-
nsum
et
ar
. da M
ISCT
Abstract
T d is
eva o ove
for is con
are prior
ou ion t
res ided
r
Keywords: Clustering; External validation; Experts knowledge
integration
1. Introduction
to justify the new trends in retailing. In the food retail,
inparticular, after an unprecedented period of hypermar-kets
growth, since the late 1970s, both in number and
from 28% to 34% in the A.C. Nielsen universe. In 1997,the
supermarkets reached the leadership and consoli-
more demanding, force the retail groups to invest inoutlets of
smaller dimension, and so in a proximity and
ARTICLE IN PRESSquality of goods and services strategy. This
investmenthas a longer run return as well as smaller economies
of
0969-6989/$ - see front matter r 2004 Elsevier Ltd. All rights
reserved.
doi:10.1016/j.jretconser.2004.11.005
Corresponding author.As in Europe, the retail sector in Portugal
is goingthrough a restructuring phase. Several authors (e.g.Birkin
et al., 2002; Dawson, 2000; Seth and Randall,1999) identify such
factors as increasing consumermobility, increasing electronic
commerce, changinghousehold size, concentration of market power,
homemarket saturation, and changes in planning legislation
dated its expansion strategy. According to the mostrecent data,
in 2001, supermarkets sales were alreadybroadly superior to the
sales in hypermarkets: 47%against just 35% of the total sales of
outlets withalimentary products (see Fig. 1).This change in food
outlet type is also found in other
European countries (Birkin et al., 2002). Much morestringent
legislation and the fact that consumers areA supplementary exercise
in cluster descriptioninvolves the investigation of the clusters in
order toestablish whether or not they can be given
substantiveinterpretations (y). Such substantive descriptions donot
make direct use of data, but require investigatorsto reect on the
results of classication studies.Gordon (1999)
market share, it is now clear that hypermarket activityhas
slowed down signicantly on behalf of the small ormedium
supermarkets (chain outlets including discountand hard discount
chains) that nowadays present alarger dynamism.In Portugal, market
share data shows that since 1996
the supermarkets were the only ones to grow simulta-neously in
the number of outlets and in the volume ofsales and, consequently,
to increase the market sharehis work is part of a supermarket chain
expansion study an
luation of outlet performance and new outlet site location.
T
a very small number of existing outlets) experts knowledge
compared for this end, the experts being required to: (1) a
tlets; (2) a posteriori: evaluate results from alternative
regress
ults from alternative dendrograms. The later approach prov
2004 Elsevier Ltd. All rights reserved.Journal of Retailing and
Co
Clustering supermark
Armando B. Mendesa,, MaMathematical Department, Azores
University, R
bDepartment of Quantitative Methods, Business Schooler Services
13 (2006) 231247
s: the role of experts
garida G.M.S. Cardosob
ae de Deus, 9501-801 Ponta Delgada, Portugal
E, Av. das Forc-as Armadas, 1649-026 Lisboa, Portugal
intended to cluster the existent outlets in order to support
the
rcome the curse of dimensionality (a large number of
attributes
sidered in the clustering process. Three alternative
approaches
i: provide values for perceived dissimilarities between pairs
of
rees; (3) interactively: help to select base variables and
evaluate
the best results according to the marketing experts.
www.elsevier.com/locate/jretconser
-
scale, which forces careful decision-making (McGol-drick, 2000;
Salvaneschi, 1996). Because smaller outletsare heterogeneous in
aspects as location, dimension, andclient behaviour, the denition
of outlet clusters isessential in outlet performance and site
evaluation.In Sections 24 the next sections, an empirical
classication of variables used in measuring super-markets
performance and the role of expert knowledgein clustering and
marketing applications is presented.The data collection phase is
described and threeapproaches for experts knowledge integration in
super-market cluster are explained. In Section 5 theseapproaches
are compared and a cluster proling ispresented. This paper nishes
with conclusions and amethodological discussion.
