Computers & Industrial Engineering 69 (2014) 1220

Contents lists available at ScienceDirect

Computers & Industrial Engineering

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

Multi-criteria inventory classication with reference items q

Banu Soylu , Bahar Akyol
Department of Industrial Engineering, Erciyes University, 38039 Kayseri, Turkey

a r t i c l e

i n f o

Article history:
Received 24 March 2013
Received in revised form 1 December 2013
Accepted 17 December 2013
Available online 25 December 2013
Keywords:
Multiple criteria analysis
Sorting problem
Inventory classication

a b s t r a c t
The ABC method is a well-known approach to classify inventory items into ordered categories, such as A,
B and C. As emphasized in the literature, it is reasonable to evaluate the inventory classication problem
in the multi-criteria context. From this point of view, it corresponds to a sorting problem where
categories are ordered. Here, one important issue is that the weights of the criteria and categorization
preferences can change from industry to industry. This requires the analysis of the problem in a specic
framework where the decision maker (expert)s preferences are considered. In this study, the preferences
of the decision maker are incorporated into the decision making process in terms of reference items into
each class. We apply two utility functions based sorting methods to the problem. We perform an experiment and compare results with other algorithms from the literature.
2014 Elsevier Ltd. All rights reserved.

1. Introduction
The ABC method has been used by business rms to manage a
huge number of distinct items, referred to as stock keeping units
(SKUs). As the size of the inventory increases, controlling the items
needs time and additional expenditure. A reasonable strategy in
this case is to classify the SKUs in terms of inspection priority.
Therefore, the SKUs with high priority may be inspected more frequently to prevent stock-out cases and the resulting losses, while
the SKUs with low priority may be inspected less frequently in order to reduce inspection costs. This tradeoff reveals that each SKU
should have an inspection priority to minimize total inventory
costs. Inventory classication is not only related to inventory management but also related to production planning. Selection of the
appropriate production planning technique should also be considered together with the inventory classication.
Traditionally, the ABC analysis classies items into three categories, namely, A (very important), B (moderately important) and
C (least important) according to the annual usage value. Here,
the priority criterion is the annual usage value which determines
the class of the SKU. Since closely monitoring all items is too costly,
a suitable inventory management policy is to concentrate more on
class A items than on classes B and C items. Here, class C items also
receive a more relaxed control policy than class B items (Silver,
Pyke, & Preterson, 1998). However, authors agree that other
characteristics of the inventory like criticality, lead time, ordering
cost, commonality, repairability, and durability may affect and
q

This manuscript was processed by Area Editor Qiuhong Zhao.

Corresponding author. Tel.: +90 352 2076666x32456.

E-mail addresses: bsoylu@erciyes.edu.tr (B. Soylu), akyolbaharr@hotmail.com (B.

Akyol).
0360-8352/$ - see front matter 2014 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.cie.2013.12.011

change the class of items (Flores & Whybark, 1987; Gvenir & Erel,
1998; Partovi & Anandarajan, 2002; Ramanathan, 2006). To take
into account these characteristics we require multiple criteria decision analysis. Evaluating the problem in the multi-criteria context
may also improve inventory investments (see Bhattacharya, Sarkar, & Mukherjee, 2007, for a comparison study of traditional and
multi-criteria inventory classication via simulation). This problem has already been addressed in the literature. Recently, Ramanathan (2006) proposed a weighted linear optimization model
where an LP model maximizing the weighted sum of criteria for
the considered inventory item is solved and the resulting weights
are assigned as favorable weights of this item. After that, Zhou and
Fan (2007) presented an extended version of Ramanathans model.
Another weighted linear optimization model was given by Ng
(2007) and Hadi-Vencheh (2010).
One different study about this problem is the case-based distance model by Chen, Kevi, Marc Kilgour, and Hipel (2008). This
was originally a sorting procedure, which classies actions into ordered classes. The decision maker (DM) assigns in advance some
reference items into each class, and a mathematical model determines the thresholds and weights. (Fuzzy) AHP based approaches
have also been presented for this problem in the literature (Partovi
& Hopton, 1993; akr & Canbolat, 2008). Genetic algorithms
(Gvenir & Erel, 1998), particle swarm optimization (Tsai & Yeh,
2008), fuzzy classication (Chu, Liang, & Liao 2008; Keskin & zkan, 2013), TOPSIS (Bhattacharya et al., 2007) and neural networks
(Partovi & Anandarajan, 2002) have also been applied to this
problem.
All the studies mentioned above present different perspectives
for the multi-criteria inventory classication. Despite the advantages of these methods, it should be noted that the importance of criteria and the categorization may change from industry to industry

