making, another condition has to be met: good explanations to the decision. A
rule extraction method based on cooperative coevolution algorithm (CCEA), is introduced to achieve
the target. To validate the proposed methodology, samples from real world were employed for experi-
ments, with results compared with conventional multi-classification support vector machine approaches
and other artificial intelligence techniques. Moreover, in terms of rule extraction, experiments on key
parameters, different methods including decompositional and pedagogical ones etc. were also conducted.
 2013 Elsevier Ltd. All rights reserved.

1. Introduction complexity of the multi-classification problem, it is tiring and


time-consuming to consider every training sample to decide its
Supplier selection is a complex multi-criteria decision making category. So, if only the ones, which are difficult to categorize
problem, on which some conflicting criteria might be considered and vital for classifier establishment, can be selected for expert
and evaluated. Decisions are complicated due to many dynamic judgment, working strength, possibility of human error and com-
factors, such as uncertain price, capacity, and demand etc. So, is- putational consumption can be reduced greatly.
sues in supplier selection have received extensive attentions, with Thirdly, in terms of supplier selection, good prediction accuracy
some valuable academic surveys accomplished (Ho, Xu, & Dey, alone cannot be considered as truly reliable decision support. Some
2010; Jain, Wadhwa, & Deshmukh, 2009). Moreover, a systemic lit- supplementary information on how a verdict has been reached is
erature review on the latest development in this area was pre- also necessary. So, the method for rule extraction, which can dis-
sented in Chai, Liu, and Ngai (2013). Among the proposed cover attribute values connected to each possible outcome, is re-
methodologies and techniques, support vector machine, one of quired. Although the models based on machine learning
promising machine learning methods, has begun to play more techniques can achieve high accuracies, the comprehensibility is
important roles, which can be seen in section 1. However, in terms limited. And it is a major drawback, which will cause reluctance
of supplier selection using the method, some problems still exist, to use the method.
which can be concluded as follows: Moreover, at the methodological level, rule extraction from sup-
Firstly, multi-classification problem should be addressed. That port vector machine models, rather than from data directly, is also
is, suppliers will usually be divided into more than two categories, an important research issue (Barakat & Bradley, 2010). Methods
while standard support vector machine based on hyper-plane can- based on standard support vector machine do not provide explana-
not handle it directly. tions or comprehensible justifications for the knowledge they
Secondly, according to conventional methodologies, based on learn. This has been shown to be one of the main obstacles imped-
machine learning techniques, experts will be required to evaluate ing their practical applications. Hence, models based on support
all training samples to determine their classification labels. Given vector machine were integrated with some other approaches to
accomplish rule extraction in different areas (He, Hu, & Harrison,
2006; Barakat, 2007; Castro, Flores-Hidalgo, Mantas, & Puche,
⇑ Corresponding author. 2007; Martens, Baesens, & Gestel, 2009; Ren, Liu, & Cao, 2011;
E-mail address: guoxues1@163.com (X. Guo).
Zhu & Hu, 2013).

0957-4174/$ - see front matter  2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.eswa.2013.09.008
2084 X. Guo et al. / Expert Systems with Applications 41 (2014) 2083–2097

In terms of the problems mentioned above, motivations of the integer NLP formulation were put forward (Rezaei & Davoodi,
research can be concluded as follows: 2012). On the other hand, Multi-objective Programming (MOP), a
kind of MP characterized by multiple and conflicting objectives
(1) Develop a model suitable for handling multi-classification functions, were applied. Especially, methods based on fuzzy MOP
problem in supplier selection. were presented in Lin (2012), Ozkok and Tiryaki (2011), Haleh
(2) Propose sample preprocessing method to ensure only the and Hamidi (2011), Yu, Goh, and Lin (2012) and Shaw, Shankar,
samples, which are difficult to categorize and vital for classi- Yadav, and Thakur (2012). As an extension or generalization of
fier establishment, are selected for expert judgment. More- MOP, Goal Programming (GP) can be applied to deal with multiple
over, sample preprocessing is also expected to reduce the and conflicting objective measures (Kull & Talluri, 2008; Tsai &
size of training set to decrease computational consumption Hung, 2009; Chen, 2011; Liao & Kao, 2011). Given the uncertain
and achieve better performance of the classifier. factors, Stochastic Programming (SP), a framework for modeling
(3) Develop techniques for rule extraction from multi-classifica- uncertainty optimization problems, was also introduced (Kara,
tion model, where the extracted rules are required to pro- 2011; Li & Zabinsky, 2011).
vide explanations to the classification results. Furthermore, models based on Artificial Intelligence (AI) have
begun to play more important roles, because they do not require
Therefore, a two-step hybridized method is proposed in the pa- formulization of the decision making process. So, they can cope
per. In the first step, multi-classification classifier is developed better with complexity and uncertainty. When the models are em-
with the thoughts of support vector domain description (SVDD) ployed, only information on characteristics of current situation,
employed. Especially, training samples, selected from original ones e.g., performance on the criteria, is required. Moreover, the AI tech-
after preprocessing, are deduced with their classification labels as- nologies subsequently make the actual trade off, based on what
signed using expert judgment. In the second step, CC-Rule method they have ‘‘learned’’ from experts or cases.
based on cooperative coevolution algorithm is introduced to Generally, Genetic Algorithm (GA), Neural Network (NN), Grey
achieve comprehensible rules extraction based on the set of sup- System Theory (GST), and Rough Set Theory (RST) have been re-
port vectors obtained in the first step. garded as the most popular methods (Chai et al., 2013). Besides
models based on GA (Yang, Wee, Pai, & Tseng, 2011; Yeh & Chuang,
2011), some simulations (Xu & Ding, 2011), and a heuristic algo-
2. Literature review rithm combining Guided GA, Pareto GA were presented (Che,
2010). Moreover, GA can be also applied as one element for deci-
2.1. Decision-making techniques applied in supplier selection sion model establishment (Lin & Yeh, 2010; Sadeghieh, Dehghan-
baghi, Dabbaghi, & Barak, 2012). On the other hand, some typical
According to current literatures, structural suppler selection is researches based on NN were presented (Celebi & Bayraktar,
usually defined as multi-criteria decision making (MCDM) prob- 2008; Wu, 2009), while approaches using adaptive neuro-fuzzy
lem. Therefore, some MCDM techniques have been employed to inference system (Guneri, Ertay, & Yucel, 2011), adaptive reso-
solve it, which can be concluded as follows: (1) Multi-attribute nance theory NN (Keskin, Ilhan, & Özkan, 2010) etc. were also
utility methods such as AHP (Levary, 2008; Chan & Chan, 2010; put forward. GST is another available technique. That is, the
Ishizaka, Pearman, & Nemery, 2012; Bhattacharya, Geraghty, & decision information can be depicted in the form of grey values
Young, 2010), ANP (Lin, Lin, Yu, & Tzeng, 2010; Tseng, Chiang, & (Sadeghieh et al., 2012; Bai & Sarkis, 2010; Tseng, 2011), while
Lan, 2009) etc.; (2) Outranking methods such as Elimination and some results can be obtained using grey relational analysis (GRA)
Choice Expressing Reality (ELECTRE) (Sevkli, 2010), Preference (Golmohammadi & Mellat-Parast, 2012; Pitchipoo, Venkumar, &
Ranking Organization Method for Enrichment Evaluation (PROM- Rajakarunakaran, 2012). RST, which can be applied to identify
ETHEE) (Chen, Wang, & Wu, 2011) etc.; (3) Compromise methods structural relationships with imprecise or noisy data, was intro-
such as Technique for Order Performance by Similarity to Ideal duced to this field (Bai & Sarkis, 2010; Chang & Hung, 2010).
Solution (TOPSIS) (Awasthi, Chauhan, & Goyal, 2010), Multi-crite- Besides the AI methods mentioned above, Case Based Reasoning
ria Optimization and Compromise Solution (VIKOR) (Chen & Wang, (CBR) (Zhao & Yu, 2011), Bayesian Networks (BN) (Dogan & Aydin,
2009) etc.; (4) other MCDM techniques such as Simple Multi-attri- 2011; Ferreira & Borenstein, 2012), Particle Swarm Optimization
bute Rating Technique (SMART) (Chou & Chang, 2008), Decision- (PSO) (Xu & Yan, 2011), Ant Colony Algorithm (ACA) (Tsai, Yang,
Making Trial and Evaluation Laboratory (DEMATEL) (Chang, Chang, & Lin, 2010), Dempster–Shafer Theory (DST) (Wu, 2009), Associa-
& Wu, 2011; Dalalah, Hayajneh, & Batieha, 2011; Buyukozkan & tion Rule (AR) (Lin, Chuang, Liou, & Wu, 2009), and Support Vector
Cifci, 2012) etc. Machine (SVM) (Guo, Yuan, & Tian, 2009) were regarded as other
Mathematical Programming (MP) techniques were also applied. alternatives.
A basic Data Envelopment Analysis (DEA) model, which can be More researchers have been focusing on integration of standard
used to evaluate the relative efficiency of DMUs according to techniques and methodologies towards more effective decision
multiple inputs and outputs, was put forward in some researches support. An evaluation model based on AHP for risk index calcula-
(Adler, Friedman, & Sinuany-Stern, 2002; Wu & Blackhurst, tion, which were then incorporated into a GP model, was presented
2009). Moreover, Linear Programming (LP), a mathematical opti- (Liao & Kao, 2011). An integrated decision model using AHP and
mization method, was employed in suppler selection. Models the Taguchi loss function was put forward (Ordoobadi, 2010).
based on simple LP (Ng, 2008; Guneri, Yucel, & Ayyildiz, 2009; And an integrated AHP and quality function deployment (QFD)
Chen, Wang, & Lu, 2011), fuzzy LP (Amin, Razmi, & Zhang, 2011; for ranking alternative suppliers was also argued (Bhattacharya
Yucel & Güneri, 2011; Lin, 2012), multi-objective LP (Ozkok & et al., 2010). Considering uncertain decision environments, some
Tiryaki, 2011; Xu & Ding, 2011) and mixed integer LP (Demirtas thoughts and methodologies based on fuzzy theories were intro-
& Üstün, 2008; Amin & Zhang, 2012) were developed. Besides LP, duced and integrated with AHP (Labib, 2011; Amid, Ghodsypour,
some other researchers studied suppler selection with methods & O’Brien, 2011; Chamodrakas, Batis, & Martakos, 2010; Chen &
based on Nonlinear Programming (NLP). Different from LP, NLP Chao, 2012; Kilincci & Onal, 2011). AHP can also be integrated with
allows for some constraints or objective functions represented as ANP, wherein triangular fuzzy values were employed to form
nonlinear equations. Besides simple application of NLP methods pairwise comparison matrix (Yucenur, Vayvay, & Demirel, 2011).
(Hsu, Chiang, & Shu, 2010), some modified ones based on mixed Multiple techniques, including AHP, TOPSIS, and DEA were
X. Guo et al. / Expert Systems with Applications 41 (2014) 2083–2097 2085