2. Clustering supermarkets and the role of experts
This work is part of an expansion study of asupermarket chain
with small and medium dimensionoutlets and is intended to cluster
the existent outlets.The classication is not just useful to
evaluate therelative performance of different locations and
outletmanagement, but also to use in analogy forecastmethods for
the identication of potential site locations(Mendes and Themido,
2004). For that purpose, severalperformance measures and other
attributes were col-lected in a framework dened in this section.
Foraddressing the high dimensionality of the data, theintegration
of expert knowledge in the clustering ofsupermarkets is
suggested.
2.1. Measuring supermarkets performance
The retailers soon realized the importance of outletlocation,
but understanding all the aspects of outletperformance, site
locations, and the consumers beha-viour, forces to collect enormous
amounts of informa-
ARTICLE IN PRESS
Hypermarkets
Supermarkets (>'s)
Supermarkets (
-
inaconuses
kuofex
doprga
ARTICLE IN PRESSRetailscale development (Hardestya and Bearden,
2004),and in expert systems and automatic methods(Turban et al.,
2004; Guijarro-Berdinas and Alonso-ranBeassessed (Moutinho et al.,
2000; Wedel and Kama-ra, 2000; Jain and Dubes, 1988). In this work,
the useexpert knowledge is suggested for non-quantitativeternal
validation.The use of expert knowledge, sometimes namedmain
knowledge, in order to evaluate the quality ofoblems solutions, is
generally applied and investi-ted in diverse areas: marketing
applications (Ow-g, 2000; Pasa, 1996; Moutinho and Brownlie,
1994),bed comparisons with other structures. In fact, cluster-g
validation is of major importance in order tocomplish the study
objectives (Gordon, 1999). Whenly a small number of observations
are available, thee of experts based external validation
becomessential, since quantitative internal reliability cannotanand
in an extensive literature review. The variables aredivided in
three groups:
The location and outlet attributes that are intended toevaluate
aspects only outlet and site dependent as theoutlet
characteristics, the accessibilities, and the imageof the chain or
the services range offered. Among theoutlet characteristics, the
commercial or sales area isthe factor of major importance, which is
emphasizedin Fig. 2 by an independent branch (Themido et al.,1998;
Salvaneschi, 1996).
The outlets influence area attributes that are relatedwith the
evaluation of the trade area (or catchmentarea) generated by the
outlet, which is essential inpotential sales forecasting. These
attributes aremainly demographic variables but also refer to
theimpact of the existent competition.
Clients characteristics that refer to their
preferences,attitudes, behaviour, socioeconomic prole and
geo-graphic location are, nally, relevant in the evaluationof
outlet performance.
Few works attempted to classify outlet and siteevaluation
variables. One good exception is the workpresented by Clarke et al.
(2003b), which is largelycoherent with Fig. 2. These authors used
cognitivemaps, based on answers of location experts from thelargest
retail chains in United Kingdom, to identify themain variables used
in location decisions. This workconrms not only the suggested
framework of thevariables but also the high volume of data required
inthe outlet/site evaluation studies.
2.2. The role of experts
Clustering methods always impose a structure to thedata, which
must be validated, using specic measures
A.B. Mendes, M.G.M.S. Cardoso / Journal oftanzos, 2002; Adelman
and Riedel, 1997). Also invisual cluster validation, there is some
kind of expert,or at least, user-assisted validation and
interpretation(Hathaway and Bezdek, 2003; Hennig and
Christlieb,2002; Jones, 1996).In what concerns marketing
applications there is a
long tradition in expert knowledge integration. How-ever, this
integration is seldom formalized, opinionsbeeing integrated in the
results in a non-explicited, andoften not reported, way. Exceptions
can, however befound in some areas of application: Hardestya
andBearden (2004) in consumer-related scale developmentresearch,
Sanders and Ritzman (2004) in forecastingintegrating judgments,
Owrang (2000) in discoveringnew knowledge in large volumes of data,
Pasa (1996) inthe use of marketing expertise in marketing
decisions,and Clarke et al. (2003b) in analogue outlet
identica-tion and site assessment.Owrang (2000), for instance, uses
domain or back-
ground knowledge, in guiding and restricting the searchfor
interesting knowledge. Several mechanisms aresuggested and compared
for this end. On the otherhand, Pasa (1996) constructs a
theoretical model thataids in evaluating marketing expertise and
observe howdifferent market conditions inuence whether compa-nies
emphasize marketing expertise in marketing deci-sions. He concludes
that greater market instability andbigger market presence increase
marketing expertisevalue.Still in the marketing literature, Clarke
et al. (2003b)
use soft systems methodologies, integrating qualita-tive
expert-based intuition with a structuring mechanismfor analogue
outlet identication and site assessment. Inthis work, authors
recommend qualitative expert-basedor soft approaches as a
complement to hard modellingsystems and they attempt to do this by
isolating thedifferences between a given new site and the
closestanalogue store outlet, based on multiple dimensions.