13

B. Soylu, B. Akyol / Computers & Industrial Engineering 69 (2014) 1220

and even from company to company. Several authors also emphasize this issue (Taylor, Sewart, & Bolander, 1981; Kampen, Akkerman, & Donk, 2012). For instance, the lead time criterion is more
important in the drug industry than in the furniture industry.
The adaptation of a methodology to these differences might be
possible by adjusting the problem parameters in accordance with
the decision maker (DM)s judgment. The DMs judgment can be
obtained directly by assigning values to the problem parameters
like weights, thresholds, etc. Alternatively, this information can
be indirectly inferred from some classication examples (see for
instance, Mousseau & Slowinski, 1998; Soylu, 2011). The latter
has some advantages since it is more understandable by the DM
and more agreeable in the case of multiple DMs.
In this study, we apply UTADIS (Doumpos & Zopounidis, 2004)
based sorting methods to the multi-criteria ABC inventory classication problem. We assume that the DM assigns in advance some
reference items for each class. It is assumed that these assignments
actually reect the characteristics of the industry involved. That
means the weights given to criteria in this industry can be inferred
from these category examples. A mathematical model constructs a
utility function and determines thresholds between classes with
respect to reference items. The aim of the model is to minimize
the total classication error over reference items. That means the
corresponding utility value of an item should place it in its correct
class otherwise the classication error occurs. Based on this information, unclassied items are placed manually. We evaluate two
function types, which are linear and piece-wise linear utility
functions.
The paper is organized as follows. We explain the utility function based approaches in Sections 2 and 3. We present computational results in Section 4. We conclude with a discussion and
further research directions in Section 5.

zji

If j 2 C Ref
then uA > Uzj P uB
B
If j 2 C Ref
then uB > Uzj
C
The following LP determines the utility function parameters, w, and
thresholds, uA and uB.

P
LP 1 :

Min f

ej
8j2C ref
A
jC ref
A j

ej
8j2C ref
B

jC ref
B j

ej

ej
8j2C ref
C
jC ref
C j

m
X
wi zji  uA ej P 0 8j 2 C ref
A

2
3

i1
m
X

wi zji  uB ej P 0 8j 2 C ref
B

i1
m
X

wi zji  uA  ej 6 d 8j 2 C ref

B

i1
m
X

wi zji  uB  ej 6 d 8j 2 C ref

C

5
6

i1
m
X

wi 1

i1

uA  uB P s

wi P 0 8i 1; 2; . . . ; m
uA ; uB P 0

In the rst approach, we utilize a classication scheme based on

a linear utility function. It is a specic version of the UTADIS method since we deal with the linear utility function rather than a
piece-wise utility function. Each SKU gets a single score from this
linear utility function and is classied by using the thresholds of
categories. The parameters of the linear utility function and thresholds are determined over reference items. For this purpose, a mathematical model is constructed and solved. The details of the
method are given below.
Without loss of generality, we assume that all criteria are of the
maximization type so the SKU with better utility is located in the
better class. SKUs are categorized into three ordered classes as CA, CB, and CC. Here CA refers to the class of SKUs with highest inspection priority and CC refers to the class of SKUs with lowest
inspection priority. Let S be the set of SKUs that should be classied, R be the set of reference SKUs and C Ref
n  C n be the set of
reference SKUs in class Cn, n = A, B, C. Let zji be the value of the ith
criterion i = 1, 2, . . . , m for SKU j e S.
Note that, in many multiple criteria problems, the ranges of
different criteria could be different. A scaling procedure might be
useful in this case. In this study, we use the following technique
to scale the ranges of criteria conveniently at the beginning of
the algorithm.

Z ji

If j 2 C Ref
then Uzj P uA
A

ej ; ej P 0 8 j 1; 2; . . . ; n

2. A linear utility function based approach

^zji  zmin
i
zmax
 zmin
i
i

In the proposed approach, a linear utility function

P
j
Uzj m
i1 wi zi is dened. Here wi is the weight of the ith criterion. Let uk be the utility threshold that distinguishes classes k
and k + 1 for k = A, B.
We expect the following three cases. Otherwise, the classication error occurs.

d 0:0001; s 0:0002 user defined small positiv e constants

The objective function of LP1 is to minimize the average classication errors over the reference SKUs. Constraint set (3) ensures that
the utility value of an SKU j 2 C ref
A should be greater than the threshold uA, otherwise a positive classication error, e
j , occurs. Constraint
sets (4) and (5) require that the utility value of an SKU j 2 C ref
B should
be in between thresholds uA and uB. Constraint set (6) ensures that
the utility value of an SKU j 2 C ref
C should be strictly less than the
threshold uB. Constraint (7) implies that the sum of all weights
should be 1 and constraint (8) requires that the threshold uA should
be strictly greater than the threshold uB. d1 and s are positive
constants to satisfy inequalities strictly. The classication errors
are explained using an example as follows. Assume that
Ref
Ref
C Ref
A fa1 ; a2 ; a3 g; C B fb1 ; b2 ; b3 g and C C fc 1 ; c2 g. If we classify them as in Fig. 1, corresponding errors occur. For instance, since
a3
a3 is classied wrongly in CC, the classication error e
a3 uA  uZ
exists.
In a specic case of the LP1, it may be possible to have a solution
with no classication error, i.e. objective value f = 0. This may happen if the DM chooses well-classied references. In this case,
thresholds may take alternative optimal values in the following
ranges.

