MANAGEMENT SCIENCE
informs
Vol. 50, No. 1, January 2004, pp. 117–131
issn 0025-1909 eissn 1526-5501 04 5001 0117
®
doi 10.1287/mnsc.1030.0165
© 2004 INFORMS
Metaheuristics with Local Search Techniques for Retail
Shelf-Space Optimization
Andrew Lim
Department of Industrial Engineering and Engineering Management, Hong Kong University of Science and Technology,
Clear Water Bay, Kowloon, Hong Kong, iealim@ust.hk
Brian Rodrigues
School of Business, Singapore Management University, 469 Bukit Timah Road, Singapore 259756, br@smu.edu.sg
Xingwen Zhang
Department of Industrial Engineering and Engineering Management, Hong Kong University of Science and Technology,
Clear Water Bay, Kowloon, Hong Kong, xwzhang@ust.hk
E
fficient shelf-space allocation can provide retailers with a competitive edge. While there has been little study
on this subject, there is great interest in improving product allocation in the retail industry. This paper
examines a practicable linear allocation model for optimizing shelf-space allocation. It extends the model to
address other requirements such as product groupings and nonlinear profit functions. Besides providing a
network flow solution, we put forward a strategy that combines a strong local search with a metaheuristic
approach to space allocation. This strategy is flexible and efficient, as it can address both linear and nonlinear
problems of realistic size while achieving near-optimal solutions through easily implemented algorithms in
reasonable timescales. It offers retailers opportunities for more efficient and profitable shelf management, as
well as higher-quality planograms.
Key words: retail; shelf allocation; metaheuristics
History: Accepted by Thomas Liebling, former department editor; received October 31, 2002. This paper was
with the authors 5 months for 2 revisions.
1.
Introduction
Naert (1988), Borin et al. (1994), Urban (1998), and
Desmet and Renaudin (1998). With a well-designed
shelf allocation system, retailers can improve inventory return on investment as well as raise consumer
satisfaction by reducing the likelihood of products
being out of stock. More significantly, the retailer can
improve the financial performance of the store and
increase profit margins while reducing manpower
costs (Buttle 1984, Fanscher 1991, Yang and Chen
1999, Yang 2001). While retailers are continually challenged to display new products for the best returns,
how space is used impacts on store operating costs
due to procurement, carrying, reshelving, and out-ofstock costs (Zufryden 1986, Drèze et al. 1994). Efficient
space management can also allow for better brand
exposure, which can encourage impulse buying and
boost incremental sales (Levy and Weitz 1995, Walters
and Bommer 1996) while providing an effective tool
for retailers to implement mixed strategies that combine low cost with differentiation (Helms et al. 1992,
Yang and Chen 1999).
Well-managed shelf-space planning provides the
basis for making category-specific merchandising
decisions where the traditional space management
tool employed is a planogram, which provides a
In the highly competitive retail industry, one of the
keys to gaining an edge is an efficient shelf allocation system where shelf space is often the retailer’s
scarcest resource. As the number of brand lines continually increases, allocating products to the supermarket shelf in the best possible arrangement poses
challenges to the retailer. Within the retail industry, user interest in shelf-space allocation is found to
be very high. A recent search of the ABI/INFORM
database resulted in over 500 references and numerous recent articles in practitioner journals such as
Advertising Age, Supermarket Business, Beverage World,
and Electronic Business.
Chen et al. (1999) emphasize the importance of
store space allocation decisions: “In the face of a
deluge of new products and the substantial profit
opportunities available through slotting allowances,
researchers have powerful incentives to make this
decision correctly” (p. 216). Indeed, there have been
a number of studies on the effects of space on sales,
and others that document the positive impact of space
allocation on product performance. These include the
works of Cairns (1963), Frank and Massy (1970),
Curhan (1973), Corstjens and Doyle (1981), Bultez and
117
118
Lim, Rodrigues, and Zhang: Metaheuristics with Local Search Techniques for Shelf-Space Optimization
shelf layout of products and a workable method by
which merchandising plans can be communicated
efficiently. Retailers, together with merchandisers,
focus on developing effective visual merchandising
plans to maximize profits on a store-by-store basis.
With effective planogramming, derived from efficient
shelf allocation algorithms, it is possible to leverage
every inch of selling space available and to capitalize on available data to meet financial targets (Zufryden 1986, Drèze et al. 1994, Borin et al. 1994, Yang
2001). Due to the problem’s complexity, only relatively simple heuristic rules have been developed and
are available for retailers to plan product-to-shelf allocation (Zufryden 1986, Yang 2001). PC-based systems
now available, such as Appollo (IRI) and Spaceman
(Nielsen), allocate space using turnover and gross
profit margins as their criteria and handling and
inventory cost as constraints. These are not effective
as global optimization tools (Desmet and Renaudin
1998) and are largely used for planogram accounting to reduce time spent on manual manipulation of
shelves (Drèze et al. 1994, Yang 2001). In the U.K.,
Galaxi, developed by Space Solutions and used in one
of its largest supermarket chains, Tesco, does not provide optimization of shelf space and operates mainly
by a manual drag-and-drop procedure, i.e., drag a
product line and drop it onto a shelf (Bai 2003).
The objectives of this paper are to present realistic
retail shelf allocation optimization models and to offer
efficient algorithms that are operationally viable. To
achieve these goals, we develop models that are practicable and easily managed, and solution techniques
that can deal with industry-size problems within reasonable timescales. Our starting point is with recent
research on this problem by Yang and Chen (1999)
and Yang (2001), who used a simplified alternative
of the well-known Corstjens and Doyle (1981) model.
We first develop a network flow solution approach to
the problem and then, using many-to-many neighborhood moves, employ a strategy of combining a strong
local search with metaheuristics. We then apply these
techniques to more complex models that address
product groupings and nonlinear profit functions.
We summarize related and previous work in §2 and
provide formulations of the linear shelf-space allocation, product grouping, and nonlinear profit function models in §3. A discussion of the most recent
solution approach and its limitations is given in §4.
The remainder of the paper begins with a description of an improved and generally applicable manyto-many neighborhood search technique that is given
in §5. We then develop a network model to obtain
upper bounds in §6 and a network flow solution in §7.
Metaheuristics, including Tabu Search and a hybrid
of Squeaky-Wheel Optimization, are developed in §8.
Experimental results and comparisons are given in §9,
and §10 provides conclusions to this work.
Management Science 50(1), pp. 117–131, © 2004 INFORMS
2.
Related and Previous Work
A recent survey by Yang and Chen (1999) highlights the lack of academic work on this subject.
Only 12 references were cited, of which 5 date to
the seventies. Further, we find no definitive shelf
allocation model(s) and, consequently, no benchmarks
are available.
Models have been developed in the past 40 years
to address the various objectives associated with
product-to-shelf allocation. In a very early work,
Cairns (1963) proposed a graphical solution to the
problem of allocating shelf space to two products to
maximize profits while taking space elasticity into
account. More recently, space optimization models
were developed by Anderson and Amota (1974),
Anderson (1979), and Hansen and Heinsbroek (1979).
Anderson and Amota (1974) proposed a model that
considered optimal brand selection and display area
allocation, given a profile of consumer brand preferences. In his work, Anderson (1979) used the relationship between a product’s share of space and its
market share to decide space allocation, and provided
an integrated theory of profit maximization of display
space. The model by Hansen and Heinsbroek (1979)
took into account space elasticity and requires a minimum shelf allocation for each product. Cross-space
influences were factored in, as were out-of-stock and
replenishment costs, and “near-optimal” solutions are
obtained using a Lagrange multiplier approach.
Corstjens and Doyle (1981) developed a geometric
programming model where both product and crossspace elasticities were considered and where profit
maximization was the objective. The model highlights
sales-space elasticities that gauge the sales response
of a given product to the space allocated to another,
and was the first space allocation model to incorporate such interdependencies. Costs were modeled as
functions of inventory investment and handling, and
constraints included store size and upper and lower
bounds on space of each category, as well as availability constraints. Their model was solved using a
branch-and-bound technique developed by Grochet
and Smeers (1979) and was an important landmark
in the development of useful space allocation models.
Space allocation was found to be dependent on product profitability, demand space elasticity, and product
cross elasticity. Corstjens and Dolye (1983) extended
the static model to a dynamic one that allows for the
anticipation of changing customer tastes and changing product growth and life cycles that could motivate retailers to allocate more space to new products
and to divest from declining ones. Zufryden (1986)
proposed a model that takes into account space elasticity, cost of sales, and demand-related marketing
variables, but neglected cross elasticity between products while fixing nonspace marketing variables. A
Lim, Rodrigues, and Zhang: Metaheuristics with Local Search Techniques for Shelf-Space Optimization
Management Science 50(1), pp. 117–131, © 2004 INFORMS
dynamic programming approach to the model was
presented. The work of Bultez and Naert (1988), Borin
et al. (1994), and Borin and Farris (1995) builds on
the work of Corstjens and Dolye and extends modeling cross elasticities in product category management
models. Bultez and Naert’s Shelf Allocation for Retailers’ Profit (SHARP) model is similar to the Corstjens
and Dolye (1981) model and optimized space allocation within a product category, taking into account
interdependencies within product groups and across
groups. They used marginal analysis and a search
heuristic to derive solutions. In Bultez et al. (1989),
space elasticities were estimated using a symmetric attraction model and later an asymmetric attraction model. The model is restricted to linear cases
because the marginal analysis used is not practical for
nonlinear models. Borin et al. (1994) addressed product assortment and space allocation in a constrained
optimization problem for which a heuristic solution
approach was employed, with the study concluding
that ignoring out of stock and product assortment
effects leads to suboptimality. More recently, Desmet
and Renaudin (1998) used the Corstjens and Dolye
model framework in an empirical study of product
category sales reponsiveness to allocated shelf space
where the model is based on a demand function linking the share of sales to the share of space allocated
to the product category. The results obtained suggest that space elasticities increase with the impulsebuying rate of the product category.
In a large-scale experimental study on shelf management, Drèze et al. (1994) concluded that the position of a product on the shelf is far more important
to determining sales than the number of facings allocated to the product, as long as a minimum threshold to avoid out of stocks was maintained. Other
recent studies involving space allocation include the
works of Urban (1998), who studied the integration of
inventory control models, product assortment models, and shelf-space allocation models; and Campo
et al. (2000), who investigated the impact of location
factors and derived optimal space allocation rules for
different location profiles.
Yang and Chen (1999) and Yang (2001) developed
a model based on the nonlinear model of Corstjens
and Dolye (1981). However, because the latter is
difficult to apply in realistic situations, Yang and
Chen proposed a simplified yet practicable alternative
model in the form of a linear multiknapsack integer
program.