Inprevious papers, the theoretical basis for managementjudgement
and intuition were underlined (Clarke andMackaness, 2001); a method
for synthesizing expertjudgements was outlined (Clarke et al.,
2000); and theuse of a system in live decision-making environments
topromote argumentation within the decision-makinggroup was
described (Clarke et al., 2003a).In the pattern recognition
literature contributions like
Liu and Samal (2002) and Pedrycz (2004) may bereferred. Liu and
Samal (2002) advocate that, bydenition, the clusters represent same
abstract conceptthat is clearly domain dependent. Pedrycz
(2004)mention the benecial aspect of incorporating domainknowledge
in the fuzzy clustering mechanisms. Forjustifying the use of this
knowledge, he suggests that anumber of essential features may not
be available orcould not be easily quantied.Another pattern
recognition example is the utilization
ing and Consumer Services 13 (2006) 231247 233of a panel of
experts to interpret classication rules,
-
ARTICLE IN PRESSRetail3. Data collection
A large number of variables were collected in order toaccount
for the diversity of attributes that may inuenceoutlets performance
evaluation. This diversity of baseclustering data is considered
essential (see for exampleWedel and Kamakura, 2000). Of all data
collectionprocedures, explained in the next sections, a total of
250variables were obtained, measured in all kind of scales,and
covering all the aspects in the suggested variableframework (Fig.
2).
3.1. In shop surveys
Shop surveys were held in two different years duringthe study.
The rst took place in 2001 and wasaccomplished in all the existent
supermarket outlets, inall days of two successive weeks, totalizing
3766 validquestionnaires. The second was conducted in 2003, in
aselected group of outlets and in selected days of theweek, in a
total of 2394 valid questionnaires.The questions included the
clients opinions regarding
the conguration of the outlet, accessibilities and
siteconguration. They provided customers demographicand
socioeconomic characterization, attitudinal andbehavioural
attributes (motivation, means of transpor-tation, choices and
preferences) and the identication ofpresented by Bay and Pazzani
(2000). Their workconcludes that many of the generated rules are
uselessor redundant and although it points for the subjectivityof
the experts interpretations, it conrms the need forthis type of
knowledge. In spite of that, few works havebeen presented
considering the explicit integration ofexpert knowledge in feature
selection and externalvalidation for the clustering analysis (see
Jain et al.,1999, for a complete survey).Several authors, as Liu
and Samal (2002) and Halkidi
et al. (2001), propose the use of external validationindexes for
measuring the degree of agreement betweenexpert delineated clusters
and the ones obtained from amathematical method.In the present
application, expert knowledge is
suggested for non-quantitative external validation of
asupermarkets clustering structure. The experts aremarketing
analysts specialized in food retail outletlocation, working with
the supermarket chain since itsorigin and being responsible for all
location andperformance studies. This group of experts could
notagree in a cluster structure for the supermarket
outlets.Clustering supermarkets was considered a difcult
andsubjective task. Therefore, alternative approaches toexpert
knowledge integration are adopted, which are notbased in expert
delineated clusters.