min8j2C Ref fuzj g P uA P max8t2C Ref fuzt g d

A

where 2 0; 1 is the scaled value of ^zji , zmin

and zmax
are the minii
i
mum and maximum values of criterion i by considering SKUs in
S [ R.

min8j2C Ref fuzj g P uB P max8t2C Ref fuzt g d

B

In other words, if the utility of an SKU is in the rst range, this SKU
can be placed either in CA or CB. We will call this set as CA/CB.
Similarly, if the utility of an SKU is in the second range, this SKU

14

B. Soylu, B. Akyol / Computers & Industrial Engineering 69 (2014) 1220

Let C1, C2, . . . , Ch be the ordered classes, where C1 is the most

preferred class while Ch is the least preferred one. Let ui zji be
the marginal utility of alternative j on criterion i and U(zj) be the
global utility value of alternative j. The marginal utility function
of each criterion is assumed to be piece-wise linear. The range of
each criterion is divided into bi subintervals g pi ; g p1
, p = 1, 2, ... , bi
i
as illustrated in Fig. 2. A heuristic algorithm is used to specify
subintervals.
According to this, the marginal utility function of each criterion
is dened as given in the following equation:
r ji 1

ui zji

wip

p1

r
zji  g i ji
r ji 1
r
gi
 g i ji

wirji

h r r 1 i
to which the zji belongs, and
where rji is the subinterval g i ji ; g i ji
p1
p
wip ug i  g i .
The global utility function of alternative zj is given in the following equation:

Uzj

m
X
ui zji

10

i1

Fig. 1. Representation of classication errors.

can be placed either in CB or CC. We will call this set as CB/CC. One of
the following two strategies can be considered in this case.
Strategy 1: Select the middle of the range as a threshold. This
strategy divides the mid region into two equal parts.

uA

uB

min8j2C Ref fuzj g max8t2CRef fuzt g d

B

2
min8j2C Ref fuzj g max8t2CRef fuzt g d
B

The classication of the non-reference alternatives is then performed by comparing U(zj) with thresholds ud , d 2; 3; . . . ; h  1 as
follows:

If

Uzj P u1 then j 2 C 1

If

ud1 > Uzj P ud then zj 2 C d

If

uh1 > Uzj then zj 2 C h

Fig. 3 illustrates the thresholds between consecutive classes and

the reference set C ref
d  Cd.
According to this, the following LP is obtained after modifying
the UTADIS for the ABC classication problem.

Strategy 2: Request the DM to further classify some SKUs belonging

to CA/CB and CB/CC. This is an interactive strategy. In this case, we can
revise the optimal solution with new references. A parametric analysis may help to revise the solution, otherwise the model should be
solved from scratch again.
After determining the utility function and thresholds, a utility
value is computed for each non-reference SKU and placed into a
suitable category as follows:

Step 1: Compute Uzj

m
X
wi zji

UTADIS :

Min f

i1

Step 2: If Uzj P uA then j 2 C A

If uA > Uzj P uB then j 2 C B

jC ref
A j

ej
8j2C ref
B

jC ref
B j

ej

ej
8j2C ref
C
jC ref
C j

11

s:t:
Uzj  uA ej P 0 8j 2 C ref
A

12

C ref
B

13

ej

ej

ej

Uz  uB
Uz  uA 
Uz  uB 

8j 2 S

ej
8j2C ref
A

P 0 8j 2

6 d 8j 2

C ref
B

14

6 d 8j 2

C ref
C

15

bi
m X
X
wip 1

16

i1 p1

uA  uB P s

17

If Uz < uB then j 2 C C
ui

3. The UTADIS method

The LP1 is an easy-to-understand and easy-to-solve model.
However, it restricts the utility function to be linear. The UTADIS
method (Doumpos & Zopounidis, 2004) denes a piece-wise linear
utility function for each criterion. The global utility of an alternative is the sum of these partial utilities. Each alternative is assigned
to a suitable class by comparing its global utility with the threshold
of each class. The threshold values and the piece-wise linear utility
functions are estimated by using some reference alternatives
which are assigned to classes in advance. For this purpose, an LP
is proposed with the objective of minimizing average classication
errors over the reference set. Below, the piece-wise utility function
construction scheme of the UTADIS method is briey reviewed. The
reader may refer to Doumpos and Zopounidis (2004) for details of
the method.

u i (g ibi +1)

ui ( gi2 )

ui ( gi1 )
0.0

g i1

gi2

Fig. 2. A piece-wise marginal utility function.

b +1

gi i

Zi

15

B. Soylu, B. Akyol / Computers & Industrial Engineering 69 (2014) 1220

u1

u2

characteristics of inventory items in each class as the probability

of observing items with higher criteria values decreases sharply
in classes B and C.
We use two types of sets, Set 1 and Set 2, where there are 50
SKUs in each class in Set 1 and there are 10, 35 and 105 SKUs in
each class in Set 2 (in accordance with the ABC classication percentages), respectively. Note that we initially know the classes of
the SKUs due to the generation procedure mentioned above. All
SKUs are also classied by each algorithm and we compute the
error rate (ER) for each of them. We perform 30 runs/algorithm.