integrated (Zeydan, Çolpan, & Çobanoglu, 2011), with some inte- in these algorithms is the choice of the data partitioning strategy,
gration with other methods such as Structural Equation Model which strongly influences the generalization performances of the
(Punniyamoorthy, Mathiyalagan, & Parthiban, 2011), MOP (Yu classifiers.
et al., 2012; Shaw et al., 2012), GRA (Wu, 2009), etc. also argued. Different from the standard multi-class SVM based on hyper-
Besides AHP, ANP were integrated with other techniques. A simple plane, some models based on hyper-sphere were also put forward.
hybrid approach, in which ANP was used to determine the weights An SVDD classification method for one-class data was proposed,
of criteria, was proposed (Chang et al., 2011). An ANP-integrated whose main idea is to find a hyper-sphere covering one-class data
mixed-integer non-linear decision model was also developed and make its radius as small as possible (Lai, Tax, Duin, Pekalska, &
(Razmi & Rafiei, 2010), while integration of ANP and QFD was ar- Paclik, 2004; Lee, Park, & Lee, 2006). SVDD is an excellent method
gued (Ho, Dey, & Lockström, 2011). In terms of information system of one-class classification with the advantages of robustness and
design, ANP, TOPSIS, and LP were integrated for ERP system (Lin, high computation speed. Aiming at multi-classification problems,
Chen, & Ting, 2011). Considering ANP and fuzzy logic hybrid a sphere-structured multi-class SVM was proposed (Tax & Duin,
approaches, ANP techniques were integrated with fuzzy LP for 1999, 2004). This method categorizes the same class data with a
selecting the best suppliers and handling the inherent uncertainty hyper-sphere to reduce the complexity of the quadratic program-
(Lin, 2012). Some researchers constructed ANP-related models ming. When adding data for a new class, only the relevant hy-
combined with other techniques such as fuzzy methods (Vinodh, per-sphere needs to be constructed.
Anesh Ramiya, & Gautham, 2011), AHP (Yucenur et al., 2011) and
DEMATEL, TOPSIS (Buyukozkan & Cifci, 2012). DEA was integrated 3. Proposed methodology
with the cross efficiency evaluation (Falagario, Sciancalepore,
Costantino, & Pietroforte, 2012), while some other researchers 3.1. Framework of the methodology
constructed DEA-related models considering a triangular fuzzy
environment (Zeydan et al., 2011; Azadeh & Alem, 2010; Chen, In the paper, a two-step hybridized method is proposed. That is,
2011). Apart from fuzzy decision environments, some researchers classification and support vectors identification is performed using
discussed utilization of DEA in handling other types of uncertainty Semi-fuzzy Support Vector Domain Description (Semi-fuzzy
(Wu, 2010; Saen, 2010). SVDD), while rules are extracted with Evolutionary Algorithm
Moreover, besides the formulated supplier selection problem (EA) employed. The framework can be illustrated in Fig. 1.
under deterministic conditions, current studies addressed practical In the first step, a multi-classification classifier based on Semi-
problems under different types of uncertainties. Some integrated fuzzy SVDD is developed to achieve high prediction accuracy, with
with uncertain decision approaches were proposed (Yang et al., support vectors identified for the second step. Especially, original
2011; Pitchipoo et al., 2012). Several AI techniques, such as GST samples are divided into two subsets: Deterministic Samples and
(Golmohammadi & Mellat-Parast, 2012), were also introduced, Fuzzy Samples, by preprocessing with semi-fuzzy kernel clustering
and integrated to deal with supplier selection problems (Bai & algorithm. Only the fuzzy samples, rather than all original ones, re-
Sarkis, 2010; Wu, 2010). quire expert judgment to assign their classification labels to form
training set.
2.2. Multi-classification methods based on SVM In the second step, rule extraction method using Cooperative
Coevolution Algorithm (CCEA), labeled as CC-Rule method, is ap-
SVM method is a promising AI technique. It has been regarded plied subsequently to evolve a set of rules based on support
as a development in statistical learning theory, though researches vectors.
on supplier selection are few (Guo et al., 2009; Sun, Xie, & Xue,
2005). Since it is originally developed for binary classification 3.2. Multi-classification classifier based on semi-fuzzy SVDD
problem (Cortes & Vapnik, 1995), extend it to multi-class classifi-
cation effectively remains an important research issue. At present, 3.2.1. Multi-classification based on SVDD
there are two main approaches to address this problem. One is the SVDD is a sphere shaped data description. Using a nonlinear
single-machine approach, in which a single large optimization transformation to map the data into a high dimensional kernel
problem is solved (Vapnik, 1998, 1999). The other is the multi- space, it can obtain a flexible and accurate data description relying
ple-machine approach, which transforms a multi-class classifica- on a small number of support vectors (Tax & Duin, 1999, 2004).
tion problem into several binary classification problems Give a data set v ¼ fxi 2 Rd ji ¼ 1; 2; . . . ; ng data points and a
(Bredensteiner & Bennett, 1999; Guermeur, 2002). However, it re- nonlinear mapping u from the input space to a kernel space, a
quires a significant amount of computational time to solve the sin- smallest sphere, which can enclose most of the mapped data
gle large optimization problem and not suitable for practical points, will be developed. The sphere can be described using the
applications (Pontil & Verri, 1998). So, some modified methods center l and radius R, just as shown in Eqs. (1) and (2).
were proposed to address multi-classification problem.
X
n
One-against-all SVM (OAA-SVM) is one of the earliest and sim- Min FðR; l; ni Þ ¼ R2 þ C ni ; ð1Þ
plest multiple-machine approaches. Recently, a reduced OAA-SVM i¼1
(R-OAA-SVM) was proposed based on sample selection to decrease
the time consumption of OAA-SVM (KreBel, 1999). under the constraints
One-against-one SVM (OAO-SVM) is another popular multiple- kuðxi Þ  lk2 6 R2 þ ni ; 8i ¼ 1; . . . ; n ð2Þ
machine approach (Guermeur, 2002). For a C-class classification
problem, it usually constructs C(C-1)/2 binary SVMs, where each where parameter C gives the trade-off between the volume of
binary SVM is trained on data points from two classes. Two inter- sphere and the accuracy of data description, and ni P 0 are slack
esting modifications on OAO-SVM are directed acyclic graph SVM variables allowing for soft boundaries.
(DAG-SVM) (Hsu & Lin, 2002) and adaptive directed acyclic graph By introducing the Lagrangian, we have
SVM (ADAG-SVM) (Bredensteiner & Bennett, 1999). X
n X
n X
n

In order to improve generalization performances and reduce LðR; l; ni ; bi ; ai Þ ¼ R2 þ C ni  bi ðR2 þ ni  kuðxi Þ  lk2 Þ  ai ni ð3Þ
i¼1 i¼1 i¼1
time consumption, some OAO-based binary tree SVM algorithms
were proposed (Kumar & Gopal, 2011). An important difference where ai P 0,bi P 0 are Lagrange multipliers.
2086 X. Guo et al. / Expert Systems with Applications 41 (2014) 2083–2097

Original Samples

Preprocess Using Semi-fuzzy Kernel Clustering algorithm

New Data Points


on Supplier Selection
Fuzzy Samples Deterministic Samples

Expert Judgment

Training Samples

Classifier Establishment

Support Vectors Classifier Applied in Multi-Classification Classifier for Supplier


Selection

The First Step

The Second Step

Rules Extraction Using CCEA Explanations to the Results

Fig. 1. Framework of the proposed methodology.