In shelf-space allocation studies, we have found
that a number of researchers used mathematical programming techniques for the models they studied
(Hansen and Heinsbroek 1979, Corstjens and Dolye
1981, Zufryden 1986, Drèze et al. 1994, Yang and
Chen 1999, Campo et al. 2000, Yang 2001), while
119
others employed statistical analysis (Anderson 1979,
Walters and Bommer 1996). In Zufryden (1986), simulation techniques were used to generate required
parameters. Less often, heuristics have been used.
Borin et al. (1994), Borin and Farris (1995), Urban
(1998), and Yang (2001) developed heuristic solutions
to their models. Borin et al. used a heuristic procedure based on simulated annealing; Urban, in an
inventory-theoretic approach to shelf-space allocation,
used a greedy heuristic and a genetic algorithm for
the solution of an integrated problem, whereas Yang
applied adjustment heuristics, which we discuss in §4.
3.
Model Formulations
3.1. Background
Corstjens and Dolye (1981, 1983) and Zufryden (1986)
gave comprehensive models for shelf allocation that
were managerially useful and are well known to
researchers in the field. In the Corstjens and Dolye
models, product space elasticities and cross elasticities are considered within a total profit objective function estimated by product demands and costs in the
form of polynomials. The basic constraints were: store
shelf capacities, product availability, lower and upperbound constraints on products, and nonnegativity
constraints. These models are well integrated and
have been used and referenced by many researchers
(Zufryden 1986, Bultez and Naert 1988, Borin et al.
1994, Borin and Faris 1995, Drèze et al. 1994, Desmert
and Renaudin 1998, Urban 1998, Campo et al. 2000,
Yang and Chen 1999, Yang 2001).
The models of Corstjens and Dolye and of Zufryden
are complex and have practical limitations (Yang
2001). Both present nonlinear or multiplicative profit
functions. On the other hand, we find technical simplifications that make retail application unrealistic.
For example, in Corstjens and Dolye (1981) there is no
provision for the requirement for integer-valued number of displayed products, and in Zufryden (1986)
product areas are taken to be multiples of shelf area
grid “slot” values. Further, these models do not provide for the location factors emphasized by Drèze
et al. (1994).
In view of these limitations, Yang and Chen (1999)
proposed a simplified integer programming model
based on the model developed by Corstjens and
Dolye. For ease of reference and continuity, we use
the notation provided by Yang and Chen (1999) in
describing the model.
3.2. The Corstjens and Dolye Model
We assume there are m shelves with length Tk for each
shelf k and that the length of the facing of a given
product i displayed on any of these shelves is ai for
i = 1 n. Let Li and Ui be the lower and upper
Lim, Rodrigues, and Zhang: Metaheuristics with Local Search Techniques for Shelf-Space Optimization
120
Management Science 50(1), pp. 117–131, © 2004 INFORMS
facings bounds for product i, respectively, and Ai be
its total availability. Let xik be the number of facings
of product i on shelf k. The demand function can then
be given by
Qik xik = i xikik
n
j=1 j=i
ij
xj
L
t=1
ytiti
where i , ik , ij , ti are parameters, where i is a
scaling constant, ik is the space elasticity of product i
on shelf k, ij is the cross elasticity between product i
and j, and ti is the elasticity of product i relative
to a
market variable, t for t = 1 L. Here, xj = m
k=1 xjk
is the total facings of product j and yti is the value
of the tth marketing variable w.r.t. to i. In Corstjens
and Dolye (1981), using survey data on five product
groups and the demand equation fitted by regression
analysis, the authors found ik to be predominantly
positive and to range between −001 and 019.
We can
write the gross cost, ci xi of product i as
i
ci xi = i m
k=1 Qik xik , where i is a sales volume
elasticity associated with the variable cost of product i and i is constant for product i If we assume
that the gross margin of product i is linear to its unit
margin, gi , we can then express a total profit function,
P , by
m
n
n
P = gi
Qik xik − ci xi
i=1
i=1
k=1
which we maximize subject to the constraints
n
i=1
ai xik ≤ Tk
Li ≤
m
k=1
m
for k = 1 m
(1)
xik ≤ Ui
for i = 1 n
(2)
Qik xik ≤ Ai
for i = 1 n
(3)
k=1
xik ∈ 0 1 2
for i = 1 n and k = 1 m
(4)
3.3. The Yang and Chen Model
Yang and Chen (1999) made simplifications to the
model of Corstjens and Dolye (1981). First, the need
for inequality (3) is removed by assuming that retailers
can prevent out-of-stock occurrences with good logistics. Second, in the absence of an estimation for i ,
the sales volume elasticity, they made the assumption
that the profit of any product is linear with respect to
a range of facings for which it is displayed by controlling the lower and upper bounds for that product.
Hence, by letting pik be the per-facing profit of product
i on shelf k, we can rewrite the profit objective to be
P=
m
n
i=1 k=1
pik xik
which we maximize subject to constraints (1), (2),
and (4).
Although researchers such as Bultez and Naert
(1988) and Drèze et al. (1994) have found that
marginal returns to space first increase and then
decrease in an S-shaped curve, we can justify the
linearity assumption for the profit function here by
the fact that retailers would want to operate on the
linear (or approximately linear) and more strongly
increasing part of the curve. This can be implemented
by using appropriate lower and upper bounds for
facings.
This integer programming multiknapsack problem
model is applicable because there are many integer programming packages available. Yang and Chen
(1999) describe a multistage approach using this
model, in view of the fact that products are commonly
Broken down into departments, categories, and items.
We shall refer to this model as the Shelf-Space
Allocation Problem (SSAP) in this work.
3.4. Product Groupings and Nonlinear Models
We study two useful extensions to the linear model.
The first deals with product associations, while the
second deals with nonlinear profit functions.
3.4.1. Product Groupings. This extension addresses the merchandising strategy of having products
or product categories placed together or apart as
arises, for example, from market basket analysis in
which tendencies of customers to purchase products together is investigated, and from which “crossselling” (cold medicines with Kleenex, beer, and diapers) and “affinity positioning” (coffee with coffee
makers) strategies can be better determined (Anonymous 1995, 1997, 1998). To model this, assuming n
products and m shelves, we represent cross-product
affinity by an n × n matrix, , where each value !ij ∈
is assigned by the manager and is predetermined.
In our experiments, we used this cross-product affinity matrix in such a way that if a facings of product i
and b facings of product j are put on the same shelf,
an additional profit defined to be !ij × mina b is
realized. The additional profit value can be defined
differently if required, and this expression was chosen
as an example of how such a profit can be implemented. From this, the extended objective is to maximize the sum of the original profit and the additional
profit (or penalty); i.e.,
P1 = P +
n−1
n
i=1 j=i+1
!ij yij
where P is the original profit function, !ij ∈ , and
yij = m
k=1 minxik xjk , subject to constraints (1), (2),
and (4).
Lim, Rodrigues, and Zhang: Metaheuristics with Local Search Techniques for Shelf-Space Optimization
Management Science 50(1), pp. 117–131, © 2004 INFORMS
Here, for a given product pair i j, the value yij is
the sum of minxik xjk for all shelves k. When there
are more facings of product i and j placed together
on the same shelf, the additional benefit achieved is
higher (or lower, in case !ij is negative). For example, given two products and three shelves, if we
place two facings of Product 1 and four facings of
Product 2 on Shelf 1, three facings of Product 1 and
one facing of Product 2 on Shelf 2, and zero facings of Product 1 and five facings of Product 2 on
Shelf 3, then y12 will equal min2 4 + min3 1 +
min0 5 = 2 + 1 + 0. This y12 value is multiplied
by the coefficient !12 to obtain the additional profit
gained in placing Products 1 and 2 together on the
same shelves. Because placing product i together with
product j is equivalent to placing product j with product i, only the values of yij when i < j are needed. In
our example, we need only compute y12 , not y21 .
3.4.2. Nonlinear Profit Functions. This extension
addresses the many situations where profit functions
that arise are nonlinear, as, for example, in the case
where the profit function follows an S-shaped profile
discussed in §3.3. Indeed, the model of Corstjens and
Dolye (1981) given here is nonlinear in its objective
with a nonlinear product resource constraint (constraint (3)). In Zufryden (1986), we find a similar
nonlinear product constraint and in Urban (1998) the
objective is to maximize a nonlinear demand function of the allocated space. In Bookbinder and Zarour
(2001), a direct product profitability is integrated into
shelf optimization where the objective is a sum of
nonlinear demand and cost functions. For nonlinear
profit functions that arise, many are multiplicative as,
for example, in the Corstjens and Dolye (1981) model.
As a simple example of a nonlinear profit function,
1/2
the function 5 · xik gives a profit of 5 units when one
facing of product i is placed on shelf k, while placing
two facings achieves 707 units of profit. We note that
this function is an example of the top of an S-shaped
curve where the derivative is decreasing. In general,
we seek to maximize profit functions of the form
P2 =
m
n
i=1 k=1
F i k xik
where F is a nonlinear function, subject to constraints
(1), (2), and (4).
4.
Yang’s Heuristics
Yang (2001) developed a heuristic algorithm commonly applied to knapsack problems to solve the
SSAP. The profit of each item per displayed length on
a particular shelf is treated as a weight, and the ranking order of weight is used as a priority index in the
process of space allocation. The algorithm consists of
121
three phases. First, a preparatory phase checks for the
feasibility of a particular problem and builds a set of
priority indexes. Second, an allocation phase allocates
available space to items one by one following a priority. This phase is further divided into two subphases,
which assure that the lower and upper-bound constraints for the number of facings of a product are not
violated. Third, in a termination phase, the objective
value of the final solution is calculated.
Further, an adjustment phase consisting of three
adjustment methods is adopted to improve solutions.
Adjustment 1 attempts to improve a solution by
swapping one facing for a pair of products allocated
on the same shelf. Adjustment 2 interchanges one facing for a pair of products allocated on two shelves.
Adjustment 3 is an extension of Adjustment 2, which
attempts to allocate shelf space still available after
interchanging facings between two products on two
shelves.
The algorithm is then improved by using an adjustment phase, with different combinations of adjustment methods, after the allocation phase.
4.1. Limitations of Yang’s Heuristics
Using a self-defined set of testing data, Yang (2001)
claimed the mean profit of solutions after the
allocation phase to be 98.2% of the optimal mean
profit. After using the adjustment phase, the average
profit ratio of the solutions obtained by his improved
heuristic to optimal solutions was 99.6%. The test
data given by Yang, however, was limited. There was
only one group of test data. The total available shelf
space was generally more than enough to allocate
all of the products in the upper bound. That is, the
lower and upper bound had no impact on the resulting feasible solution. Also, the test data scales were
limited, where, for example, a range of 0–3 units for
product-facings bounds was used, which had little
impact on optimal solutions obtained. In the data,
all shelves were assumed to have the same lengths
and all products had the same lower and upper
bounds. Further, it is not possible to determine how
the heuristics perform with the small problem sizes
used, and how CPU time is affected when the three
adjustment methods are added was not described.