ER

P
j
# of misclassified SKUs
8j2S x

150
150
8
1;
if
the
class
of
an
SKU j and the class
>
<

where xj

u h1

Fig. 3. Illustration of ordered classes and thresholds.

wip P 0 8i 1; 2; . . . ; m; 8p 1; 2; . . . ; bi
ej ; ej P 0 8j 1; 2; . . . ; n
uA ; uB P 0
d 0:0001; s 0:0002 user defined small positiv e constants
The explanation of constraints is the same with LP1 except for the
utility function denition.
4. Computational results
In the rst part of this section, we design an experiment to analyze the classications by the R-model (Ramanathan, 2006), ZFmodel (Zhou & Fan, 2007), case-based model (Chen et al., 2008),
linear utility model and UTADIS (Doumpos & Zopounidis, 2004).
The brief review of rst three methods is given in Appendix A.
When running the UTADIS method, we use the Heuristic 2 to dene the number of subintervals (the reader may refer to Doumpos
& Zopounidis, 2004 for the denition of Heuristic 2). We used
Cplex 12.1 solver (Cplex, 2010) for the solution of all models. In
the second part of this section, we also apply the utility function
based models to the ABC classication problem given in the literature and discuss results.
4.1. Experimental analysis
For our experiments, we consider a system with 150 SKUs, and
we select three evaluation criteria, which are annual usage, average unit cost and lead time, from the literature. The criteria values
of each SKU in each class are generated articially from lognormal
distribution with the following mean and standard deviation.

Class CA ! lA 4;

rA 4
Class CB ! lB 2; rB 2
Class CC ! lC 1; rC 1

Lognormal distribution is used extensively in engineering,

economics, and biological/chemical sciences since it represents
well the occurrences of events. Since the distribution is skewed
with a long tail to the right, it may also represent the

>
:

17

assigned by the algorithm are different

0;

Otherwise

Since the linear utility model, UTADIS and the case-based model
require reference SKUs, all 150 SKUs are also assumed as references for these algorithms.
Table 1 presents the mean error rate and standard deviation results of each algorithm over 30 runs. We also test different k levels
of the ZF-model. According to these results UTADIS gives the best
performance in both sets. In the ZF-model, k = 0.25 is better than
other k levels. For further statistical comparisons, we apply the
paired-t test to analyze whether the mean ER of UTADIS is signicantly different from the mean ER of another method. Actually, we
test the following hypothesis.

H0 : lUTADIs lMethod
H1 : lUTADIs lMethod
In both sets, p-values are less than 0.05, which means the H0
hypothesis is rejected, concluding that UTADIS is signicantly better than the others except for the ZF-model k 0:50 in Set 2.
When we compare Set 1 and Set 2, we observe that the error rate
decreases signicantly in Set 2. This is because in Set 2 the cardinalities of classes A and B decrease but the cardinality of class C increases. As the sizes of two sets decrease, the probability of
misclassications due to classes A and B also decreases, i.e. misclassications due to A ? B, B ? A, A ? C, B ? C. However, since
the size of the class C increases, the probability of misclassications (due to C ? B, C ? A) increases. The decrement in misclassications is higher than the increment.
Additionally, we perform an analysis on inventory costs and ll
rate as presented in Teunter, Babai, and Syntetos (2010) and Lajili,
Babai, and Ladhari (2012). Here it is assumed that the system is
controlled with an (s, Q) policy. We compare the total holding cost
and ll rate for the classication of each method. The following
notation (Lajili et al., 2012) and corresponding settings are used
in inventory cost calculations.
N
n
Dj

rj
Lj
Qj
CSLj
kj

Number of SKUs in the inventory system (150)

Number of classes (3)
Demand of SKUj
Standard deviation of demand of SKUj. (1 * Dj)
Lead time of SKUj
q
2W j Dj
Order quantity of SKUj. Q j
hj
Cycle service level of SKUj. (99% j e CA, 95% j e CB, 90%
j e CC)
Safety factor for each SKUj.

U1 CSLj 2:326 J 2 C A ; 1:645 J 2 C B ; 1:282 J 2 C C

G(kj)

Loss function of the standard normal distribution.

k2
j

e 2  kj 1  Ukj 
Gkj p1
2p
(continued on next page)

16

B. Soylu, B. Akyol / Computers & Industrial Engineering 69 (2014) 1220

Table 1
Error rate results of methods.
Method

Set 1

Set 2

Mean

Std. dev.

p-Value

Mean

Std. dev.