@L @L
Setting @R
¼ 0 and @l
¼ 0 respectively, leads to In terms of the multi-classification problem, suppose a set of
X
n training samples fðxi ; yi Þgni¼1  v  w is given, where xi 2 v denotes
bi ¼ 1 ð4Þ an input pattern and yi 2 w denotes its output class respectively.
i¼1 The central idea is utilizing the information of the domain descrip-
tion generated by SVDD to estimate the distributions of each par-
and
titioned class data and then accomplish multi-classification (Zhu,
X
n
Chen, & Liu, 2003; Lee & Lee, 2005).
l¼ bi uðxi Þ ð5Þ
i¼1

By eliminating the variables R, l, ni and ai, the Lagrangian can be 3.2.2. Sample preprocessing based on semi-fuzzy kernel clustering
turned into the Wolfe Dual form, just as shown in Eqs. (6) and (7). Just as mentioned above, establishment of sphere according to
the training set is essential to multi-classification based on SVDD
X
n X
n
(Lee & Lee, 2007; Wang & Lai, 2013). A sphere is only determined
max W ¼ bi Kðxi ; xi Þ  bi bj Kðxi ; xj Þ ð6Þ
bi
i¼1 i;j¼1
by support vectors, but all the samples will have to be taken into
X
n consideration in standard algorithms (Tax & Duin, 1999, 2004;
subject to bi ¼ 1; 0 6 bi 6 C 8i ¼ 1; . . . ; n ð7Þ Lee & Lee, 2005). So, if the samples, which are more liable to be-
i¼1 come support vectors, can be identified in advance, the size of
training set will be reduced to cause lower complexity and compu-
where the dot products u(xi)u(xj) are replaced by the correspond-
tational consumption. Just as shown in Fig. 2, the sphere can be
ing kernel function K(xi,xj).
determined according to the samples in the Edge Zone. On the
Moreover, the kernel radius function, which is defined by the
other hand, in terms of practical supplier selection, it is difficult
Euclidian distance of u(x) from l, is given as Eq. (8).
to determine whether the samples located in this area belong to
RðxÞ ¼ kuðxÞ  lk the class or not, and supports from experts are necessary.
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi To accomplish Edge Zone detection, semi-fuzzy kernel cluster-
u X n X n
u ing algorithm, which is developed based on standard Semi-fuzzy
¼ tKðx; xÞ  2 bi Kðxi ; xÞ þ bi bj Kðxi ; xj Þ ð8Þ
i¼1 i;j¼1 Clustering algorithm (Selim & Ismail, 1984; Kamel & Selim,
1991), is employed.
So, the radius of the sphere is defined as Eq. (9). Application of this algorithm is introduced as follows.
Let lji 2 [0, 1] denote the membership of xi, i = 1, . . . ,n, to the
R ¼ maxfRðxi Þjxi is a support vectorg ð9Þ j-th class, whose center of clustering is vj, where j = 1, . . . ,c and
X. Guo et al. / Expert Systems with Applications 41 (2014) 2083–2097 2087

Based on the results derived using semi-fuzzy kernel clustering


algorithm, the original samples can be divided into two subsets:
Deterministic Samples and Fuzzy Samples, according to the follow-
ing partition rule.
If $lji = 1 for j 2 {1, . . . ,c}, then xi is a Deterministic Sample,
else xi is a Fuzzy Sample. Sets of Deterministic Samples and Fuzzy
Samples are represented as v1 and v2 respectively.
If xi 2 v1, xi can be classified as data point belonging to class j
directly without extra judgment.
Otherwise, it is a fuzzy sample, and difficult to categorize due to
fuzzy memberships. Therefore, expert consideration and judgment
is necessary.

3.2.3. Procedures for classifier establishment


Therefore, the procedures for training set generation can be
concluded as follows.
Fig. 2. Sphere specification in kernel space.

Step 1. Collect raw data to generate original samples depicted as


vectors, which are represented according to criteria, just as
m 2 (1, 1) denotes the weighting exponent. The mathematical
illustrated in Fig. 3.
formulation (Bezdek, 1980) of Semi-fuzzy Kernel Clustering
Step 2. Accomplishing sample preprocessing, divide them into
model takes the form shown as Eq. (10) by introducing kernel
two subsets: Deterministic Samples and Fuzzy Samples, just
function K(,).
as illustrated in Fig. 4. Classification labels of deterministic sam-
X
c X
n
ples are assigned using semi-fuzzy kernel clustering algorithm,
min JðU; VÞ ¼ ðlji Þm  kUðxi Þ  Uðv j Þk2
while the ones of fuzzy samples remain null.
j¼1 i¼1
Step 3. Generate training set by assigning classification labels of
X
c X
n
¼ ðlji Þm ðKðxi ; xi Þ  2Kðxi ; v j Þ þ Kðv j ; v j ÞÞ ð10Þ fuzzy samples based on expert judgment, just as illustrated in
j¼1 i¼1 Fig. 5(a).
X
c
S:t lji ¼ 1; i ¼ 1; . . . ; n
j¼1 Just as shown in Fig. 5, different from conventional methods,
training set is derived from part samples, rather than all original
Some other functions can be expressed as Eqs. (11) and (12). ones. Therefore, experts can focus on considering categories of re-
 m1
1 duced samples.
1
dji
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Moreover, fuzzy samples can be identified to be located in Edge
lji ¼  m1 ; where dji ¼ Kðxi ;xi Þ  2Kðxi ; v j Þ þ Kðv j ; v j Þ ð11Þ
Pc 1
1
Zone, and are more liable to be selected as support vectors, just as
j¼1 dji shown in Fig. 6.
Pn
i¼1 ð lji Þm xi So, the sphere can be specified according to training samples,
vj ¼ Pn m ð12Þ which are located in Edge Zone, to reduce computational consump-
i¼1 ðlji Þ
tion, but also achieve better performance. Then, procedures of mul-
So, the procedures can be concluded as follows: ti-classification using Semi-fuzzy SVDD can be concluded as
follows.
Step 1. Set iterative stopping condition e 2 (0, 1), maximum
iterating time T, and parameter k 2 (0, 1]. Step 1. Generate training set according to the procedures for
Step 2. Initialize clustering center V = {v1,v2, . . . ,vc}. training set generation.
P  m1 1

Step 3. Compute dji, and Let Li ¼ cj¼1 d1ji ; i ¼ 1; 2; . . . ; n.


Step 4. Set dpi = min{dji, "1 6 i 6 n}
For each data point, 1 6 i 6 n, do

1; j ¼ s
If $dji = 0 for j = s then set uji ¼ Original Samples Rrepresented According to Criteria
0; j – s; 81 6 j 6 c Classification Labels
 m1
1

Else If
1
dpi

Li
1; j ¼ p
P k, then uji ¼ x11 x12 x13 x14 x1j NULL
0; j – p; 81 6 j 6 c
Else compute lji according to Eq. (11) x21 x22 x23 x24 x2j NULL
End if
End if x31 x32 x33 x34 x3j NULL
End do
Step 5. Compute new lji x41 x42 x43 x44 x4j NULL

Step 6. If max jlji  l ji j < e or maximum iterating time is
reached, then terminate x51 x52 x53 x54 x5j NULL
Ă

Else compute new clustering center


 
V ¼ v 1 ; v 2 ; . . . ; v c according to Eq. (12), and turn to
Step3. xi1 xi2 xi3 xi4 Ă xij NULL
End if
Fig. 3. Graphical representation of original samples.
2088 X. Guo et al. / Expert Systems with Applications 41 (2014) 2083–2097

The results of rule extraction in classification can be repre-


sented by a set of rules containing at least one prototype for each
class (Stoean & Stoean, 2009). The decomposition of each potential
problem is performed by assigning to each population the task of
establishing rule(s) for one certain class. In the research, a rule is
considered to be a first order entity shown in Eq. (13), where
e1,e2, . . . ,ew are the attributes of every sample, h1,h2, . . . ,hw
are the values in the definition domain of each indicator, and
k 2 {1, 2, . . . , c} denotes the class.
IF ðe1 ¼ h1 Þ AND . . . ANDðew ¼ hw Þ THEN class ¼ k ð13Þ