There were limitations in the heuristic. For example,
in the allocation phase, to satisfy the lower bound
for all products, the algorithm can allocate a product
to a shelf for which there may be far better profit
if stocked with other products. Indeed, this is why
Yang’s (2001) algorithm requires an adjustment phase.
However, the adjustment phase itself may not be sufficient to allocate higher-priority products to their preferred shelves because if the length of the facings
of products varied widely, which is often the case,
then the three adjustments used could hardly make
Lim, Rodrigues, and Zhang: Metaheuristics with Local Search Techniques for Shelf-Space Optimization
122
Management Science 50(1), pp. 117–131, © 2004 INFORMS
improvements. For example, if product j occupies
shelf k, which is the shelf designated for product i,
and the facing of j is 20, while that for i is 400, then
all three adjustments fail to make any improvement.
However, in this case, it would certainly be possible
to interchange 20 units of product j with 1 unit of
product i for more profit.
5.
New Neighborhood Moves
To address some of their shortcomings, we extend the
three adjustment methods to multiple-facings adjustments of products. These many-to-many moves are a
natural extension of the adjustment methods and are
better suited to achieve product allocation to shelves.
We describe these new neighborhood moves:
Multishift Move. This move attempts to improve
the solution by swapping multiple facings for a pair
of products allocated on the same shelf. Figure 1
illustrates this move: Shift (remove) two out of three
pieces of Product B out, and shift five pieces of
Product E in if all constraints are fulfilled, for a profit
increase of $5.
Multiexchange Move. This move interchanges multiple facings for a pair of products allocated on two
shelves. Figure 2 shows two pieces of Product B on
Shelf 1 interchanged with five pieces of Product E
on Shelf 2 for an additional profit of $25 + $2 − $20−
$5 = $2, if all constraints are fulfilled.
Multiadd and Exchange Move. This move is an extension of the second move. The key point here is that
after interchanging a number of facings between two
products on two shelves, there may be enough shelf
space left over to be reallocated to other products. Figure 3 illustrates this when two pieces of Product B on
Shelf 1 are exchanged with five pieces of Product E
Figure 1
on Shelf 2, and there is still space left on Shelf 1 for
one unit of Product F. The total profit increases after
this operation.
The detailed description and verification of these
moves are straightforward, and hence omitted here.
All three moves ascertain whether the profit after
the move is greater than before. However, in using
these moves in our algorithms, we look for increases
in residual space after the moves are made even if
profit remains unchanged.
5.1.
Improved Yang’s Heuristics with New
Neighborhood Moves
For the purpose of evaluation, we replaced Yang’s
adjustment moves with the new neighborhood
moves and applied these to the algorithm given by
Yang. There are six combinations of orderings in
total, and performance testing of these showed
that the (Multiexchange)-(Multiadd and Exchange)(Multishift) combination option has the highest performance increase. Experimental performance using
these new moves is provided in §9, where they are
compared with the other algorithms.
6.
Upper Bounds for the SSAP
In this section, we provide a method to obtain upper
bounds for the problem. As the SSAP is NP-hard
(Yang 2001), we first make the simplifying assumption that all products are of unit length. We call this
form of the SSAP the Unit-Length Shelf-Space Allocation Problem (ULSSAP) and show that the ULSSAP
can be transformed into an equivalent Minimum Cost
Flow Problem (MCFP).
The MCFP is defined on a digraph with node and
directed arc sets. Each arc is associated with a cost
Multishift Move
$10
Shelf 1
Product
A
Product
A
Product
A
$10
Product B
Products not in
any shelves
Prod
E
$10
Product B
Prod
E
Prod
E
Product B
Prod
E
Product
C
Prod
E
Multi-Shift Operator
$10
Shelf 1
Product
A
Product
A
Product
A
Removed
products
Product B
$5
$5
$5
$5
$5
Prod
E
Prod
E
Prod
E
Prod
E
Prod
E
Product B
Product B
Product
C
Lim, Rodrigues, and Zhang: Metaheuristics with Local Search Techniques for Shelf-Space Optimization
123
Management Science 50(1), pp. 117–131, © 2004 INFORMS
Figure 2
Multiexchange Move
$10
Shelf 1
Product
A
Product
A
Product
A
Product B
$1
$1
Prod
E
Product D
Shelf 2
$10
Prod
E
Product B
$1
$1
Prod
E
Prod
E
Product
C
$1
Prod
E
Multi-Exchange Operator
Shelf 1
Product
A
Product
A
Product
A
$5
$5
$5
$5
$5
Prod
E
Prod
E
Prod
E
Prod
E
Prod
E
Product D
Shelf 2
and capacity. The flow through each edge is nonnegative and not more than its capacity. Each node, x,
is associated with a demand or supply value f x
that should be equal to the net flow at the node for
any valid flow. The MCFP is to find such a valid
flow on the network for which total cost is minimized (Ahuja et al. 1993) subject to flow bound constraints. Orlin (1988) has provided a fast polynomial
time algorithm to solve the MCFP with time complexity Om log nm + n log n, where n is the number
of nodes and m is the number of arcs in the digraph.
Here, we show that the ULSSAP can also be solved
Figure 3
$1
$1
Product B
Product B
Product
C
with polynomial time complexity via a transformation into a MCFP.
To achieve this, we implement three transformations. First, we transform the ULSSAP into a MCFP
with directed arcs associated with lower and upper
bounds, where the upper bound is taken as its capacity. Second, we transform the MCFP with lower
bound into a MCFP without lower bound. Third, we
transform the MCFP into a Minimum Cost Maximum
Flow Problem (MCMFP) to find a maximum flow, F ,
such that for any other maximum flow, F , the cost of
flow F is less than or equal to the cost of flow F .
Multiadd and Exchange Move
$10
Shelf 1
Product
A
Product
A
Product
A
Product B
$1
Prod
E
Product D
Shelf 2
$10
$1
Prod
E
Product B
$1
$1
Prod
E
Prod
E
Product
C
$1
F
Prod
E
Unallocated
Product
Multi- Add & Exchange Operator
$5
Shelf 1
Product
A
Product
A
Product
A
Prod
E
$5
$5
Prod
E
$1
Shelf 2
Product D
$5
Product B
$5
$5
Prod
E
$5
$5
$5
Prod
E
Prod
E
F
$1
$5 B $5
Product
Product
C
Lim, Rodrigues, and Zhang: Metaheuristics with Local Search Techniques for Shelf-Space Optimization
124
1
2
3
Figure 4
Length
Lower
bound
Upper
bound
Profit
Shelf 1
Profit
Shelf 2
1
1
1
3
2
1
5
7
5
3
4
5
4
2
2
Network After Transformation Phase I
-4
2
[1
,5
]
,6
]
6
-2
-5
5
]
,8
[0
3
-2
The ULSSAP is transformed into a MCMF from the
source node S to the destination node T . If the flow
amount of the maximum flow is less than the sum of
the lower bounds of all products, then the ULSSAP
has no solution; that is, there is insufficient shelf space
to satisfy all the lower-bound constraints. As long
as there is sufficient shelf space, the cost obtained
from the MCMFP is the negative of the optimal profit
value that is achievable from the ULSSAP.
For the example we have given here, using the
MCMF model, we find a maximum profit (upper
bound) of 55.
6.3. Proof of Correctness
For any instance of the ULSSAP, the transformed network model captures all the constraints. The capacity
constraint is given by the edge capacity of the arc
from a shelf node to the destination node. The total
amount of flow from all product nodes into a shelf
node is the same as that from the shelf node into the
destination node. Because the flow through an arc is
not allowed to exceed the capacity of the arc, the total
occupied length of a given shelf cannot exceed the
capacity of the shelf.
Figure 5
Final Network
S
[0,3
]
[0
,2
]
-3
]
1
,2
]
-4
4
[0
[0
-4
,6
[0,5]
0
2
[0
[0,6]
6.2. Transformation Phases II and III
After Transformation Phase I, to transform the MCFP
with lower bound into a MCFP without lower bound
and then into the MCMFP, we follow standard steps
as given by Ahuja et al. (1993), for example. Figure 5
illustrates the final network for the given example of
two products and three shelves.
[0
[2,7]
[0,1
6.1. Transformation Phase I
The purpose here is to transform the ULSSAP into
a MCFP with a lower bound. For a ULSSAP with
m shelves and n products, a graph of m + n + 2 nodes
is created such that Node 1 to node n represent Product 1 to n, respectively (product nodes), and such
that node n + 1 to node m + n represent Shelf 1 to
m, respectively (shelf nodes). For the remaining two
nodes, Node 0 is the source node and node m+n+1 is
the destination node of the network flow. The demand
or supply values f x of all nodes x are set to 0.
From the source node to each product node k, with
1 ≤ k ≤ n, an arc from 0 to k 0 k, is created with
lower bound Lower(k) and capacity Upper(k), and the
cost is set to 0.
From each product node to each shelf node, there
is an arc without lower and upper bounds. The cost
of the arc is the negative of the corresponding profit;
i.e., for an arc from a node representing product i to
a node representing shelf j, the cost of the arc is the
negative of the profit achieved by placing one item
of product i into shelf j, that is, −pij . Finally, from
each shelf node to the destination node, an arc with
cost 0 is established. The capacity of the arc is the
corresponding shelf capacity; i.e., for an arc from a
shelf node representing the shelf j to the destination
node, the capacity of the node is Capacity(j).
Figure 4 illustrates the network after Transformation Phase I for the given example of two products
and three shelves. In this figure, each (L U ) label adjacent to an arc indicates that L and U are the lower
and upper bound, respectively. An X next to an arc
indicates that the cost of the arc is X. With this labeling, the default value for L is 0, for U it is infinity,
and for X it is 0.
4
-4
0
We use a simple example to illustrate the transformation steps of a ULSSAP into an equivalent MCFP.
Assume we have two shelves and three products. The
capacities for Shelves 1 and 2 are 6 and 8, respectively.
Product properties and profits are given in Table 1.
-3
1
5]
Product
Product Properties and Profits
[3
,
Table 1
Management Science 50(1), pp. 117–131, © 2004 INFORMS
,4
5
3
-2
T
6
-2
-5
]
]
[0
,8
]
Lim, Rodrigues, and Zhang: Metaheuristics with Local Search Techniques for Shelf-Space Optimization
125
Management Science 50(1), pp. 117–131, © 2004 INFORMS
The lower and upper-bound constraints are captured by the corresponding lower and upper bounds
of the arc from the source node to the product node.