p-Value

R-model

0.36263

0.02744

0.000

0.27683

0.03208

0.000

ZF-model
k = 0.0
k = 0.25
k = 0.50
k = 0.75
Case-based
Linear U.
UTADIS

0.34290
0.33460
0.34087
0.34647
0.39020
0.32890
0.25267a

0.03463
0.03273
0.03342
0.02867
0.04069
0.03890
0.03508

0.000
0.000
0.000
0.000
0.000
0.000

0.25107
0.24330
0.24737
0.25847
0.2779
0.30390
0.23197a

0.02927
0.03395
0.03800
0.03308
0.06270
0.04808
0.05134

0.000
0.022
0.103
0.039
0.001
0.000

FRj
FRT

Fill rate of SKUj. FRj 1 

rj

p
Lj

Qj

4.2. An illustrative example

Gkj

PN
FRj Dj
j1
Overall ll rate of the inventory system. FRT P
N
j1

hj
Wj
CT

Dj

Inventory holding cost of SKUj. (20% of the unit cost)

Unit ordering cost of SKUj. (1)
p
P
Total safety stock inventory cost. C T N
Lj
j1 hj kj rj

We only multiply the annual usage criterion of each SKU by

1000 to obtain more meaningful inventory cost values instead
of
p
normalized values. Here, the inventory cost of an SKUj hj kj rj Lj
would be different depending on the class assigned. This difference
is actually due to the safety factor (kj) which decreases from classes
A through C. Therefore, we expect higher cost if the SKUj is located
in the class A rather than in other classes. On the other hand, we
also analyze the ll rate, which is the fraction of demand directly
satised from stock on hand. The ll rate is an important indicator
to measure the customer satisfaction. Since the loss function of the
standard normal distribution gives lower G(kj) values for the higher values of kj, the ll rate of an SKUj in class A would be higher
than in other classes.
The simulation results are presented in Table 2. In terms of
inventory cost, case-based sorting gives the lowest value in both
sets. Inventory costs in Set 2 are much lower than in Set 1. This
is an expected result since the safety factor (kj) decreases from
class A through C. A disadvantage of this inventory cost calculation
is that it does not take shortage cost (i.e. criticality) into account.
The ZF-model k 0:50 and UTADIS give the best ll rate in Sets
1 and 2, respectively. Furthermore, rather than analyzing cost-service objectives separately Teunter et al. (2010) suggest an alternative cost approach where the ll rate objective is translated into a
penalty cost per backorder. However, we did not consider this issue in this study.

In this study, we apply the utility function based methods to the

ABC inventory classication problem given in the literature and
compare results with models by Ramanathan (2006), Zhou and
Fan (2007) and Chen et al. (2008). Flores, Olson, and Dorai (1992)
present the data including 47 inventory items and corresponding
criteria values given in Table 3 on a hospital inventory management problem. We consider the annual usage value, average unit
cost, lead time and critical factor as criteria in this case to be able
to make the results compatible with the results of other algorithms. Higher values of these criteria indicate to be located in
higher classes. In some previous studies, the critical factor criterion
was omitted by some authors due to its discontinuity and nonlinearity. To alleviate these disadvantages Torabi, Hate, and Pay
(2012) proposed a modied version of existing DEA based models.
Table 3 presents the criteria values, reference SKUs and nal classications. Reference SKUs for the case-based, linear and UTADIS
models are dened by assuming that the DM has the linear utility
function given in Eq. (18) which is actually evolved from AHP analysis (Flores and Whybark, 1987). Based upon the utility values, rst
10 items have been classied as A, next 14 items have been classied as B and the last 23 items have been classied as C. The SKUs
marked with a, b and c in Table 3 are randomly selected as reference SKUs based on the AHP classication.

zj1  5:12
Uz 0:07872
210  5:12

!
!
zj2  25:38
zj  1
0:40999 3
5840:64  25:38
71
!
j
z  0:01
0:41969 4
1:0  0:01

0:09161

18

Table 2
Inventory cost and ll rate results of methods.
Method

Inventory cost

Fill rate

Set 1

Set 2

Set1

Set 2

Mean

Std. dev.

Mean

Std. dev.

Mean

Std. dev.

Mean

Std. dev.

R-model

12046

1505

4703

600

0.98627

0.00091

0.98157

0.00097

ZF-model
k = 0.00
k = 0.25
k = 0.50
k = 0.75

12019
12072
12085
12069

1552
1536
1522
1511

4682
4728
4730
4719

646
638
614
612

0.98637
0.98657
0.98663a
0.98637

0.00113
0.00097
0.00085
0.00093

0.98163
0.98190
0.98170
0.98163

0.00103
0.00076
0.00095
0.00096

Case-based
Linear U.
UTADIS

11548a
12098
12000

1595
1538
1520

4662a
5069
4920

649
679
656

0.98390
0.98363
0.98613

0.00271
0.00299
0.00148

0.98120
0.98187
0.98417a

0.00262
0.00232
0.00109

indicates the best of all algorithms

17

B. Soylu, B. Akyol / Computers & Industrial Engineering 69 (2014) 1220

Table 3
Classications of 47 SKUs by different methods.
Item no
a

1
2a
3
4c
5
6
7
8
9
10b
11
12
13a
14
15
16
17
18
19
20
21
22
23
24
25c
26
27c
28
29b
30
31
32
33
34c
35
36b
37
38
39
40
41
42
43
44
45b
46
47

Average unit cost

Annual dollar usage

Lead time

Critical factor

AHP

R-model

ZF (0.5)

Case based

Linear U.