3.3.1. Procedures for acquiring classification rules

(a) Training stage


As the task is to build c rules, one for each class, c popula-
tions are considered, each with the purpose of evolving
one of the c individuals.
Representation
Each individual, which represents one rule, in each popula-
tion follows the same encoding as a sample from the data
Fig. 4. Samples partition using semi-fuzzy kernel clustering algorithm.
set. And individuals representing simple IF–THEN rules have
the condition parts in the attribute space and the conclu-
Step 2. Accomplish sphere specification based on the set of sions in the class space.
training samples obtained in Step 1 for each class. Initialization
Step 3. Accomplish multi-classification of the testing samples The values of genes of all individuals are initialized following
based on classifier (Zhu et al., 2003; Lee & Lee, 2005) estab- a uniform distribution intervals of the corresponding attri-
lished in Step 2. butes in the data set. In case the considered data set is nor-
malized, the values for the genes are initialized in the [0, 1]
3.3. Rule extraction method using CCEA interval, again following a uniform distribution.
Fitness evaluation
Based on the results obtained using Semi-fuzzy SVDD, rules To evaluate an individual from a certain population, a collab-
extraction method based on CCEA (Potter & De Jong, 2000), labeled orating rule from each of the other populations is selected
as CC-Rule method, is developed in the section. according to the collaborator selection pressure choice for

Fig. 5. Comparison of proposed training set generation method and conventional method.
X. Guo et al. / Expert Systems with Applications 41 (2014) 2083–2097 2089

Sphere Surface

Center Vj Center Vj

Deterministic Sample
Support Vector
Fuzzy Sample

(a) Samples Partition According to Class j (b) Sphere Specification Based on Fuzzy Samples

Fig. 6. Sphere specification in semi-fuzzy SVDD.

a number of times equal to Collaboration Pool Size (cps). In (c) Testing stage
addition to the classical CCEA types (optimistic assignment, After the stop condition is reached, it is disposed of c popu-
pessimistic assignment and hedge assignment), a novel lations of rules that had been evolved against the training
assignment, which can be seen in Appendix, is applied in set. In order to form a complete set of rules, an item from
the research. Departing from the current individual, for each each population has to be chosen. Rules may be selected ran-
sample s, multiple sets of rules are formed and applied in domly, the best ones may be considered or a selection
order to predict its class. All rules within a set have different scheme may be applied.
outcomes. Scores are computed for each of the possible out-
comes, in the following manner: When a rule set is applied 3.3.2. Application in rule extraction
to a sample, a certain outcome is established for it. The score If pedagogical method (Martens, Baesens, Gestel, & Vanthienen,
of that outcome is increased by unity. Each of the cps sets of 2007) is followed, fidelity is calculated in Eq. (17) and expressed as
rules is applied to s. Finally, the class of s is concluded to be the percentage of the identically labeled samples (Huysmans,
the class that obtains the highest score. The fitness of the Baesens, & Vanthienen, 2006), where xi is the current observation,
considered rule is computed as the correctly labeled cases yci is its outcome obtained by Semi-fuzzy SVDD, and yi is the one
over the total number. reached by CCEA. Contrary, if the behavior is decompositional
Independently of the chosen algorithm for evaluating rule (Martens et al., 2007), the known labels are taken as the real ones
performance, the distance between individuals and samples from the samples that represent support vectors, just as shown in
has to be computed to determine which rule is closest to Eq. (18), while accuracy is therefore calculated as Eq. (19).
each sample. In the paper, normalized Manhattan was 
adopted for distance calculation, where xi = (xi1, xi2, . . . ,xiw) fidelity ¼ Prob yci ¼ yi jxi 2 X ð17Þ

represents a sample from the training set, while h = (h1, fidelity ¼ Prob yreal ¼ yi jxi is a support vector ð18Þ
i
h2, . . . ,hw) represents an individual (or rule), and al, bl denote  real
accuracy ¼ Prob yi ¼ yi jxi 2 X ð19Þ
the lower, upper bounds of the l-th attribute, just as shown
in Eq. (14). In the study, the set of support vectors identified using Semi-
fuzzy SVDD is given as training set for rule extraction. So, decom-
X
w
jhl  xil j positional method is followed, and Eqs. (18) and (19) should be
dðh; xi Þ ¼ ð14Þ applied.
l¼1
bl  al
Overall, hybridization of Semi-fuzzy SVDD and CC-Rule method
(b) Select and variation operators can be illustrated in Fig. 7.
In the study, intermediate recombination is applied, with
two parents p and q selected. The value of gene i of the off- 4. Experiments
spring o is obtained according to Eq. (15), where r is a uni-
formed distributed random number over [0, 1]. 4.1. Data sets

oi ¼ pi þ r  ðqi  pi Þ ð15Þ A data set, from one of the largest household appliances
manufacturing companies in China (Song, 2001; Sun, 2008),
Mutation with normal perturbation is considered, and a statistically was introduced and labeled as Data Set A. Because of its business
chosen gene i of an individual p is changed according to Eq. (16). activities, the company needs a certain amount of raw materials
and has to coordinate with a large number of suppliers in Chinese
bi  ai mainland. This company has a committee in supply chain manage-
pi ¼ pi þ r  ð16Þ
ms ment (SCM) unit including about 45 members, who decide the
choice of supplier. So, 31 active members were selected as experts
where r is a random number with normal distribution, bi and ai are for interview and inquiry.
the upper and lower bounds of the i-th attribute in the data set and This research considered some previous literatures to conclude
ms is the mutation strength parameter. some alternative items for further argument. Besides review
2090 X. Guo et al. / Expert Systems with Applications 41 (2014) 2083–2097

Fig. 7. Hybridization of semi-fuzzy SVDD and CC-Rule method.

literature (Kannan, Sivakumar, Joseph, & Murugesan, 2013), some were used to test the proposed methodology. Therefore, 672
other researches on supplier selection of similar enterprises in samples were employed in the experiments. After raw data on
context of China were referred to (Qian, 2011; Li, Ming, & Zhong, samples were distributed to the experts, Delphi method was
2011; Fan & Wang, 2012; Tan, Wang, & Jiao, 2012; Wu & Cao, applied to determine their categories. A consensus process
2012; Gong & Hong, 2012; Ji & Zhang, 2013; Xiao & Guo, 2013), was performed through the questionnaires. Each inquiry was
and the preferred evaluation criteria were listed in Table 1. accompanied by providing information of proceeding round of
Based on the preliminary list of items derived from reference replies. Experts were encouraged to change previous replies
materials mentioned above and the case company’s opinions, a regarding those of others. Based on the procedures mentioned
Delphi expert questionnaire was designed and released to the ex- above, the samples were all categorized, and the data set was
perts to determine the criteria. If the opinions are not consistent, derived.
the Delphi questionnaire must be modified and reissued until the Moreover, another data set having been argued in Guo et al.
results are converged. The questionnaire was based on Likert scale (2009), was also introduced and labeled as Data Set B. Perfor-
to represent importance of each criterion. The criterion must be mances in Data Set A and B were both measured with a seven-
omitted if it has lower total score, and can be preserved otherwise. point scale, just as listed in Table 3.
The final criteria has five dimensions including ‘‘Quality,’’
‘‘Cost,’’‘‘Delivery,’’ ‘‘Service,’’ and ‘‘Corporate social responsibility’’, 4.2. Experiments set
which were concluded in Table 2.
All indicators listed in Table 2 were applied as the factors in In terms of the proposed methodology, whether it produces via-
supplier selection. The samples, generated from historical data, ble results and how appropriate parameters should be determined

Table 1
Criteria for supplier selection.

Research Evaluation items


Qian (2011) Finance, Consistency, Relationship, Flexibility, Technology Capability, Service, Reliability, Price
Li et al. (2011) Quality Level, Production Capacity, Lead Time, Storage Capacity
Fan and Wang (2012) Quality, Delivery Performance, Price, Reaction to Demand Change, Service of Support
Tan et al. (2012) Quality, Delivery On Time, Price, Service
Wu and Cao (2012) Quality, Delivery, Price, Technical Capability, Financial Position, Past Performance Attribute, Facility, Flexibility, Service
Gong and Hong (2012) Cost, Quality, Reliability of Delivery, Lead Time, Delivery on Time
Ji and Zhang (2013) Cost, Quality, Delivery, Service, Social Responsibility, Response
Xiao and Guo (2013) R&D, Cost, Quality, Service, Response
X. Guo et al. / Expert Systems with Applications 41 (2014) 2083–2097 2091

Table 2
Criteria involved in data set A.