As the only incoming arc for a product node is from
the source node, and all outgoing arcs are from the
product node to all the shelf nodes, the amount of
flow through the incoming arc is the total length
occupied by that product. Further, as the amount of
flow through each incoming arc is not less than the
lower bound and not more than the upper bound, the
lower and upper-bound constraints for the product
are satisfied.
The maximum profit constraint is satisfied by the
fact that the MCFP generates a network flow that
produces maximum total cost and at the same time
satisfies the capacity constraint and the lower and
upper-bound constraints. As the cost for the MCFP is
minimal and negative, its negative is the maximum
profit for the ULSSAP.
Lastly, if no solution exists for the ULSSAP, the
MCMF model would give a maximum flow value less
than the sum of lower-bound values of all the products and infeasibility would be identified. This completes the proof of correctness of the representation of
the ULSSAP by the MCFP.
6.4. Catering for Additional Constraints
Besides the capacity and bound constraints, the network flow model given here is easily extended for the
ULSSAP to handle other types of constraints related
to lower or upper bounds. One possible constraint
type could limit the amount of a product allocated
onto a shelf. For example, we may require that at least
three and at most five items of Product 1 are placed on
Shelf 2. This constraint could be catered for by setting
correct lower and upper bounds for arcs from product to shelf nodes. The network model thus has the
advantage of being able to cater for this and similar
constraints arising in shelf allocation.
7.
Network Flow Model
We have shown that a network flow model can be
used to solve the ULSSAP to achieve maximum profit.
For the general SSAP, products are of variable lengths,
and because the problem is NP-hard, we do not
expect that the polynomial-time MCFP will be sufficient to solve it directly.
7.2. Integrality Satisfaction
Although integral, solutions for the ULSSAP may not
be integral solutions for the SSAP. For example, for
a product of length 8, the flow model may allocate
12 unit-length spaces of it to a shelf which comprises
of 15 product items. This forces fractions of product
items to be allocated onto shelves. For each allocation,
the fractional part is truncated. For example, if product i is allocated shelf space L on shelf j, after truncation, product i would have space L−remL Length(i),
where rema b is the remainder when a is divided
by b. The fractional space remL Length(i) is added
to the free space of shelf j. After truncation, the
capacity and upper-bound constraints will not be violated, although the lower-bound constraint can be.
For example, if product k has lower bound Lower(k) =
3 and length Length(k) = 4 and it is allocated to three
shelves with space allocations of 5, 6, and 2, then
although the total allocated space of 13 = 5 + 6 + 2
exceeds the lower-bound requirement of 12 = 3 × 4,
after truncation the remaining space is 4, 4, and 0, and
the lower-bound constraint for product k cannot be
satisfied.
To resolve the lower-bound violation for product k,
the following procedure is used. From the source to
the sink, we attempt to find a least-cost path to augment the flow for product k until its lower bound is
reached. If this is not possible, a maximum cost path
from sink to source is located via edges representing
products above their lower-bound requirements. Flow
Figure 6
Network After Transformation Phase I with Variable Product
Lengths
-3
1
5]
7.1.
-4
[3
,
Adaptation of the ULSSAP for
Nonunit Lengths
We transform the SSAP into a network flow model,
using its ULSSAP as a transition step. This transformation is similar to that employed to transform
the ULSSAP. The difference here is in the first transformation phase. For arcs from the source node to
product nodes, the lower and upper bounds are no
longer the corresponding lower and upper bounds
of the products. For the arc from the source node
to a product node representing product k, the lower
bound is (Length(k) × Lower(k)) and the upper bound
is (Length(k) × Upper(k)). For an arc from a product
node representing product i to a shelf node representing shelf j, the cost is −pij divided by Length(i).
We illustrate with the example given above, with
three products and two shelves. When the lengths are
changed to 1, 2, 1 for Products 1, 2, and 3, respectively,
the resulting network graph is shown in Figure 6.
4
[0
,6
]
-2
[4,14]
0
2
[1
,5
]
6
-1
-5
5
3
-2
,8
[0
]
Lim, Rodrigues, and Zhang: Metaheuristics with Local Search Techniques for Shelf-Space Optimization
126
Management Science 50(1), pp. 117–131, © 2004 INFORMS
is reduced along this maximum cost path without violating the lower-bound requirement of any product,
and the procedure is repeated to augment product k.
This is done until a feasible solution is found or no
backflow from sink to source can be found. In the
latter case, the algorithm will return a failure, as no
feasible solution can be found.
If there is a feasible solution, a similar procedure is
applied. Here, we choose the product and shelf combination that provides highest profit per unit length.
We will attempt to increase the placement of the product on that shelf by one. This is repeated until no
improvement is possible.
We provide experimental results of the network
flow method and comparisons with other algorithms
in §9.
While no one approach can address every problem
that arises, we seek to develop techniques that can
be applied to a range of problems. As we know, the
SSAP is linear and already NP-hard in complexity.
This suggests the development of heuristics for this
and other more intractable problems. Specifically, in
the next part of this study, we develop artificial intelligence techniques that have not been applied to shelf
allocation problems. These include Tabu Search (TS)
and Squeaky-Wheel Optimization (SWO). Although
other hybrids were developed and tested, their performance was no better than that of TS and SWO, to
which we limit our discussion here. We first apply
the heuristics developed to the (linear) SSAP and then
extend their application to the product grouping and
nonlinear models.
Elite Candidate List. Careful and intelligent organization in choosing a candidate list is the key when
faced with the SSAP’s huge search space. We select
a candidate list for iterations by selecting the best K
(elite) nodes within each iteration and, in the next run,
evaluate these K best nodes to select the K best nodes
among the K × K candidate nodes to be the next starting nodes. There are always K nodes evaluated in an
iteration.
Probabilistic Evaluation. To avoid local optima, evaluations are translated into probabilities of selection.
The evaluation function for selecting candidate nodes
combines a satisfaction factor that is used to tune
search directions together with a random factor.
Given a list of candidate nodes, the nodes are sorted
by decreasing profit value. The selection process starts
from a node with the highest profit value and selects
the node with a random probability. The process goes
on until K nodes have been selected. The reason for
having the random factor is to ensure that the best
K nodes are not always selected so as to avoid local
optima.
Tabu Reactive Strategy. This strategy is aimed at preventing repetition. For the SSAP, we disallow repeat
exchanges whenever an exchange/shift/add has been
applied to products and shelves.
Restarting with Elite Solutions. Whenever there are
insufficient suitable K candidates selected for current iteration, we restart. Restarting can be done with
the K elite nodes either from the previous iteration or
from preceding iterations. Restarting can effectively
help prevent searches from repeating.
Referent-Domain Optimization Strategy. This is a
learning strategy: Products selected often during
search are given increased weights.
An outline of the TS algorithm follows:
8.1. Tabu Search
The iterative TS metaheuristic was developed by
Glover (Glover and Laguna 1997). Its strategy is to
avoid being trapped in cycles by forbidding moves
which take the solution, in successive iterations, to
points in the solution space already visited. To avoid
retracing steps, the method records recent moves in
Tabu lists (Tabu Memory) which forces the search to
explore new areas of the search space. At the point of
initialization the solution space is checked in a process of “diversification,” but as candidate locations
and local optima are found, the search becomes more
focused in a process of “intensification.” In effect, TS
is basically deterministic, although probabilistic elements can be augmented to it as we do here.
For the SSAP, we first obtained an initial solution
using the original heuristic without adjustments, and
then applied TS strategies, including the following:
TABU_SEARCH
Run original allocation heuristic to get an
initial solution
Update the Elite List to contain the initial solution
while (current iteration < max iteration)
for each current node of the Elite List
for each products i, j and shelves k, l of
current node
Exchange:
evaluate (all possible exchanges)
remove Tabu active nodes
return best K nodes with exchanges
Shift:
evaluate (all possible shifts)
remove Tabu active nodes
return best K nodes with shift performed
Add:
evaluate (all possible products that can be added)
remove Tabu active nodes
8.
Metaheuristics
Lim, Rodrigues, and Zhang: Metaheuristics with Local Search Techniques for Shelf-Space Optimization
127
Management Science 50(1), pp. 117–131, © 2004 INFORMS
return best K nodes with add performed
select the best K nodes for current node
Update Elite List to contain best K chosen from
all nodes
if (Elite List contains less than K nodes)
Restart
Experimental results for the TS approach, together
with comparisons with other approaches studied, are
provided in §9.
8.2. Squeaky-Wheel Optimization
In SWO, introduced by Joslin and Clements (1999), a
construction algorithm first processes each element of
a solution in an order that is determined by priorities assigned to each element based on certain criteria. The solution is then examined to determine which
elements are positioned disadvantageously. These elements are deemed to “squeak” because they contribute negatively to the objective function of the solution. These “trouble” elements are then advanced to
the front in the ordered priority list so that the construction algorithm handles them earlier when the
next solution is constructed. This process of constructing, analyzing, and reordering is repeated, producing a variety of candidate solutions to the problem
at hand. In favorable situations, near-optimal or even
optimal solutions can be found with this procedure.
The basic approach in SWO is to form a ConstructAnalyze-Prioritize three-component cycle. The “constructor” uses priorities assigned to construct a solution, employing a greedy algorithm. The “analyzer”
assigns a numerical value as a “blame” factor to each
element that has contributed to the shortcomings in
the solution constructed in the previous step. The
“prioritizer” then modifies the priority list according to the blame factor assigned for each element, by
moving elements with greater blames to the front of
the list. We propose an enhanced five-phase algorithm
for the SSAP that is a hybrid developed by combining the SWO technique with local search heuristics.
Because the greedy heuristic by itself may not generate good solutions, SWO fine-tunes the solutions and
reorders the priority list so that “trouble” elements
can be handled earlier. This process can be viewed
as jumping between two search spaces: the solution
and prioritization spaces, where a small change in
prioritization space can result in a large transformation in the solution space. This allows SWO to find
good solutions rapidly. However, there are also many
deterministic limitations that need to be considered
carefully in designing such an algorithm (Joslin and
Clements 1999).
We designed the new five-phase SWO with Local
Search (SWOL) to overcome possible limitations. Two
new main components are added to the core SWO
Figure 7
The Five-Phase SWO with Local Search
Initialization
Initialize
ju
Ad
st
Analyzer
Bla
m
Adjustor
(Local Search)
So
lut
ion
e
Prioritizer
Constructor
o
Pr i
ze
riti
cycle: a “special constructor” that is only used for
generating the initial solution, and a “local search
adjuster” that serves to enhance constructed solutions. This extended five-phase cycle is illustrated in
Figure 7.