UTADIS

49.92
210
23.76
27.73
57.98
31.24
28.2
55
73.44
160.5
5.12
20.87
86.5
110.4
71.2
45
14.66
49.5
47.5
58.45
24.4
65
86.5
33.2
37.05
33.84
84.03
78.4
134.34
56
72
53.02
49.48
7.07
60.6
40.82
30
67.4
59.6
51.68
19.8
37.7
29.89
48.3
34.4
28.8
8.46

5840.64
5670
5037.12
4769.56
3478.8
2936.67
2820
2640
2423.52
2407.5
1075.2
1043.5
1038
883.2
854.4
810
703.68
594
570
467.6
463.6
455
432.5
398.4
370.5
338.4
336.12
313.6
268.68
224
216
212.08
197.92
190.89
181.8
163.28
150
134.8
119.2
103.36
79.2
75.4
59.78
48.3
34.4
28.8
25.38

2
5
4
1
3
3
3
4
6
4
2
5
7
5
3
3
4
6
5
4
4
4
4
3
1
3
1
6
7
1
5
2
5
7
3
3
5
3
5
6
2
2
5
3
7
3
5

1
1
1
0.01
0.5
0.5
0.5
0.01
1
0.5
1
0.5
1
0.5
1
0.5
0.5
0.5
0.5
0.5
1
0.5
1
1
0.01
0.01
0.01
0.01
0.01
0.01
0.5
1
0.01
0.01
0.01
1
0.01
0.5
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01

A
A
A
C
B
C
C
C
A
B
B
B
A
B
A
C
B
A
B
B
A
B
A
A
C
C
C
C
B
C
B
B
C
C
C
B
C
C
C
C
C
C
C
C
B
C
C

B
C
C
A
C
C
C
C
C
C
C
C
C
A
A
A
B
A
A
B
B
B
B
C
C
C
B
A
A
C
B
C
B
A
C
C
B
C
B
B
C
C
B
C
A
C
B

A
A
A
C
B
C
C
B
A
A
C
B
A
A
C
C
C
A
B
B
C
B
B
C
C
C
C
A
A
C
B
C
B
B
C
C
B
C
B
B
C
C
C
C
B
C
C

A
A
A
C
B
C
C
C
A
B
B
C
A
C
B
C
C
B
C
C
B
C
B
B
C
C
C
C
B
C
C
B
C
C
C
B
C
C
C
C
C
C
C
C
B
C
C

A
A
A
C
B
B
B
C
A
B
B
B
A
B
B
B
B
B
B
B
B
B
B
B
C
C
C
C
B
C
B
B
C
C
C
B
C
B
C
C
C
C
C
C
B
C
C

A
A
B
C
C
C
C
C
B
B
B
C
A
B
B
C
C
B
C
C
B
B
B
B
C
C
C
B
B
C
B
B
C
C
C
B
C
C
C
C
C
C
C
C
B
C
C

C Ref
for case based model, linear utility model and UTADIS.
A

C Ref
for case based model, linear utility model and UTADIS.
B

C Ref
for case based model, linear utility model and UTADIS.
C

First, we should note that we cannot determine a winner by looking

at classications. We can only infer some results by analyzing the
differences in each algorithms classications. Accordingly, with
the assumption of the DM has the utility function given in Eq.
(18) the linear utility model gives the closest classication to
AHP. Number of matches with respect to AHP are 18, 29, 35, 38
and 34 over 47 items for the other algorithms, respectively. We observe that an SKU is placed into adjacent classes, i.e. either A or B or
B or C, by all models except the R-model. If a SKU is placed into far
classes, like A and C, by the two methods, this implies that these
two algorithms are very inconsistent with each other. Fig. 4 illustrates the distribution of items into classes. Since there is no rule
about class sizes in the case-based model, linear utility model and
UTADIS, their class sizes are changeable.
Although the utility based classication procedures depend on
the chosen reference set, they are not so sensitive to the small
changes in the reference set. Assume that we have an optimal

Fig. 4. Number of SKUs placed into each class by each algorithm.

18

B. Soylu, B. Akyol / Computers & Industrial Engineering 69 (2014) 1220

Table 4
Classication results (average of 5-fold cross-validation, %).
AHP

LP1
A

Avg.

Min.

Max.

Avg.

Min.

Max.

Avg.

Min.

Max.

A
B
C

100
5
0

100
0
0

100
25
0

0
95
4

0
80
0

0
100
2

0
0
96

0
0
80

0
0
100

Avg. accuracy

97.0

50
0
0

100
30
0

0
70
13

0
50
0

0
100
50

10
17
87

0
0
50

50
50
100

UTADIS
A
B
C

90
13
0

Avg. accuracy

82.3

Bold indicates correct classication percentages

solution to the LP1 or UTADIS model and all SKUs are classied
according to the optimal solution. If the following modications
are done in the reference set, the nal classication will not change.
 If a new reference SKUj e Cn is added to the C ref
(for instance,
n
SKU11 e CB is added to the C ref
),
then
the
nal
classication
will
B
not change.