Category Definition
Quality
Q1 Reject rate Percentage of rejected incoming material by quality control
Q2 Management commitment Degree of importance given to quality function in management
Q3 Process improvement The application of process improvement activities
Q4 Warranties and claim policies Existence of warranties and claim policies provided by supplier
Q5 Quality assurance The attainment of quality assurance
Cost
C1 Price performance value Supplier’s general delivery and quality performance for its product price
C2 Compliance with sectoral price behavior Consistence of price level of supplier with the sectoral average
C3 Transportation cost Sum of unit variable and allocated fixed transportation costs occurring as a result of product supply
Delivery
D1 Order fulfill rate Compliance with the predetermined order quantities
D2 Lead time Time between placement and arrival of an order
D3 Order frequency Frequency of orders
Service
R1 Responsiveness The ability of scheduling and changing orders
R2 Stock management Compliance of supplier’s stock management system with production line
R3 Willingness Willingness to share expertise and resolve conflicts
R4 Design capability Capability of developing new designs and speed of development
Corporate social responsibility
E1 The interests and rights of employees Labor relations, human rights and interests of employees
E2 The rights of stakeholders The interests and right of shareholders, such as consumers, communities and other related ones
E3 Information disclosure Non-financial information disclosure
E4 Respect for the policies Comply with local regulations and policies

Table 4
Values for the parameters.
Table 3
Parameter Value
Scale of performance.
Classifier Kernel function Gaussian Kernel
Scale Abbreviation Threshold (k) 0.9
Very poor VP CC-Rule method Population size 100
Poor P Recombination probability 0.5
Medium Poor MP Mutation probability 0.6
Fair F Mutation strength 0.01
Medium good MG Number of generations 120
Good G Collaboration pool size 2
Very good VG

Table 5
was argued. In the experiments, the values for parameters, listed in Experimental results of data set A.
Table 4, were tuned manually. Data set Classifier CC-Rule method CC
In the research, all features of data points ranged form 0 to 1
Train (%) Valid (%) Fidelity (%) Valid (%)
after Min–max transformation. The experiments were conducted
on platform of Matlab 7.0. To evaluate the prediction performance, 1 71.21 69.17 92.31 68.21 66.17
2 69.17 68.13 93.15 66.26 63.12
10-fold cross validation, which has shown good performance in 3 70.18 70.01 95.21 69.17 63.27
model selection, was followed, and about ten percent of the data 4 74.28 72.29 95.91a 70.11 66.74
points for each class were selected as testing samples to validate 5 80.91 79.99a 95.86 76.17 70.38
the proposed methodology. So, the results discussed in the paper 6 81.12a 79.99a 94.17 75.18 69.88
7 75.68 74.39 92.11 73.12 72.91a
were all shown as average values.
8 76.37 75.89 93.10 74.89 70.26
9 79.49 78.09 92.18 76.99a 72.17
10 72.09 71.15 92.82 70.17 68.91

4.3. Experimental results Avg. 75.05 73.91 93.68 72.03 68.38


a
Denoted the best performance.
To compare the classification accuracy of each algorithm, hit-
ratio, which is defined according to the samples classified cor-
rectly, is applied. Classification accuracies of training set were la-
beled as ‘‘Train’’, while ‘‘Valid’’ denoted those of testing samples. Then, the proposed methodology was compared with some
Moreover, fidelity, labeled as ‘‘Fidelity’’, was also computed based other Multi-class SVM (MSVM) techniques, which were listed in
on the set of support vectors. Additionally, the CCEA method for Table 7.
classification (Potter & De Jong, 2000; Stoean & Stoean, 2009), As shown in Table 7, the proposed methods based on thoughts
labeled as CC, was also applied to test how the proposed of hyper-sphere, achieves better performance than conventional
classifier helps it to boost performance. The results were shown MSVM models. Moreover, as one of modified models, some results
in Tables 5 and 6. implied that the proposed methodology has better generalization
2092 X. Guo et al. / Expert Systems with Applications 41 (2014) 2083–2097

Table 6 much fewer samples to determine their categories in practical sup-


Experimental results of data set B. plier selection process.
Data set Classifier CC-rule method CC Furthermore, sample preprocessing is accomplished using
Train (%) Valid (%) Fidelity (%) Valid (%) semi-fuzzy kernel clustering algorithm. Since the Edge Zone is
determined by threshold k, greater value of k will lead to more can-
1 70.01 68.14 93.41 67.29 65.87
2 70.17 69.93 94.15 66.86 64.28
didates for support vectors. Especially, when k = 1, the algorithm
3 72.15 70.99 96.29a 70.17 62.27 will be transformed to standard SVDD because almost all data
4 75.66 73.19 94.91 71.17 64.29 points are candidates for support vectors. Hence, model with too
5 82.01a 77.70 96.13 77.17a 69.09 large threshold may be little different from standard SVDD, while
6 80.92 79.98a 95.23 74.38 68.14
a too small value will have poor ability due to lack of essential
7 79.78 74.69 94.11 72.91 71.99a
8 75.47 73.89 92.19 73.99 69.98 training samples. Thus, issues on choosing the appropriate thresh-
9 80.49 79.11 93.16 75.81 71.99 old were discussed by empirical trials.
10 74.09 73.28 94.87 71.26 69.92 In the following experiments, the model with various threshold
Avg. 76.08 74.09 94.45 72.10 67.78 values was tested based on different data sets, just as shown in
a Figs. 9 and 10.
Denoted the best performance.
The results on Data Set A illustrated in Fig. 9(a) showed that
ability and much less computational complexity, which can be par- the proposed method achieved the best performance with the
tially measured with training time labeled with ‘‘Time’’, than stan- threshold of 0.8, because some outliers can be eliminated from
dard SVDD. the original data set to improve generalization ability. Just as
Especially, a statistical method t-test was employed based on shown in Fig. 10, with decline of threshold, more samples were
the results of CC and CC-Rule method, and the difference between eliminated from training set to reduce training time. Hence,
the results (Valid) obtained using them proved to be significant. smaller values of threshold will lead to less computational con-
Therefore, it can be concluded that the proposed classifier helps sumption, while classification accuracy may be decreased due to
CC to boost performance, while the latter can provide explanations lack of necessary training samples. Overall, threshold selection,
to the classification results. which involves complex tradeoffs between computational con-
As far as Semi-fuzzy SVDD was concerned, subsets of determin- sumption and classification accuracy, is essential to the proposed
istic samples and fuzzy samples derived using sample preprocess- methodology.
ing method were considered, and the distribution can be shown in On the other hand, experimental results on CC-Rule method
Fig. 8. were discussed in details. It can be noticed the amount of samples
The analysis results shown in Fig. 8 implied that less than half of representing support vectors, was larger than the misclassified
the samples are selected as fuzzy samples (44% in data set A and training samples, but significantly less than the entire training
48% in data set B). That is, experts will be required to consider set, just as shown in Fig. 11. The results implied that a major reduc-

Table 7
Experimental results of different methods.

Type Technique Data set A Data set B


Valid (%) Time (second) Valid (%) Time (second)
Prior AI approach GA Yeh and Chuang (2011) 64.71 4.21 65.77 3.98
NN Wu (2009) 65.25 2.15 66.71 2.09
GST Sadeghieh et al. (2012) 71.28 3.28 72.65 3.16
RST Chang and Hung (2010) 71.27 3.31 72.15 3.26
ACA Tsai et al. (2010) 69.29 4.31 70.18 3.68
CC Stoean and Stoean (2009) 68.38 4.19 67.78 3.96
Conventional MSVM One-Against-All 72.16 3.61 73.11 3.31
One-Against-One 72.12 2.57 72.19 2.12
H-PSVM Guo et al. (2009) 73.03 2.41 73.19 1.99
The sphere-based classifier Standard SVDD 73.78 1.14 74.08 1.11
Proposed method classifier 73.91 1.09 74.09 1.02
CC-Rule method 72.03 1.78 72.10 1.17

The training time of CC-Rule method was the time consumed in rules extraction phase, and Valid denoted the accuracy in terms of the testing samples.

298.9, 44% 245.2, 48%

373.1, 56% 266.8, 52%

Deterministic Sample Fuzzy Sample


Deterministic Sample Fuzzy Sample

(a) Data Set A (b) Data Set B

Fig. 8. Partition of samples after preprocessing.


X. Guo et al. / Expert Systems with Applications 41 (2014) 2083–2097 2093

Fig. 9. Experimental results of generalization ability on data sets.

Fig. 10. Experimental results of training time on data sets.

168
122
288 188

216
202

Misclassified T raining Samples Support Vectors Other T raining Samples Misclassified T raining Samples Support Vectors Other T raining Samples

(a) Results on Data Set A (b) Results on Data Set B

Fig. 11. Partition of samples prior to rule extraction.

tion in number of training samples for rule extraction was per- time and accuracy. Another important observation was that, when
formed prior to application of CC-Rule method. a certain threshold (cps = 6 for Data Set A and 5 for Data Set B) was
As shown in Fig. 12, the cps parameter directly influenced the surpassed, no further gain in accuracy was reached with much
runtime of program that implements the approach. The final test- more training time consumption.
ing accuracy was positively affected by the increase in this value at Besides the decompositional method, pedagogical method is
the same time, but a balance had to be established between run- another one available for rule extraction. According to it, a
2094 X. Guo et al. / Expert Systems with Applications 41 (2014) 2083–2097

Fig. 12. Experimental results on CC-Rule method according to different cps parameters.