Initialization. Although SWO finds solutions rapidly, it does not guarantee the feasibility of solutions
generated. For the SSAP with multiple constraints, the
constructed solution tends to violate several hard constraints. To obtain feasibility of solutions, it is important that the initial solution be feasible. Therefore, a
good initial solution is important.
Blame-Factor Design. The effectiveness of the blame
factor is key to the success of SWOL. We have found
that using a satisfaction factor is appropriate. A satisfaction factor can be described as a subjective value
of how satisfactory the current allocation of a product
and a shelf is when compared with its most favorable
allocation.
To control the scalability of construction so that feasible solutions are constructed, other dynamic values,
including the allocation number and an infeasibility penalty, are combined with the satisfaction factor.
These are described in greater detail in the next section. We also attempt to control the effect misassigned
blame has on the quality of solutions. Due to the
way initial solutions are constructed, it is possible for
the lower bound of solutions to be violated. Whenever this happens, we have an indication that, in the
current solution, the blame values assigned for some
products are too low to fulfill its lower-bound requirement. To remedy this situation, we assign additional
blame value to these elements so as to force reallocation in the next iteration.
Local Search. SWO effectively avoids local optima
traps by moving between the two search spaces.
Although this characteristic of making large moves is
an advantage of SWO, it is also a limitation, causing
SWO to be poor at making small tuning moves in
the solution space. To improve on this, we combined
SWOL to explore neighborhoods of good solutions.
128
Lim, Rodrigues, and Zhang: Metaheuristics with Local Search Techniques for Shelf-Space Optimization
Another reason for employing local search is that
in certain situations, some elements should be handled badly (given lower priorities) to achieve a good
overall solution. SWO alone is unable to achieve good
solutions under such situations because high blame
is assigned to these elements, leading the search further from optimal because the constructor will handle them first. By performing local search before the
analysis component, these elements are made to keep
their low priorities.
Furthermore, the adoption of adjustments in local
search will allow the constructor to avoid taking
excessive time in evaluating unnecessary alternatives
if the previous configuration already possesses the
characteristics of a good solution.
8.2.1. Implementation. Initial Constructor. We use
Yang’s (2001) initial heuristic to get the initial feasible solution. This allocation algorithm consists of two
steps. The algorithm first checks whether total available shelf space is large enough to allocate products
in the minimum amounts required. The first round
of allocation aims to fulfill the minimum number of
allocations required for each product (referred to as
lower-bound constraints), starting with the product
having the greatest unit profit, pik /ai , to the product
with the lowest. The second round uses a greedy
algorithm to allocate each type of product until its
maximum allocation is reached, starting from the one
having the greatest unit profit to the product with the
lowest. A feasible solution is then obtained. The initial
constructor can also use a network flow solution.
Analyzer. We experimented with various methods
of assigning blame. The blame factors that result in
achieving near-optimal value performance comprise
a satisfaction factor, an allocation number, and an
infeasibility penalty. Satisfaction factors are assigned
in such a way that a poorly allocated product in
current iteration is very likely to be blamed highly
so as to get a better allocation in the next iteration.
In the implementation, each product i is associated
with a shelf best K where it achieves highest profit. If
product i is placed on shelf k, the satisfaction factor
pi best K − pik /ai is assigned. The allocation number is
the current number of facings for product i on shelf k.
It is used to ensure that the solution constructed will
not be far away from infeasibility because the starting point (initial solution) is feasible. Each subsequent
reallocation will not change the total allocation for
each product among all shelves; it will only change
the distribution among shelves. Thus, the allocation
number guides the next SWOL iteration in allocating
the number of facings of the product. However, when
the allocation is changed, it is possible that the resulting solution becomes infeasible. When this happens,
an additional allocation number is added to these
“exceptional” products with high blame value so as
Management Science 50(1), pp. 117–131, © 2004 INFORMS
to ensure they can be reallocated with more facings in
the next iteration. For pik ranging from 200 to 7,000,
the infeasibility penalty was set to be 105 .
Prioritizer. The prioritizer builds an object to record
all information from combined evaluation for each
nonzero placing. Next, it assigns allocation numbers
for each blame factor recorded in this object, together
with information on the current product or shelf and
the ideal product or shelf. It then goes on to construct
the priority list using the object by sorting the products in order of blame value from highest to lowest.
For products with zero-value blame factors, a random order will be assigned so that they will be allocated differently for subsequent iterations. A restart
is applied whenever solutions are detected to form
cycles within the solution space.
Constructor. The constructor builds a sequence of
allocations based on the priority list guided by the
prioritizer. The constructor takes an object one at a
time from the beginning to the end of the priority list.
There are many different options involved in allocation. The best option tested was chosen.
Local Search. This can be the greedy starting algorithm with the three proposed adjustment moves.
9.
Experimental Results and Analysis
9.1. SSAP
For problems of small size, an optimal profit value Po
is obtained from exhaustive enumeration, where To is
the actual time taken to compute this profit. For problems of large size, a profit upper-bound value Pub is
obtained from the network flow model, where Tub is
the time taken to compute this upper bound. For performance comparisons, the upper bound is treated as
optimal although any such upper bound is likely to be
higher than the actual optimal value. The actual profit
obtained and time taken by a heuristic are denoted by
Ph and Th , respectively.
Simulated problems were generated to test the performance of the adjustment neighborhood moves and
the other algorithms. To have accurate and complete
performance comparisons between the two sets of
adjustment methods, new parameter sets are simulated. In total, there are five sets of parameters generated: m n: the ordered pair of shelves and products;
Tk : the length of the shelf k; Li : the lower bound for
the amount of facings of product i; 0i : the upper and
lower bound difference for the total facings of product i; and ai : the length per facing of product i.
To ensure that our samples were general enough,
we used a random generator for the parameter
sets, within a set range, with a normal distribution.
Moreover, to see how selection of parameter ranges
can affect solution performance, different ranges are
tested for each parameter set.
Lim, Rodrigues, and Zhang: Metaheuristics with Local Search Techniques for Shelf-Space Optimization
129
Management Science 50(1), pp. 117–131, © 2004 INFORMS
Table 2
Parameter
m n
ai
Li
i
Tk
Table 3
Parameter Values and Ranges for Large Problems
Value
Range
5 10 5 30 5 50 5 100 10 30 10 50,
10 100 30 50 30 100 [9 values]
Random1 A
A = 5 10 30 50 100 300 [6 values]
Random0 L
L = 0 10 30 50 100 [5 values]
Random0
= 10 30 50 100 300 [5 values]
RandomTl /4 Tu Tl = i Li · ai /m Tu = i Li + i · ai /m
In experimentation with Yang’s (2001) results, we
implemented his original heuristic (Original Heuristic) without adjustments and then ran the first two
phases (Preparation and Allocation) of his algorithm,
before applying 1 his adjustment methods (Yang’s
Adjustments) and 2 our neighborhood moves (New
Neighborhood Moves). The remaining algorithms
implemented are as described in the preceding
sections: network flow, TS, and the five-phase SWOL.
We present the results for large problems first because
they are more significant.
9.1.1. Results for Large Problems with Limited
Shelf Space. To evaluate the performance of all the
algorithms when the number of shelves and products
is large and the shelf spaces are limited, we choose
the parameter sets to be as shown in Table 2.
For problems of large size, Tl and Tu values are cho
sen as Tl = i Li · ai /m and Tu = i Li + 0i · ai /m,
respectively, which are the total shelf length required
to allocate all products with the amount equal to their
lower bounds or upper bounds, respectively. Thus,
the total shelf space is generally sufficient to allocate
all products and satisfy the lower-bound constraint,
but not enough (limited) to allocate all the products to
reach the upper bounds. The profit pik has a uniform
distribution between 0 and 10.
For each the nine m n pairs, 6 · 5 · 5 · 1 = 150 test
cases were run (see Table 2), giving a total of 1350
test cases. Profit and time analysis for large problems are given in Table 3. The results for the nine
m n pairs are shown in Table 4. From the results
given in Tables 3 and 4, the new neighborhood moves
heuristic, the network flow approach, and the fivephase SWOL all outperform the original heuristic and
Yang’s adjustments. The five-phase SWOL obtains the
Table 4
Profit and Time Analysis for Large Problems (Time in
Milliseconds, 10−3 s)
Original heuristic
Yang’s heuristic
New neighborhood
moves
Tabu search
Five-phase SWOL
Network flow
Average
Ph
Average
(Ph /Pub ) (%)
2403057
2411434
2445856
9674
9724
9891
2407813
2456715
2460330
(Average Pub )
9726
9952
9946
Average
Th
Average
(Th /Tub )
198
4888
42181
004
004
015
2646649
2951272
4240279
(Average Tub )
35094
8760
10
best performance ratio of 99.52% with an average Th
of 2951272 milliseconds, well within reasonable limits.
Parameter Testing. Experiments were conducted for
different parameter sets: A, L, and 0. In all of these,
our new neighborhood moves, network flow, and
five-phase SWOL heuristics perform much better than
Yang’s method, while the TS, the weakest of the new
algorithms, does slightly better than Yang’s (2001)
method. We provide an analysis for the parameter L
below for large problems in Figure 8.
Performance for changes in the other parameters, A
and 0, were similar to performance for changes in L.
9.1.2. Results for Small Problems. Similar testing was carried out for instances when the number
of shelves and products is small. In this case, for m
ranging from 2 to 6 and n from 2 to 10, an exhaustive
enumeration algorithm was used to compute actual
optimal profits (Po ). A total of 800 test cases were
used. In these, Yang’s (2001) heuristics fared better,
with an average Ph /Po of 97.86%. All new heuristics performed better than Yang’s, with the five-phase
SWOL again achieving the best results with an average Ph /Po of 99.16%. Yang’s method obtained an average for Th /To of 0.87, while the five-phase SWOL
obtained an average of 48.55 with an average Th of
106.59 milliseconds. Details are given in Table 5.
9.2.
Product Groupings and Nonlinear
Profit Functions
The techniques developed for the linear case were
adapted to deal with these models. Specifically, we
Detailed Profit Analysis for Algorithms for Large Problems
Algorithm\m n
5 10
5 30
5 50
5 100
10 30
10 50
10 100
30 50
30 100
Original heuristic
Yang’s adjustments
New neighborhood
moves
Tabu search
Network flow
Five-phase SWOL
Best performance
9509
9580
9828
9647
9702
9904
9696
9746
9931
9715
9778
9949
9640
9683
9857
9687
9725
9891
9731
9779
9927
9673
9718
9837
9768
9800
9893
9762
9891
9934
5-p SWOL
9748
9968
9977
5-p SWOL
9734
9981
9979
Network
9736
9991
9987
Network
9668
9925
9942
5-p SWOL
9710
9957
9954
Network
9736
9979
9967
Network
9676
9880
9899
5-p SWOL
9769
9929
9941
5-p SWOL
Lim, Rodrigues, and Zhang: Metaheuristics with Local Search Techniques for Shelf-Space Optimization
130
Management Science 50(1), pp. 117–131, © 2004 INFORMS
Figure 8
Performance of Algorithms When the Value of L Changes
Original Heuristic
Y ang's A djustments
New Neigbhd. Moves
Tabu Search
Netw ork Flow
5-phase SWOL
100.00%
99.50%
Ph / Pub
99.00%
98.50%
98.00%
97.50%
97.00%
96.50%
96.00%
0
10
30
50
Values of the Parameter L
100
10.
adapted the TS and the five-phase SWOL heuristics,
which were found to be suitable for these problems
and gave good results.