Remark 1. The optimal solution will always satisfy the new constraint added to the LP1 or UTADIS.

 If we remove a reference item SKU j 2 C ref

n of which constraint in
LP1 or UTADIS model is nonbinding (inactive) at the optimal
solution (for instance, the constraint regarding SKU 2 2 C ref
is
A
nonbinding at the optimal solution), then the nal classication
will not change.

Remark 2. The optimal solution will not change if nonbinding

constraints (among the constraint sets (3)(6)in LP1 or the constraint sets (12)(15) in UTADIS) are removed from the LP1 or
UTADIS.
An ideal methodology should not also be so sensitive to the considerable changes in the reference set. We applied the cross-validation approach (Stone, 1974) to analyze the sensitivity of methods
to the different reference selections. For this purpose, 47 SKUs
are randomly split into 5 mutually exclusive folds where there
are 10 SKUs in two folds and 9 SKUs in three folds. At each replication k, the fold k is used as the validating set and the remaining sets
are used as the training sets. The LP1 and the UTADIS models are
generated over the training set and validated over the validating
set. Assuming that the DM has the utility function given in Eq.
(18), we computed the classication accuracies of both methods
given in Table 4. According to these results, the LP1 model classies the 100% of SKUs in class A of validating set correctly while
the UTADIS model classies the 90% of them correctly. Although
both methods have over 80% average accuracy rate, the LP1 gives
the highest accuracy (97%). Concerning the accuracy rates, the
LP1 is less sensitive to the changes in the reference sets for this
example.
We also perform the service-cost performance analysis as stated in Lajili et al. (2012). Our parameter settings are given in Section 4.1. Fig. 5 illustrates the cost versus FRT (overall ll rate)
graph for each method. According to this gure, except for the ZF
(0.5) model, the other algorithms are nondominated with each

Fig. 5. Cost versus FRT values of each algorithm.

other. For instance, in terms of cost criterion the R-model is the

best one. On the other hand, in terms of the FRT criterion the linear
utility model is the best one. UTADIS and case-based model are
better than others in one of the criteria. For justication, we also
include results for two extreme cases: All SKUs are assigned in
class A and all SKUs are assigned in class C. When all SKUs are
assigned in class A, the ll rate is highest but the inventory cost
is also highest. Opposite of this case is observed when all SKUs
are assigned in class C.
5. Discussion
The multi-criteria ABC classication problem has been receiving
more attention from experts both in theory and in practice. Most
studies so far have evaluated the problem in a very general framework. However, specic industry characteristics may inuence the
resulting classications. Although the evaluation criteria set is
common in general, criteria priorities and class discrimination
rules are industry or company specic. To take these specications
into consideration, we need the DMs judgment. In this study, this
information is inferred from reference items provided by the DM,
and two types of utility functions (linear and piece-wise linear)
are generated. Classication methods are also evaluated in terms
of inventory cost and ll rate. There are different inventory policies
in the literature. As a further research study, the effect of the
multi-criteria ABC analysis on these policies should also be
analyzed together. Additionally, the DMs may have other types of

B. Soylu, B. Akyol / Computers & Industrial Engineering 69 (2014) 1220

utility functions depending on the industry. Different utility functions should be considered and evaluated.
Acknowledgement
This study was supported by the Erciyes University, Scientic
Research Unit under Grant No. BAP-FBY-11-3457.
Appendix A
R-model (Ramanathan, 2006): assigns favorable weights for
each SKU by using the following model. The objective function of
the model is the inventory score (IS(j)) of the corresponding SKUj
(IS(j)). SKUs are classied by selecting user-dened thresholds.

Max ISj

m
X
wi zji

19

8j 2 S

i1

s:t:
m
X
wi zji 6 1 8j 2 S
i1

wi P 0 8i 1; 2; . . . ; m
ZF model (Zhou & Fan, 2007): Least favorable weights are also computed by another model in addition to R-model. Finally, a combined
index CIk j is computed.

Min BIj

m
X
wbi zji

8j 2 S

i1

s:t:
m
X
wbi zji P 1 8j 2 S
i1

wbi P 0 8i 1; 2; . . . ; m
CIk j k

Fig. A1. Distances to the ideal SKU.