Fig. 13. Experimental results of performance on rule extraction methods.


X. Guo et al. / Expert Systems with Applications 41 (2014) 2083–2097 2095

correct data input–output mapping was established using all the This research was sponsored by Project 985–3 of Xi’an Jiaotong
training samples. So, it should also be important to compare University.
performances of the different methods, and the results can be
shown in Fig. 13. Appendix A
The fidelity values indicated that the pedagogical and decompo-
sitional methods both behave very similar to classifier based on Require: A rule m from a population
Semi-fuzzy SVDD, in the sense that the fidelity indicated relatively Ensure: The evaluation of the rule
high percents. High fidelity is very much desired, since it implies Begin
that the approach can learn the relationship between the values For each samples s in the training set do
of the attributes and the outcomes capably. Set the score for each possible outcome of s to 0;
Although pedagogical method proved to be more successful End for
than decompositional method in the results illustrated in For i = 1 to cps (Collaboration pool size) do
Fig. 13(a) and (b), the latter has a certain advantage. That is, it re- select a random collaborating rule from each population
duced by a great amount of training samples for rule extraction. different from that of c according to the collaborator
Just as shown in Fig. 13(c), this can be translated into a significant selection pressure parameter;
reduction in computational consumption. For each sample s in the training set do
find the rule r from the set of all collaborators that is
closest to s
5. Conclusions and discussions found class for s = class of r
increase the score of found class for s by one
A two-step hybridized method, by which classifier establish- End For
ment and rule extraction for supplier selection can be accom- End For
plished, is proposed in the paper. And combination of the success = 0;
advantages of two important tools has been achieved. For each sample s in the training set do
Semi-fuzzy SVDD has demonstrated its ability to solve multi- If the real class of s equals the class with the highest score
classification problems in supplier selection by using only for s then
necessary resources, with a solid mathematical background. From s is correctly classified;
experimental results, it can be concluded that generalization success = success + 1;
performance of the classifier based on Semi-fuzzy SVDD is better End If
than standard SVDD, which outperformed many kinds of conven- End For
tional MSVM algorithms argued in prior literatures. Moreover, as
accuracy ¼ number 100 success
of training samples
a modified sphere-based classifier, the method has much less
computational consumption than standard SVDD. End
CC-Rule method has shown its applicability in rule extraction.
Its performance was also tested, with pedagogical method com-
References
pared. The approach based on pedagogical method outperforms
CC-Rule method in classification accuracy, while fidelity values of Adler, N., Friedman, L., & Sinuany-Stern, Z. (2002). Review of ranking methods in the
the two methods behave very similar. But, the computational con- data envelopment analysis context. European Journal of Operational Research,
sumption is reduced by CC-Rule method. 140(2), 249–265.
Amid, A., Ghodsypour, S. H., & O’Brien, C. (2011). A weighted max-min model for
Accordingly, supplier selection process can be improved using
fuzzy multi-objective supplier selection in a supply chain. International Journal
the methodology. Besides explanations to the decisions, a new of Production Economics, 131(1), 139–145.
training set generation method is put forward. Only part of the Amin, S. H., Razmi, J., & Zhang, G. (2011). Supplier selection and order allocation
based on fuzzy SWOT analysis and fuzzy linear programming. Expert Systems
samples, rather than all the original ones, require expert judgment
with Applications, 38(1), 334–342.
to decide their categories and are selected as training ones to Amin, S. H., & Zhang, G. (2012). An integrated model for closed-loop supply chain
accomplish SVDD specification, which was illustrated in Fig. 5, configuration and supplier selection: Multi-objective approach. Expert Systems
with experimental results shown in Fig. 8. and Applications, 39(8), 6782–6791.
Awasthi, A., Chauhan, S. S., & Goyal, S. K. (2010). A fuzzy multi-criteria approach for
The underlying assumption of the proposed method is that evaluating environmental performance of suppliers. International Journal of
sufficient boundary points identified in advance, could support Production Economics, 126(2), 370–378.
a close boundary around the target data but too many ones Azadeh, A., & Alem, S. M. (2010). A flexible deterministic, stochastic and fuzzy data
envelopment analysis approach for supply chain risk and vendor selection
might cause over-fitting and poor generalization ability. So, the problem: Simulation analysis. Expert Systems with Applications, 37(12),
‘‘edge information’’ is exploited in the research. That is, the sup- 7438–7448.
port vectors, which are selected from samples located in Edge Bai, C., & Sarkis, J. (2010). Integrating sustainability into supplier selection with grey
system and rough set methodologies. International Journal of Production
Zone, are applied not only in classifier establishment, but also Economics, 124(1), 252–264.
in rule extraction. Hence, it is one of the promising directions Barakat, N. H., & Bradley, A. P. (2007). Rule extraction from support vector
to seek methods supporting sample selection, by which the can- machines: A sequential covering approach. IEEE Transactions on Knowledge and
Data Engineering, 19(6), 729–741.
didates of support vectors can be identified with some outliers
Barakat, N., & Bradley, A. P. (2010). Rule extraction from support vector machines: A
eliminated to achieve better generalization ability. Furthermore, review. Neurocomputing, 74, 178–190.
since the extracted rules are required to provide explanations Bezdek, J. C. (1980). A convergence theorem for the fuzzy ISODATA clustering
algorithm. Transactions on Pattern Analysis and Machine Intelligence, 12(1),
to the results, tradeoff among fidelity, accuracy, and comprehen-
235–251.
sibility should be argued. Bhattacharya, A., Geraghty, J., & Young, P. (2010). Supplier selection paradigm: An
integrated hierarchical QFD methodology under multiple-criteria environment.
Applied Soft Computing Journal, 10(4), 1013–1027.
Acknowledgment Bredensteiner, E. J., & Bennett, K. P. (1999). Multi-category classification by
support vector machines. Computational Optimization and Applications, 12(1–3),
The authors appreciate the valuable comments of editor and the 53–79.
Buyukozkan, G., & Cifci, G. (2012). A novel hybrid MCDM approach based on fuzzy
anonymous reviewers, which have positively contributed to the DEMATEL, fuzzy ANP and fuzzy TOPSIS to evaluate green suppliers. Expert
quality of the paper. Systems with Applications, 39(3), 3000–3011.
2096 X. Guo et al. / Expert Systems with Applications 41 (2014) 2083–2097