9.2.1. Results for Product Groupings. We adapted
our TS and the five-phase SWOL heuristics to deal
with this model by modifying the cost evaluation
functions for each operation they performed. Experiments were conducted with parameters set as in
Table 2, where m n = 5 10 5 30 5 100
10 30 10 100 30 50 30 100; A = 10 30 50;
L = 0 10 20; and 0 = 10 20. In all, 126 test cases
were run with pik uniformly distributed between 200
and 7000 and !ij uniformly distributed over −200
and 200. Results were compared against the best
profit, Pmax , obtained for both methods. The fivephase SWOL performed best again, with an average Ph /Pmax value of 99.85% and an average Th of
101723 milliseconds.
9.2.2. Results for Nonlinear Profit Functions. We
adapted the TS and five-phase SWOL heuristics to
address this class of nonlinear problems through
Table 5
Profit and Time Analysis for Small Problems (Time in
Milliseconds, 10−3 s)
Original heuristic
Yang’s heuristic
New neighborhood
moves
Tabu search
Network flow
Five-phase SWOL
Exact enumeration
Average
Ph
Average
(Ph /Po ) %
5246
5362
5400
9584
9786
9848
5358
5293
5444
5479
(Average Po )
9773
9641
9916
100
Average
Th
122
129
135
532485
306
10659
846350
(Average To )
modifying cost evaluation functions for each operation performed by removing the linear assumption. For our experiments, we tested polynomials and
found that the heuristics were able to handle these
well, giving good results in reasonable times. For
3/2
example, taking F = cij · xij , where the coefficient cij
has a uniform distribution between 200 and 7000,
and the product-shelf range and parameters are those
for the 126 test cases for product groupings, we compared profits against the best profit instance, Pmax ,
obtained for both methods and all instances. The fivephase SWOL performed best again, with an average Ph /Pmax value of 99.97% and an average Th of
3772 milliseconds.
All experiments were run on a Pentium 4, 2.5 GHz
PC with 512 Mb of RAM.
Average
(Th /To )
084
087
090
368433
162
4855
10
Conclusion
In this paper, we explored the potential of certain metaheuristics and their hybrids for solving a
basic shelf allocation problem originally proposed
by Corstjens and Dolye (1981, 1983) and recently
simplified by Yang and Chen (1999). In our study,
we developed exact upper bounds though a network
model which we used to determine the efficiency of
each technique. Also, we developed a network flow
model for the SSAP that gave good results and could
be exploited for its network structure.
As a starting point of our study, we developed
neighborhood moves which could possibly find application in a wider range of shelf allocation problems,
as they are easily applied. We found these “manyto-many” moves to be well suited for this problem,
having worked well in improving Yang’s basic hillclimbing algorithm significantly. These new neighborhood moves were important in the development of
techniques given here and provided the good initial
solutions crucial to achieving high-quality solutions.
The potential myriad of constraints and competing objectives in shelf allocation problems, emanating
from the prerogative needs of the retailing industry
and the need to improve on current results, motivated the exploration of various AI techniques in this
work. Specifically, a number of metaheuristics and
hybrids were studied. These included TS and SWO.
To exploit these, we used the strategy of combining
a strong local search technique with a metaheuristic,
and experimentation has shown that this hybridization strategy worked better than if we used only
a metaheuristic. Indeed, the strategy of embedding
a local search within an extended five-phase SWOL
metaheuristic worked extremely well for the problem
at hand. As a step towards addressing more complex models and for testing the techniques developed,
Lim, Rodrigues, and Zhang: Metaheuristics with Local Search Techniques for Shelf-Space Optimization
Management Science 50(1), pp. 117–131, © 2004 INFORMS
we were able to apply some of the heuristics developed here to product grouping and nonlinear models,
where they continued to work well.
Near-optimal results are important to shelf allocation and even more so when product-shelf numbers are low. In practical applications, including that
of producing planograms, algorithms must consistently provide near-optimal or optimal solutions for
all ranges considered. The implementation of the
five-phase SWOL is significant in that it provides
near-optimal solutions for low-number product-shelf
ranges while giving consistently good results in all
ranges tested, for the linear as well as the nonlinear
problems studied within reasonable timescales.
Acknowledgments
The authors thank the anonymous referees for invaluable
suggestions, which helped improve the content of this work
and its presentation. Andrew Lim and Xingwen Zhang
acknowledge the funding support received from the Logistics and Supply Chain Management Institute, Hong Kong
University of Science and Technology, Brian Rodrigues
acknowledges funding support received from the WhartonSingapore Management University Research Center, Singapore Management University.
References
Ahuja, R. K., T. L. Magnanti, J. B. Orlin. 1993. Network Flows: Theory,
Algorithm and Applications. Prentice-Hall, Englewood Cliffs, NJ.
Anderson, E. E. 1979. An analysis of retail display space: Theory
and methods. J. Bus. 52(1) 103–118.
Anderson, E. E., H. N. Amota. 1974. A mathematical model for
simultaneously determining the optimal brand collection and
display area allocation. Oper. Res. 3(6) 58–63.
Anonymous. 1995. Every transaction tells a story. Chain Store Age
Executive 71(3) 50–62.
Anonymous. 1997. New data on “frozen” consumers can improve
promos: A. C. Nielsen. Frozen Food Age 45(11) 36–42.
Anonymous. 1998. Data mining is more than beer and diapers.
Chain Store Age 74(6) 64–68.
Bai, R. 2003. Personal communication. Automated Scheduling,
Optimization and Planning Research Group, University of
Nottingham, U.K.
Bookbinder, J. H., F. Zarour. 2001. Direct product profitability and
retail shelf-space allocation models. 22(2) 183–209.
Borin, N., P. Farris. 1995. A sensitivity analysis of retailer shelf management models. J. Retailing 71(2) 153–171.
Borin, N., P. Farris, J. Freeland. 1994. A model for determining retail
product category assortment and shelf space allocation. Decision Sci. 25(3) 359–384.
Bultez, A., P. Naert. 1988. SHARP: Shelf allocation for retailers
profit. Marketing Sci. 7(3) 211–231.
131
Bultez, A., P. Naert, E. Gijsbrechts, P. V. Abelle. 1989. Asymmetric
cannibalism in retail assortments. J. Retailing 65(2) 153–192.
Buttle, F. 1984. Retail space allocation. Internat. J. Physical Distribution Material Management 14(4) 3–23.
Cairns, J. P. 1963. Allocate space for maximum profits. J. Retailing
39(2) 43–55.
Campo, K., E. Gijsbrechts, T. Goossens, A. Verhetsel. 2000. The
impact of location factors on the attractiveness and optimal
space shares of product categories. Internat. J. Res. Marketing 17
255–279.
Chen, Y., J. D. Hess, R. T. Wilcox, Z. J. Zhang. 1999. Accounting
versus marketing profits: A relevant metric for category management. Marketing Sci. 18 208–229.
Corstjens, M., P. Dolye. 1981. A model for optimizing retail space
allocations. Management Sci. 27(7) 822–833.
Corstjens, M., P. Dolye. 1983. A dynamic model for strategically
allocating retail space. J. Oper. Res. Soc. 34(10) 943–951.
Curhan, R. C. 1973. Shelf space allocation and profit maximization
in mass retailing. J. Marketing 37(3) 54–60.
Desmet, P., V. Renaudin. 1998. Estimation of product category sales
responsiveness to allocated shelf space. Internat. J. Marketing
Res. 15 443–457.
Drèze, X., S. J. Hoch, M. E. Purk. 1994. Shelf management and space
elasticity. J. Retailing 70(4) 301–326.
Fanscher, L. A. 1991. Computerized space management: A strategic
weapon. Discount Merchandiser 31(3) 64–65.
Frank, R., W. Massy. 1970. Shelf position and space effects on sales.
J. Marketing Res. 7 59–66.
Glover, F., T. Laguna. 1997. Tabu Search. Kluwer Academic
Publishers, Dordrecht, The Netherlands.
Grochet, W., Y. Smeers. 1979. A branch-and-bound method for
reversed geometric programming. Oper. Res. 27(5) 982–996.
Hansen, P., H. Heinsbroek. 1979. Product selection and space allocation in supermarkets. Eur. J. Oper. Res. 3 474–484.
Helms, N. M., P. J. Haynes, S. D. Capple. 1992. Competitive strategies and business performance within the retailing industry.
Internat. J. Retail Distribution Management 20(5) 3–14.
Joslin, D. E., D. P. Clements. 1999. Squeaky wheel optimization. J.
Artificial Intelligence Res. 10 353–373.
Levy, M., B. Weitz. 1995. Retailing Management. Irwin, Chicago, IL.
Orlin, J. B. 1988. A faster strongly polynomial minimum cost flow
algorithm. Proc. 20th ACM Sympos. Theory Comput. Chicago,
IL, 377–387.
Urban, T. L. 1998. An inventory-theoretic approach to product
assortment and shelf-space allocation. J. Retailing 74(1) 15–35.
Walters, R. G., W. Bommer. 1996. Measuring the impact of product and promotion-related factors on product category price
elasticities. J. Bus. Res. 36 203–216.
Yang, M. 2001. An efficient algorithm to allocate shelf space. Eur. J.
Oper. Res. 131 107–111.
Yang, M., W. Chen. 1999. A study of shelf space allocation and
management. Internat. J. Production Econom. 60–61 309–317.
Zufryden, F. S. 1986. A dynamic programming approach for product selection and supermarket shelf-space allocation. J. Oper.
Res. Soc. 37(4) 413–422.