ISj  IS
BIj  BI

 1  k
IS  IS
BI  BI

where IS* = maxj2S {IS(j)} and IS = minj2S{IS(j)} and

BI* = maxj2S {BI(j)} and BI = minj2S{BI(j)}.
0 6 k 6 1 is a control parameter which may reect the preference of
the DM on the favorable index and the least favorable index.
Case-based model (Chen et al., 2008): In this model, the preferences of the DM are synthesized from the cases selected by the DM
as in UTADIS. Two articial SKUs, A+ ideal SKU and A anti-ideal
SKU, are set as given in Figs. A1 and A2. The distance of an SKU
from A+ and A determines its class. A weighted distance function
is applied. If an SKU is close to A+ and faraway from A-, then it deserves to be located in higher classes. The weights and thresholds
are determined by using reference SKUs according to ideal and
anti-ideal SKUs separately. So, two classications are available
for each SKU. By using joint matrix in Fig. A3 the nal class of an
SKU is determined.
References

Fig. A2. Distances to the anti-ideal SKU.

Fig. A3. The joint matrix for classications according to A+ and A.

Bhattacharya, A., Sarkar, B., & Mukherjee, S. K. (2007). Distance-based consensus

method for ABC analysis. International Journal of Production Research, 45(15),
34053420.
akr, O., & Canbolat, M. S. (2008). A web-based decision support system for multicriteria inventory classication using fuzzy AHP methodology. Expert Systems
with Applications, 35, 13671378.
Chen, Y., Kevi, W. L., Marc Kilgour, D., & Hipel, K. W. (2008). A case-based distance
model for multiple criteria ABC analysis. Computers & Operations Research, 35,
776796.
Chu, C. W., Liang, G. S., & Liao, C. T. (2008). Controlling inventory by combining ABC
analysis and fuzzy classication. Computers & Industrial Engineering, 55,
841851.
Cplex (2010). IBM ILOG Cplex 12.1 optimizer users manual.
Doumpos, M., & Zopounidis, C. (2004). A multicriteria classication approach based
on pairwise comparisons. European Journal of Operational Research, 158,
378389.
Flores, B. E., Olson, D. L., & Dorai, V. K. (1992). Management of multicriteria
inventory classication. Mathematical and Computer Modelling, 16(12), 7182.
Flores, B. E., & Whybark, D. C. (1987). Implementing multiple criteria ABC analysis.
Journal of Operations Management, 7(1), 7984.
Gvenir, H. A., & Erel, E. (1998). Multicriteria inventory classication using a genetic
algorithm. European Journal of Operational Research, 105, 2937.
Hadi-Vencheh, A. (2010). An improvement to multiple criteria ABC inventory
classication. European Journal of Operational Research, 201, 962965.
Kampen, T. J. V., Akkerman, R., & Donk, D. P. V. (2012). SKU classication: A
literature review and conceptual framework. International Journal of Operations
& Production Management, 32(7), 850876.
Keskin, G. A., & zkan, C. (2013). Multiple criteria ABC analysis with FCM clustering.
Journal of Indusrtrial Engineering, 2012, 17.
Lajili, I., Babai, M. Z., & Ladhari, T. (2012). Inventory performance of multi-criteria
classication methods: An empirical investigation. In Proceedings of 9th
international conference of modeling, optimization and simulation MOSIM12.

20

B. Soylu, B. Akyol / Computers & Industrial Engineering 69 (2014) 1220

Mousseau, V., & Slowinski, R. (1998). Inferring an ELECTRE TRI model from
assignment examples. Journal of Global Optimization, 12, 157174.
Ng, W. L. (2007). A simple classier for multiple criteria ABC analysis. European
Journal of Operational Research, 177, 344353.
Partovi, F. Y., & Anandarajan, M. (2002). Classifying inventory using an articial
neural network approach. Computers & Industrial Engineering, 41, 389404.
Partovi, F. Y., & Hopton, W. E. (1993). The analytic hierarchy process as applied to
two types of inventory problems. Production and Inventory Management Journal,
35(1), 1319.
Ramanathan, R. (2006). ABC inventory classication with multiple-criteria using
weighted linear optimization. Computers & Operations Research, 33, 695700.
Silver, E. A, Pyke, D. F, & Preterson, R. (1998). Inventory management and production
planning and scheduling. John Wiley & Sons.
Soylu, B. (2011). A multi-criteria sorting procedure with Tchebycheff utility
function. Computers & Operations Research, 38(8), 10911102.

Stone, M. (1974). Cross-validation choice and assessment of statistical predictions.

Journal of the Royal Statistical Society B, 36, 111147.
Taylor, S. G., Sewart, S. M., & Bolander, S. F. (1981). Why the process industries are
different. Production & Inventory Management Journal, 22(4), 924.
Teunter, R. U., Babai, M. Z., & Syntetos, A. A. (2010). ABC classication: Service levels
and Inventory costs. Production and Operations Management, 19(3), 343352.
Torabi, S. A., Hate, S. M., & Pay, B. S. (2012). ABC inventory classication in the
presence of both quantitative and qualitative criteria. Computers & Industrial
Engineering, 63, 530537.
Tsai, C. Y., & Yeh, S. W. (2008). A multiple objective particle swarm optimization
approach for inventory classication. International Journal of Production
Economics, 114, 656666.
Zhou, P., & Fan, L. (2007). A note on multi-criteria ABC inventory classication using
weighted linear optimization. European Journal of Operational Research, 182,
14881491.