Castro, J. L., Flores-Hidalgo, L. D., Mantas, C. J., & Puche, J. M. (2007). Extraction of Ho, W., Xu, X., & Dey, P. K. (2010). Multi-criteria decision making approaches for
fuzzy rules from support vector machines. Fuzzy Sets and Systems, 158, supplier evaluation and selection: A literature review. European Journal of
2057–2077. Operational Research, 202(1), 16–24.
Celebi, D., & Bayraktar, D. (2008). An integrated neural network and data Hsu, B., Chiang, C., & Shu, M. (2010). Supplier selection using fuzzy quality data and
envelopment analysis for supplier evaluation under incomplete information. their applications to touch screen. Expert Systems with Applications, 37(9),
Expert Systems with Applications, 35(4), 1698–1710. 6192–6200.
Chai, J., Liu, J. N. K., & Ngai, E. W. T. (2013). Application of decision-making Hsu, C. W., & Lin, C. J. (2002). A comparison of methods for multi-class support
techniques in supplier selection: A systematic review of literature. Expert vector machines. IEEE Transactions on Neural Networks, 13, 415–425.
Systems with Applications, 40(10), 3872–3885. Huysmans, J., Baesens, B., & Vanthienen, J. (2006). Using rule extraction to improve
Chamodrakas, I., Batis, D., & Martakos, D. (2010). Supplier selection in electronic the comprehensibility of predictive models. Technical report KU Leuven.
marketplaces using satisficing and fuzzy AHP. Expert Systems with Applications, Ishizaka, A., Pearman, C., & Nemery, P. (2012). AHP sort: An AHP-based method for
37(1), 490–498. sorting problems. International Journal of Production Research, 50(17),
Chan, F. T. S., & Chan, H. K. (2010). An AHP model for selection of suppliers in the 4767–4784.
fast changing fashion market. International Journal of Advanced Manufacturing Jain, V., Wadhwa, S., & Deshmukh, S. G. (2009). Select supplier-related issues in
Technology, 51(9–12), 1195–1207. modeling a dynamic supply chain: Potential, challenges and direction for future
Chang, B., Chang, C., & Wu, C. (2011). Fuzzy DEMATEL method for developing research. International Journal of Production Research, 47(11), 3013–3039.
supplier selection criteria. Expert Systems with Applications, 38(3), 1850–1858. Ji, J., & Zhang, W. (2013). Suppliers selection based on emergent loss in supply chain.
Chang, B., & Hung, H. (2010). A study of using RST to create the supplier selection Science and Technology Management Research (8), 238–245.
model and decision-making rules. Expert Systems with Applications, 37(12), Kamel, M. S., & Selim, S. Z. (1991). A threshold fuzzy C-means algorithm for semi-
8284–8295. fuzzy clustering. Pattern Recognition, 24(9), 825–833.
Che, Z. H. (2010). A genetic algorithm-based model for solving multi-period supplier Kannan, G., Sivakumar, R., Joseph, S., & Murugesan, P. (2013). Multi-criteria decision
selection problem with assembly sequence. International Journal of Production making approaches for green supplier evaluation and selection: A literature
Research, 48(15), 4355–4377. review. Journal of Cleaner Production, 20, 1–18.
Chen, T. (2011). Bivariate models of optimism and pessimism in multi-criteria Kara, S. S. (2011). Supplier selection with an integrated methodology in unknown
decision-making based on intuitionistic fuzzy sets. Information Sciences, environment. Expert Systems with Applications, 38(3), 2133–2139.
181(11), 2139–2165. Keskin, G. A., Ilhan, S., & Özkan, C. (2010). The fuzzy ART algorithm: A categorization
Chen, Y. (2011). Structured methodology for supplier selection and evaluation in a method for supplier evaluation and selection. Expert Systems with Applications,
supply chain. Information Sciences, 181(9), 1651–1670. 37(2), 1235–1240.
Chen, Y., & Chao, R. (2012). Supplier selection using consistent fuzzy preference Kilincci, O., & Onal, S. A. (2011). Fuzzy AHP approach for supplier selection in a
relations. Expert Systems with Applications, 39(3), 3233–3240. washing machine company. Expert Systems with Applications, 38(8), 9656–9664.
Chen, L. Y., & Wang, T. (2009). Optimizing partners’ choice in IS/IT outsourcing KreBel, U. H. G. (1999). Pairwise classification and support vector machines. Advances
projects: The strategic decision of fuzzy VIKOR. International Journal of in Kernel Methods: Support Vector Learning. Cambridge, MA: The MIT Press, pp.
Production Economics, 120(1), 233–242. 255–268.
Chen, T., Wang, H., & Lu, Y. (2011). A multi-criteria group decision-making approach Kull, T. J., & Talluri, S. (2008). A supply risk reduction model using integrated multi-
based on interval-valued intuitionistic fuzzy sets: A comparative perspective. criteria decision making. IEEE Transactions on Engineering Management, 55(3),
Expert Systems with Applications, 38(6), 7647–7658. 409–419.
Chen, Y., Wang, T., & Wu, C. (2011). Strategic decisions using the fuzzy PROMETHEE Kumar, M. A., & Gopal, M. (2011). Reduced one-against-all method for multi-class
for IS outsourcing. Expert Systems with Applications, 38(10), 13216–13222. SVM classification. Expert Systems with Applications, 38, 14238–14248.
Chou, S., & Chang, Y. (2008). A decision support system for supplier selection based Labib, A. W. (2011). A supplier selection model: A comparison of fuzzy logic and the
on a strategy-aligned fuzzy SMART approach. Expert Systems with Applications, analytic hierarchy process. International Journal of Production Research, 49(21),
34(4), 2241–2253. 6287–6299.
Cortes, C., & Vapnik, V. N. (1995). Support-vector networks. Machine Learning, 20, Lai, C., Tax, D., Duin, R., Pekalska, E., & Paclik, P. (2004). A study on combining image
273–297. representations for image classification and retrieval. International Journal of
Dalalah, D., Hayajneh, M., & Batieha, F. (2011). A fuzzy multi-criteria decision Pattern Recognition and Artificial Intelligence, 18(5), 867–890.
making model for supplier selection. Expert Systems with Applications, 38(7), Lee, J., & Lee, D. (2005). An improved cluster labeling method for support vector
8384–8391. clustering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 27(3),
Demirtas, E. A., & Üstün, O. (2008). An integrated multi-objective decision making 461–464.
process for supplier selection and order allocation. Omega, 36(1), 76–90. Lee, D., & Lee, J. (2007). Domain described support vector classifier for multi-
Dogan, I., & Aydin, N. (2011). Combining Bayesian networks and total cost of classification problems. Pattern Recognition, 40(4), 41–51.
ownership method for supplier selection analysis. Computers and Industrial Lee, S. W., Park, J., & Lee, S. W. (2006). Low resolution faces recognition based on
Engineering, 61(4), 1072–1085. support vector data description. Pattern Recognition, 39(9), 1809–1812.
Falagario, M., Sciancalepore, F., Costantino, N., & Pietroforte, R. (2012). Using a DEA Levary, R. R. (2008). Using the analytic hierarchy process to rank foreign suppliers
cross efficiency approach in public procurement tenders. European Journal of based on supply risks. Computers and Industrial Engineering, 55(2), 535–542.
Operational Research, 218(2), 523–529. Liao, C., & Kao, H. (2011). An integrated fuzzy TOPSIS and MCGP approach to
Fan, C., & Wang, X. (2012). Supplier evaluation based on synthesis method of supplier selection in supply chain management. Expert Systems with
principal component analysis, data envelopment analysis and analytic Applications, 38(9), 10803–10811.
hierarchy. Journal of Tongji University, 40(12), 1899–1904. Li, X., Ming, Y., & Zhong, J. (2011). A study of supplier selection method based on
Ferreira, L., & Borenstein, D. (2012). A fuzzy-Bayesian model for supplier selection. QFD and Kano model. Journal of Systems & Management, 20(5), 589–594.
Expert Systems with Applications, 39(9), 7834–7844. Lin, R. (2012). An integrated model for supplier selection under a fuzzy situation.
Golmohammadi, D., & Mellat-Parast, M. (2012). Developing a grey-based decision- International Journal of Production Economics, 138(1), 55–61.
making model for supplier selection. International Journal of Production Lin, C., Chen, C., & Ting, Y. (2011). An ERP model for supplier selection in electronics
Economics, 137(2), 191–200. industry. Expert Systems with Applications, 38(3), 1760–1765.
Golmohammadi, D., & Mellat-Parast, M. (2012). Developing a grey-based decision Lin, R., Chuang, C., Liou, J. J. H., & Wu, G. (2009). An integrated method for finding
making model for supplier selection. International Journal of Production key suppliers in SCM. Expert Systems with Applications, 36(3 PART 2),
Economics, 137(2), 191–200. 6461–6465.
Gong, X., & Hong, Y. (2012). Method of supplier selection based on DEA and game Lin, Y., Lin, C., Yu, H., & Tzeng, G. (2010). A novel hybrid MCDM approach for
theory. Machinery Design & Manufacture (11), 223–225. outsourcing vendor selection: A case study for a semiconductor company in
Guermeur, Y. (2002). Combining discriminant models with new multi-class SVMS. Taiwan. Expert Systems with Applications, 37(7), 4796–4804.
Pattern Analysis & Applications, 5, 168–179. Lin, Y., & Yeh, C. (2010). Optimal carrier selection based on network reliability
Guneri, A. F., Ertay, T., & Yucel, A. (2011). An approach based on ANFIS input criterion for stochastic logistics networks. International Journal of Production
selection and modeling for supplier selection problem. Expert Systems with Economics, 128(2), 510–517.
Applications, 38(12), 14907–14917. Li, L., & Zabinsky, Z. B. (2011). Incorporating uncertainty into a supplier selection
Guneri, A. F., Yucel, A., & Ayyildiz, G. (2009). An integrated fuzzy-lp approach for a problem. International Journal of Production Economics, 134(2), 344–356.
supplier selection problem in supply chain management. Expert Systems with Martens, D., Baesens, B., & Gestel, T. V. (2009). Decompositional rule extraction from
Applications, 36(5), 9223–9228. support vector machines by active learning. IEEE Transactions on Knowledge and
Guo, X. S., Yuan, Z. P., & Tian, B. J. (2009). Supplier selection based on hierarchical Data Engineering, 21(2), 178–191.
potential support vector machine. Expert Systems with Applications, 36(3 PART Martens, D., Baesens, B., Gestel, T. V., & Vanthienen, J. (2007). Comprehensible credit
2), 6978–6985. scoring models using rule extraction from support vector machines. European
Haleh, H., & Hamidi, A. (2011). A fuzzy MCDM model for allocating orders to Journal of Operational Research, 183, 1466–1476.
suppliers in a supply chain under uncertainty over a multi-period time horizon. Ng, W. L. (2008). An efficient and simple model for multiple criteria supplier
Expert Systems with Applications, 38(8), 9076–9083. selection problem. European Journal of Operational Research, 186(3), 1059–1067.
He, J., Hu, H.-J., & Harrison, R. (2006). Rule generation for protein secondary Ordoobadi, S. M. (2010). Application of AHP and taguchi loss functions in supply
structure prediction with support vector machines and decision tree. IEEE chain. Industrial Management and Data Systems, 110(8), 1251–1269.
Transactions on Nanobioscience, 5(1), 46–53. Ozkok, B. A., & Tiryaki, F. (2011). A compensatory fuzzy approach to multi-objective
Ho, W., Dey, P. K., & Lockström, M. (2011). Strategic sourcing: A combined QFD and linear supplier selection problem with multiple-item. Expert Systems with
AHP approach in manufacturing. Supply Chain Management, 16(6), 446–461. Applications, 38(9), 11363–11368.
X. Guo et al. / Expert Systems with Applications 41 (2014) 2083–2097 2097