Inf Syst Front (2015) 17:159–175
DOI 10.1007/s10796-012-9392-7
Antecedents of cognitive trust and affective distrust
and their mediating roles in building customer loyalty
Jung Lee & Jae-Nam Lee & Bernard C. Y. Tan
Published online: 8 November 2012
# Springer Science+Business Media New York 2012
Abstract The present research investigates how trust
and distrust differently mediate in customer perceptions
of various web features in the process of building customer loyalty. Assuming trust and distrust are different
in their psychological aspects, we propose that trust is a
cognitively active construct, whereas distrust is an affectively active construct. To support this proposal, we
select six antecedents of trust and distrust and hypothesize
their different relationships as follows: 1) Antecedents
with capability-based elements, such as site convenience
and content relevance, are associated with trust; 2)
Antecedents with relationship-affecting elements, such
as customer involvement and web fraud, are associated
with distrust; 3) Antecedents with both elements, such
as content truthfulness and customer responsiveness, are
associated with both trust and distrust. A survey is
conducted on 279 online shopping mall users in
Korea, and the result shows that most of the foregoing
hypotheses are supported. The finding suggests: 1) Trust
emerges when customers expect positive result with
confidence, thereby implying that it is cognitively activated;
J. Lee
Bang College of Business, KIMEP University,
4 Abay Avenue,
Almaty 050100, Kazakhstan
e-mail: junglee@kimep.kz
J.-N. Lee (*)
Korea University Business School,
Anam-Dong Seongbuk-Gu,
Seoul 136-701, Korea
e-mail: isjnlee@korea.ac.kr
B. C. Y. Tan
Department of Information Systems,
National University of Singapore,
13 computing Drive,
Singapore, Singapore 117417
e-mail: btan@comp.nus.edu.sg
2) Distrust emerges when customers suspect that the seller has
a vicious motivation, thereby implying that it is affectively
activated. From these premises, the present study contributes
to the literature by showing how trust and distrust are different, and why they should be managed differently to establish
customer loyalty.
Keywords Trust . Distrust . Cognition . Affect . Customer
loyalty
1 Introduction
The importance of trust in online businesses has long
been discussed in voluminous research. Trust is an
expectation that the other party will not opportunistically behave by taking advantage of the situation (Gefen et
al. 2003b). Trust reduces business transaction complexity (Luhmann 1979), facilitates transactions among business parties (Moorman et al. 1993), and accelerates
economic exchanges, which all lead to high scales of
sales and profit (Barney and Hansen 1994). Trust also
has a positive impact on various essential business
factors, such as familiarity (Gefen et al. 2003a), subjective norm (Awad and Ragowsky 2008), and privacy
(Kim 2008). Therefore, trust is certainly an important
element, especially in an online environment (Ridings et
al. 2002).
In recent years, the presence of distrust in online businesses has attracted the attention of researchers due to its
destructive impact on the success of businesses (McKnight
et al. 2002). Simply put, distrust is a negative feeling on the
conduct of another person (Lewicki et al. 1998). It is an
emotional repulsion between people and can also be a fear
that one party would not care about or might hurt the other
(Grovier 1994). Distrust suppresses economic transactions
between parties (Bigley and Pearce 1998). It blocks further
160
business exchanges, especially in online businesses where
transactions are not interpersonal.
Understanding trust and distrust, such as how they emerge
and diminish and how they are related with each other, is
considered a high priority by researchers and practitioners
because of their crucial effect on businesses (Pavlou and
Gefen 2004; Lee and Choi 2011). The antecedents, mediators,
and consequences of trust and distrust have been discussed in
various contexts and perspectives (Dimoka 2010), whose aim
is to guide organizations on how to enhance trust, avoid
distrust, and outperform their competitors (Cho 2006).
However, despite a significant body of research on trust
and distrust (Komiak and Benbasat 2008), how the two can
be distinguished from each other has not been fully discussed. For example, in the past, distrust has been considered as simply the opposite of trust (Lewicki et al. 1998). At
present, however, trust and distrust are accepted as not
necessarily opposite concepts (McKnight and Choudhury
2006). They may independently emerge from the same
person (Lewicki et al. 1998) and can be manifested in
different mechanisms (Cho 2006).
The concept of distrust, possibly as a distinct entity from
trust, needs to be investigated for its critical but opposite
impact compared with trust. When two factors with opposite
effects exist, such as trust and distrust, their dynamics must
be identified to avoid misunderstanding between the cause
and the effect. Without a clear understanding of the roles of
trust and distrust, attributing trust or distrust to the success
of a business is difficult. This particular issue on the formation of trust and distrust has been highlighted in recent
research (Dimoka 2010; Komiak and Benbasat 2008).
This paper aims to identify the characteristics of trust and
distrust, especially their psychological aspects. It proposes
that trust and distrust are different in psychological status.
Therefore, people who trust and people who distrust will
respond differently to various external stimuli in online
businesses. Through a testing of the associations between
certain stimuli and trust and distrust, trust and distrust are
anticipated to be differently positioned in the cognitive and
affective dimensions and hence elicit different responses to
external factors and different mediations of customer
loyalty.
The study is organized as follows. First, the cognitive–
affective dimensions of the human mindset are reviewed. A
research model of trust and distrust in an online business
context is then proposed. Various antecedents with different
characteristics are selected to highlight the psychological
differences between trust and distrust. Large survey data
from Korean Internet shopping mall users are collected
and analyzed using structural equation modeling to validate
the hypotheses. Finally, the results of the analysis are interpreted, and the theoretical contribution and practical implications of the study are discussed.
Inf Syst Front (2015) 17:159–175
2 Theoretical background
2.1 Cognitive and affective dimensions of the human
mindset
The psychological foundations of human beings have been
extensively discussed in various studies (Fredrickson 2001;
Russell 2003) because these foundations convincingly explain why individuals behave the way they do in certain
situations. These foundations also provide a theoretical lens
for people to understand better how they think and act,
which is generally difficult to observe from the outside.
A number of researchers, including psychologists, have
investigated the brain activities of humans and have
developed strong foundations as reflected in various theories and practices (Fredrickson 2001). In the development of their studies, researchers used various ideas and
advanced tools, such as f-MRI, which were of great help
(Dimoka 2010).
The use of cognition and affect as two important, exhaustive, but highly distinguishable psychological units that mediate consumer perception on behavior has been repeatedly
practiced in the literature (Homburg et al. 2007). According to
the cognitive–affective systems theory of personality, individuals differ on how they categorize and encode situational
stimuli and on how these encodings activate such stimuli
through the complicated mediation of cognition and affect
on human behavior (Mischel and Shoda 1995). In addition,
McAllister (1995) and Chua et al. (2008) investigate the
configuration of the cognitive and affective dimensions of
trust as important business factors with distinguishable effects.
The cognitive dimension of the human mind captures
how people judge and assess an object based on facts and
evidence (Chua et al. 2008). This dimension connotes the
straightforward and conscious process of being aware of an
event. Cognition is a knowledge-based, immediate understanding of exogenous stimuli, such as a partner’s capability
and potential (McAllister 1995). It can also subsequently
form beliefs or expectations based on reason and rationale,
such that it works as a base for an individual’s further
behavior. The cognitive aspect of an individual emphasizes
the perception based on reason and evidence. Chua et al.
(2008) calls it as “referring to [the] head.”
The affective dimension captures how people “feel”
about a subject. Affect is a state of mind that arises from
one’s own emotion and a sense of others’ feeling and
motives (Chua et al 2008). It is an emotional bond between
parties which does not necessarily result from reasoning and
understanding but from feeling and sense (Morrow et al.
2004). Unlike cognition, however, affect is divided into two
types, namely, positive affectivity such as happiness and
pride (Fredrickson 2001), and negative affectivity such as
worry and anger, which need to be discussed separately
Inf Syst Front (2015) 17:159–175
because of their different functions ( Watson and Clark
1984). Chua et al. (2008) calls affect as “referring to [the]
heart.”
Regarding the relationship between cognition and affect,
whether cognitive activity is a necessary pre-condition of
emotion has long been an issue of discussion. In some
studies, the causal relationship between cognition and affect—that cognition influences affect—has been proposed
and exercised (Chang and Chen 2009; McAllister 1995).
These studies argue that the cognition unit (i.e., expectation
and beliefs) should be formed first based on encoded information in order to build affect. The expectations and beliefs
then shape affective responses with psychological reactions.
Based on this view, cognition comes before affect.
However, cognition and affect do not necessarily have
causal relations; rather, they are independent units. Zajonc
(1980) has argued that affect is precognitive in nature as it
occurs without any extensive perceptual and cognitive processes. In addition, Hoch and Loewenstein (1991) propose
that feelings of desire that consumers often experience in
shopping situations may occur with little or no cognition.
Berkowitz (1993) proposes that an experiential system,
which is affective in nature, and a rational system, which
is cognitive in nature, tend to operate in parallel in any given
task. All these studies relax the causality between cognition
and affect, but they also admit that affective reactions can
happen automatically without active higher-order cognitive
processes.
3 Research model and hypotheses
A research model of trust-distrust in online business context
is proposed (Fig. 1). Electronic commerce has now become
one of the largest business sectors (Nielson report 2008) and
numerous studies have attempted to identify the roles of
Fig. 1 Research model
161
trust and distrust in online business to understand the behavior of online consumers (McKnight et al. 2002; Pavlou
and Gefen 2004). Further, as antecedents of trust and distrust, site convenience, content relevance, content truthfulness, customer responsiveness, customer involvement and
web fraud are selected. The selection is made among the key
factors in online business (Dai et al. 2008; Tam and Ho
2006) in a way to highlight the difference between trust and
distrust. Trust and distrust are then set as cognitively- and
affectively- activated mediators, respectively. Finally, customer loyalty is selected as a dependent variable for its
significance in online business (Srinivasan et al. 2002). In
the following sections, the variables and their relations will
be described in detail.
3.1 Business value: Customer loyalty to the website
Loyalty refers to the overall attachment and deep commitment to a product, service, brand, or organization of a buyer
(Oliver 1999). Loyal customers not only spend more than
usual customers, but they also act as enthusiastic advocates
for the business (Harris and Goode 2004). Therefore, loyal
customers are considered as the most important profit source
of online businesses (Flavián et al. 2006). In the literature,
many sources of customer loyalty have been identified. For
example, factors such as customer satisfaction (Chang and
Chen 2009) and trust (Aydin and Özer 2005) are found to be
the important cursors of customer loyalty. Functional factors, such as convenience (Srinivasan et al. 2002) and website quality (Chang and Chen 2009), are also found to be the
antecedents of loyalty.
To form a definition, a group of researchers have attempted
to integrate the behavioral aspect of loyalty with the attitudinal
aspect. For example, Assael (1992) and Srinivasan et al.
(2002) view loyalty as a “favorable attitude toward the brand
(product, seller) resulting in consistent purchase over time.”
162
Keller (1993) defines it as “favorable attitude manifested by
repeated purchasing.” However, when researchers empirically
validate customer loyalty, behavioral aspects have been the
critical indicators of loyalty affecting the feasibility of
measurement.
In the present study, the referent of customer loyalty is
the website and not the seller or product. This is because the
long-term relationships of most online shopping mall users
are built based on the website and not on the specific
product or seller. For example, Amazon.com users visit
and purchase from Amazon.com because they have faith
in the website. Amazon.com users may change the product
they purchase or the sellers from which they buy, but they
hardly change the website from which they purchase.
Therefore, the current study views that loyalty is built toward the website and defines it as “the overall attachment to
a certain website with favorable attitude manifested by
repeated purchasing.”
3.2 Mediators: Cognitive trust and affective distrust
To conceptualize trust and distrust, this study investigates
the psychological aspects of both factors and proposes that
trust is a cognitively activated construct, whereas distrust is
affectively activated. Such difference can be argued based
on their roles that trust encourages customers to buy the
product, whereas distrust restrains customers from buying,
hence making them passive in the business.
In online businesses, trust is a factor that makes customers actively participate. It changes customer behavior from
not buying to buying. Customers, on the other hand, become
active when expecting positive results. Customers think,
assess, and judge the situation of a business, and if the
desired outcome is anticipated, customers build trust and
become active. In other words, unless an evident reason can
be based on rationales, customers are usually passive in the
business and do not purchase products. The tendency of
most customers is to make business decisions based on
reasoning and evidence before they decide to enter transactions with their partners. This process is cognitive in
nature.
In contrast, distrust is a factor that inactivates customers.
It blocks, inhibits, and restrains business transactions and
changes customers from being active to passive or keeps
them passive. Unlike trust, it is not only a negative anticipation of the result but can also be in various forms including “just a feeling of insecurity,” which makes customers
passive. Most customers are risk aversive (Mas-Colell et al.
1995). Hence, although the transaction has no evident reason to result in damage, customers often become passive
and decide not to buy.
Based on the idea that trust is a cognitively activated
construct and distrust is an affectively activated construct,
Inf Syst Front (2015) 17:159–175
the details of the psychological statuses of trust and distrust
are explained in the following sections.
3.2.1 Cognitively-activated trust
In online businesses, trust refers to the belief of customers
that the seller will transact in a manner consistent with the
customers’ expectations (Pavlou and Gefen 2004). Such
expectation can be formed based on the sellers’ competency
and credibility (Dimoka 2010), which conceptually include
various capabilities such as professionalism and punctuality
(McKnight et al. 2002). Trust is driven by knowledge on the
trustee and sensitive to the calculated risks and rewards
(Johnson and Grayson 2005). Hence, it often reflects the
functional competency and predictability of the trustee
(Rempel et al. 1985). Dimoka (2010) also recently stated
that trust is associated with brain areas linked to anticipating
rewards, predicting the behavior of others, and calculating
uncertainty.
These characteristics indicate that building trust is a
cognitively activated situation. Forming trust relies on
rational evaluation and available knowledge and is based
on good reasons (Jeffries and Reed 2000). Trust influences
customers to initiate and continue business with their partners based on the knowledge he acquires about the trustee,
reasoning, and judgment and not on their feelings and
hunches. Various studies on trust also stated that trust is
initiated by cognition and not affection (McAllister 1995).
Cognitively initiated trust may lead, but not necessarily, to
affectively initiated trust.
Trust has been considered critical in establishing longterm business relationships (McKnight et al. 2002). Once
trust between parties is formed, the relationship between
parties can last longer by overcoming uncertainties and
reducing perceived risks (Moorman et al. 1993). Naturally,
trust can be argued to strengthen customer loyalty, as many
previous studies have indicated (Harris and Goode 2004;
Sirdeshmukh and Singh 2002). Therefore, the following
hypothesis is proposed:
H1: Trust is positively associated with customer loyalty in
online businesses.
3.2.2 Affectively-activated distrust
In literature, distrust is defined as a strong negative feeling
regarding the conduct of another person (Lewicki et al.
1998). It is the concern that the other person does not care
about one’s welfare and thus may act in a harmful manner
(Grovier 1994). More intuitively, distrust is a frantic, fearful,
frustrated, and vengeful feeling (McKnight and Choudhury
2006). These negative characteristics of distrust suppress
transactions and result in business failures once it emerges.
Inf Syst Front (2015) 17:159–175
Distrust is an affectively activated construct that generates a repulsive force between two parties. To emerge, an
affectively activated construct does not have to go through
higher order cognitive processes (Zajonc 1980) because the
feeling often arises with little or no cognition (Hoch and
Loewenstein 1991). For example, distrust can emerge only
when the other person is suspected of being capable of
doing harmful things (Sitkin and Roth 1993). With or without strong cognitive processes, distrust, which is an affectively activated notion, can arise.
Distrust affects the behavior of an individual because it is
an affectively strong notion. Affection is one of the most
important psychological statuses that determine the subsequent action (Mischel and Shoda 1995). In this case, distrust
is a construct with a negative impression. Therefore, it also
has a negative impact on customer behavior. More intuitively, when customers suspect the competency and intention of
sellers, their fears and worries will eventually decrease their
loyalty to the sellers. Therefore, the following hypothesis is
established:
H2: Distrust is negatively associated with customer loyalty
in online businesses.
3.3 Antecedes of trust: Capability-based factors
Trust is formed when an individual can have positive
expectations on the result of the transaction. It emerges
when a partner appears to have sufficient competence to
complete the tasks on time, as promised, and has credibility
such that an individual does not have to worry about the
future. Trust is strongly determined by the capabilities of a
partner. Numerous studies have identified the partner’s specific capabilities, such as know-how and knowledge, as the
main foundations of the trust (McKnight et al. 2002).
3.3.1 Site convenience
The concept of convenience has been practiced in various
business contexts for a long time (Berry et al. 2002; Seiders
et al. 2000). However, one generic feature of convenience is
that it saves the time and effort of customers. While efficiency, which also involves a reduction in time and costs of
the business, considers the organization as referent, convenience involves a reduction in the time and effort of individuals, not of the organization, and considers the customer
as the referent.
In online business literature, convenience has been especially important because it is one of the powerful drivers that
lead consumers from offline to online purchases. Many
people started to do online shopping because of the convenience involved (Bhatnagar et al. 2000). Online shopping
mall users face various situations, and each time, the feature
163
of convenience is of different levels whenever they access,
search, and pay, among others. Accordingly, numerous ecommerce studies have been conducted on convenience and
have identified the sub-concepts, antecedents, results, and
dimensions of convenience (Fassnacht and Koese 2006;
Seiders et al. 2007). The original concept of convenience
is simple but it has well developed with enriched contexts in
e-commerce.
Websites that are conveniently designed help customers a
lot, and business is also smoothly facilitated. In turn, customers who experience convenience when purchasing from
a website will have positive expectations about the business
because of the time and effort they saved in transactions and
the easy achievement of their goal. A conveniently designed
website is one of the competencies of online shopping malls,
which has also become the foundation of the development
of trust for customers. From this, the following hypothesis is
proposed:
H3: Site convenience is positively associated with trust in
online businesses.
3.3.2 Content relevance
Relevance refers to the meaningful relation between two
objects under a certain circumstance, which are engaged in
a common activity and have a purpose (Allwood 1984). The
two relevant objects are not simply related with each other
but are “meaningfully” connected. In online businesses, this
meaningful connection is practiced when a customer
searches for “relevant” product information from data on
the web. Customers with certain purchase purposes look for
relevant product information because not all web data are
relevant to customers. Hence, content relevance in this study
would be defined as the meaningful relation between product information on the web and the customers whose purpose is finding product information about the product or
service that he is interested in.
For these reasons, content relevance — how much the
information on the web is relevant to the objective of the
web browsing — is considered important in online businesses. It was proposed as a sub-concept of information
quality in Delone and Mclean’s IS success Model (1992)
and also considered as an important factor in web personalization (Tam and Ho 2006). A relevant web content enhances the accuracy and speed of information processing of the
customers, thereby making the decision-making process
more effective. Eventually, a relevant web content increases
user acceptance (Tam and Ho 2006) and exerts a positive
effect on the flow experience of the user (Jiang et al. 2009).
Relevant web content, including product information,
expectedly increases the positive expectation of customers
regarding the business transaction. It helps customers make
164
decisions faster and more accurately. In turn, customers will
perceive the higher level of information quality on the web.
From this, we propose the following hypothesis:
H4: Content relevance is positively associated with trust in
online businesses.
3.4 Antecedes of distrust: Relationship-affecting factors
Distrust is formed when it is observed that the partner has
bad motivations or does damage on purpose. Distrust can be
defined as feelings of hurt, suspiciousness, or fear of a
partner’s intentions and future behavior. It is not based on
capability or skill of a partner that can be objectively
assessed; rather, it is more related to the sincerity and
motivation to continue the relationship. Thus, relationshipaffecting factors (e.g., involvement and responsiveness) that
show the motivation and intentions of a partner become the
foundations of distrust (Goodman et al. 1995).
3.4.1 Customer involvement
Customer involvement embraces the number and type of
activities that a customer engages in together in the course
of their regular economic transactions with organizations
(Goodman et al. 1995). It determines the strength and depth
of the relationship between the customer and the organization (Goodman et al. 1995). For example, the customers can
increase their involvement level by actively participating in
the service and product design. Tightening the feedback
loop between consumption and product can also enhance
customer involvement level (Lundkvist and Yakhlef 2004).
The impersonality of online business makes customer involvement more difficult than in offline businesses.
However, services such as special contract, active customization or refund flexibility, can increase the level of customer involvement between online sellers and customers.
Customer involvement is a relationship-affecting factor
that represents the total quality (strength and depth) of the
relationship between the customer and the organization.
Frequent involvement by the customers strengthens the
bond between the customer and the seller or organization.
Customers would intuitively feel that they have a good
relationship with the sellers if they are fully acknowledged
and deeply involved in business operations. These activities
affect the customers in their affective dimension, that is, the
greater the extent of the customers’ involvement, the greater
their sense of control, thereby lessening fear and worry in
their transactions. Based on this premise, we derive the
following hypothesis.
H5: Customer involvement is negatively associated with
distrust in online businesses.
Inf Syst Front (2015) 17:159–175
3.4.2 Web fraud
Web fraud refers to a broad category of crimes conducted by
online users through website transactions. Due to the impersonality, anonymity, and information asymmetry between
online sellers and customers (Turban et al. 2010), the web
environment is vulnerable to a variety of fraudulent acts,
such as misrepresentation, failure to ship, and improper
handling (Chua et al. 2007). Web fraud is different from a
seller’s typographical error or simple information mistake
due to confusion; it refers to an intentional misconduct in a
web-based business transaction to deceive customers.
Auctions are commonly the settings where vast majority of
online fraudulent acts are committed (Abbasi et...
Purchase answer to see full
attachment