Pitchipoo, P., Venkumar, P., & Rajakarunakaran, S. (2012). A distinct decision model Tseng, M. (2011). Green supply chain management with linguistic preferences and
for the evaluation and selection of a supplier for a chemical processing industry. incomplete information. Applied Soft Computing Journal, 11(8), 4894–4903.
International Journal of Production Research, 50(16), 4635–4648. Tseng, M., Chiang, J. H., & Lan, L. W. (2009). Selection of optimal supplier in supply
Pontil, M., & Verri, A. (1998). Support vector machine for 3-D object recognition. chain management strategy with analytic network process and choquet
IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(6), 637–646. integral. Computers and Industrial Engineering, 57(1), 330–340.
Potter, M. A., & De Jong, K. A. (2000). Cooperative coevolution: An architecture for Vapnik, V. N. (1998). Statistical learning theory. NewYork: JohnWiley& Sons.
evolving co-adapted subcomponents. Evolutionary Computation, 8, 1–29. Vapnik, V. N. (1999). An overview of statistical learning theory. IEEE Transactions on
Punniyamoorthy, M., Mathiyalagan, P., & Parthiban, P. (2011). A strategic model Neural Networks, 10(5), 988–999.
using structural equation modeling and fuzzy logic in supplier selection. Expert Vinodh, S., Anesh Ramiya, R., & Gautham, S. G. (2011). Application of fuzzy analytic
Systems with Applications, 38(1), 458–474. network process for supplier selection in a manufacturing organization. Expert
Qian, Z. (2011). On the BP neural network and its application in supplier selection Systems with Applications, 38(1), 272–280.
and evaluation. Industrial Engineering and Management, 16(3), 1–7. Wang, C.-D., & Lai, J.-H. (2013). Position regularized support vector domain
Razmi, J., & Rafiei, H. (2010). An integrated analytic network process with mixed description. Pattern Recognition, 46(3), 41–51.
integer non-linear programming to supplier selection and order allocation. Wu, D. (2009). Supplier selection: A hybrid model using DEA, decision tree and
International Journal of Advanced Manufacturing Technology, 49(9–12), neural network. Expert Systems with Applications, 36(5), 9105–9112.
1195–1208. Wu, D. (2009). Supplier selection in a fuzzy group setting: A method using grey
Ren, Y., Liu, X., & Cao, J. (2011). A parsimony fuzzy rule-based classifier using related analysis and dempster-shafer theory. Expert Systems with Applications,
axiomatic fuzzy set theory and support vector machines. Information Sciences, 36(5), 8892–8899.
181, 5180–5193. Wu, D. D. (2010). A systematic stochastic efficiency analysis model and application
Rezaei, J., & Davoodi, M. (2012). A joint pricing, lot-sizing, and supplier selection to international supplier performance evaluation. Expert Systems with
model. International Journal of Production Research, 50(16), 4524–4542. Applications, 37(9), 6257–6264.
Sadeghieh, A., Dehghanbaghi, M., Dabbaghi, A., & Barak, S. (2012). A genetic Wu, T., & Blackhurst, J. (2009). Supplier evaluation and selection: An augmented
algorithm based grey goal programming (G3) approach for parts supplier DEA approach. International Journal of Production Research, 47(16), 4593–4608.
evaluation and selection. International Journal of Production Research, 50(16), Wu, J., & Cao, Q. (2012). A method for choosing green supplier under uncertain
4612–4630. decision—making. Operations Research and Management Science, 21(1), 220–225.
Saen, R. F. (2010). Developing a new data envelopment analysis methodology for Xiao, H., & Guo, P. (2013). Research on supplier selection based on principal
supplier selection in the presence of both undesirable outputs and imprecise component analysis and BP neural network. Computing Technology and
data. International Journal of Advanced Manufacturing Technology, 51(9–12), Automation, 32(1), 130–133.
1243–1250. Xu, J., & Ding, C. (2011). A class of chance constrained multi-objective linear
Selim, S. Z., & Ismail, M. A. (1984). Soft clustering of multidimensional data: a semi- programming with bi-random coefficients and its application to vendor
fuzzy approach. Pattern Recognition, 17(5), 559–568. selection. International Journal of Production Economics, 131(2), 709–720.
Sevkli, M. (2010). An application of the fuzzy ELECTRE method for supplier Xu, J., & Yan, F. (2011). A multi-objective decision making model for the vendor
selection. International Journal of Production Research, 48(12), 3393–3405. selection problem in a bi-fuzzy environment. Expert Systems with Applications,
Shaw, K., Shankar, R., Yadav, S. S., & Thakur, L. S. (2012). Supplier selection using 38(8), 9684–9695.
fuzzy AHP and fuzzy multi-objective linear programming for developing low Yang, P. C., Wee, H. M., Pai, S., & Tseng, Y. F. (2011). Solving a stochastic demand
carbon supply chain. Expert Systems with Applications, 39(9), 8182–8192. multi-product supplier selection model with service level and budget
Song, H. (2001). Cases on modern logistics and supply chain management. Beijing: constraints using genetic algorithm. Expert Systems with Applications, 38(12),
Economic Press. 14773–14777.
Stoean, C., & Stoean, R. (2009). Evolution of cooperating classification rules with an Yeh, W., & Chuang, M. (2011). Using multi-objective genetic algorithm for partner
archiving strategy to underpin collaboration. Studies in Computational Intelligence. selection in green supply chain problems. Expert Systems with Applications,
Intelligent Systems and Technologies – Methods and Applications (vol. 1). Springer, 38(4), 4244–4253.
pp. 47–65. Yucel, A., & Güneri, A. F. (2011). A weighted additive fuzzy programming approach
Sun, L. (2008). The advanced manufacturing mode of Chinese enterprise: T mode of for multi-criteria supplier selection. Expert Systems with Applications, 38(5),
Haier. Beijing: Tsinghua University Press. 6281–6286.
Sun, H.L., Xie, J.Y., & Xue, Y.F. (2005). An SVM-based model for supplier selection Yucenur, G. N., Vayvay, Ö., & Demirel, N. Ç. (2011). Supplier selection problem in
using fuzzy and pairwise comparison. In Proceedings of 2005 international global supply chains by AHP and ANP approaches under fuzzy environment.
conference on machine learning and cybernetics (vol. 6, pp. 3629–3633). International Journal of Advanced Manufacturing Technology, 56(5–8), 823–833.
Tan, Di., Wang, X., & Jiao, X.-C. (2012). Supplier selection based on uncertain Yu, M., Goh, M., & Lin, H. (2012). Fuzzy multi-objective vendor selection under lean
linguistic information. Industrial Engineering Journal, 15(3), 46–51. procurement. European Journal of Operational Research, 219(2), 305–311.
Tax, D. M. J., & Duin, R. P. W. (1999). Support vector domain description. Pattern Zeydan, M., Çolpan, C., & Çobanoglu, C. (2011). A combined methodology for
Recognition Letters, 20(11–13), 1191–1199. supplier selection and performance evaluation. Expert Systems with Applications,
Tax, D. M. J., & Duin, R. P. W. (2004). Support vector data description. Machine 38(3), 2741–2751.
Learning, 54(1), 45–66. Zhao, K., & Yu, X. (2011). A case based reasoning approach on supplier selection in
Tsai, W., & Hung, S. (2009). A fuzzy goal programming approach for green supply petroleum enterprises. Expert Systems with Applications, 38(6), 6839–6847.
chain optimization under activity-based costing and performance evaluation Zhu, M.L., Chen, S.F., & Liu, X.D. (2003). Sphere-structured support vector machines
with a value-chain structure. International Journal of Production Research, 47(18), for multi-class pattern recognition, Lecture Notes in Computer Science, vol.
4991–5017. 2639, pp. 589–593.
Tsai, Y. L., Yang, Y. J., & Lin, C. H. (2010). A dynamic decision approach for supplier Zhu, P., & Hu, Q. (2013). Rule extraction from support vector machines based on
selection using ant colony system. Expert Systems with Applications, 37(12), consistent region covering reduction. Knowledge-Based Systems, 42, 1–8.
8313–8321.

Вам также может понравиться