A.B. Mendes, M.G.M.S. Cardoso / Journal of234the
competition.Quantitative variables as the percentage of
customersthat come from home, the average monthly expenses in
theoutlet or the average value of purchase were madeavailable
through the survey. When no signicantdifferences were observed
between the average valuesreferring to 2001 and 2003, average
values yielding fromthe two surveys were used. In the rare cases
were pairedby outlet sample t statistic where lower than 5% (e.g.
forthe mensal expenses in food), the most recent value
wasused.Client segmentation with data from both surveys was
performed resulting in two segments. The segments
werecharacterized and termed as preferential costumers andeventual
costumers (Cardoso and Mendes, 2002). Inconsequence, the percentage
of preferential costumerswas included as a new variable in the
study.
3.2. The mystery shopping programme
A mystery shopping programme (e.g. Wilson, 2001),was
accomplished with a visit to the outlet of anincognito analyst that
observed visible aspects, did abuy, and evaluated several aspects
in ordinal scales.These in loco observations were performed in all
thechain outlets and in some of the most importantcompeting
shops.They used a check list with several location attributes,
outlet characteristics, accessibilities, outlet visibility,
andsome related with competition and characterization ofthe
influence area. The variables are mainly nominal butsome are
evaluations of some aspects of the outlet andservice in an ordinal
scale of nine points.Coordinated with the mystery shopping
programme
the outlets GPS coordinates were also collected alongwith their
nearby competitors. This GPS coordinatesand the mystery shopping
data were loaded in aGeographical Information System and used to
deneinuence areas, and to calculate variables used in theoutlet
characterization and clustering.
3.3. Census, geographic and competition data
A large number of quantitative variables are availablefrom the
national geographical base of census 2001 data.This is high-quality
demographic data, accessible inseveral disaggregation degrees, and
ready to use in aGeographical Information System. To include this
data,inuence areas must be dened along with criteria forgeospatial
intersection between these areas and thedemographic areas.
Influence, trade or catchment areas can be dened asan area
around the outlet from where it is likely to drawclients. Several
methods have been suggested for itsdelimitation (e.g. Boots, 2002;
Birkin et al., 2002;McMullin, 2000), in the present case, shortest
paths
ing and Consumer Services 13 (2006) 231247polygons and
multiplicative weighted Voronoi diagrams
-
4.1
ARTICLE IN PRESS
oi di
Retailwere applied (Fig. 3). The latter method
allows,simultaneously, incorporating the outlet attractivityand the
competition in the outlet proximities (Bootsand South, 1997). A
data base with the location of morethan 600 food outlets in
Portugal was necessary for themethod implementation.Using space
interaction procedures, percentage values
and densities were calculated for all existing outlets. Bythe
end of this process, and in spite of having made acareful selection
of the census data, it resulted in morethan half a thousand
variables. To reduce this numberthe Pearson correlation coefcient
matrix were calcu-lated and the variables with more very strong
correla-tions were deleted.Fig. 3. Shortest path polygons (left)
and multiplicative weighted Voron
area and demographic polygons shaded by population
density).4.
secsmfoTowe
1.
2.A.B. Mendes, M.G.M.S. Cardoso / Journal ofMethodological
approach
In spite of the data abundance following from lasttion, the
number of outlet supermarkets was veryall. This fact hindered the
process of variable choicer the outlet clustering and respective
characterization.overcome this difculty, three different
proceduresre considered for experts knowledge integration:
In a priori integration approach, the experts wererequired to
compare pairs of outlets and evaluatetheir dissimilarities in a
perceptual scale. The dissim-ilarities matrix was then directly
used in the Wardshierarchical clustering procedure. Finally, the
selec-tion of clusters proling variables relied on regres-sions
over MDS dimensions.In a posteriori approach, the integration
wasaccomplished by evaluating alternative clusteringstructures
derived from a supervized learning proce-dure using regression
trees. In this approach the base
opisowhacve
sositThobinvexususdeotavwe
fupr. A priori experts knowledge integration
In this approach the integration of the expertsvariables choice
relied on the regression tree proce-dure although experts required
diverse regression treeparameterizations and target variables.
3. In the last procedure an interactive process wasadopted: the
experts knowledge was considered insuccessive stages regarding the
choice of clusteringvariables and the evaluation of clustering
results. Thisprocess is termed interactive integration.
WardsHierarchical procedure was based in alternative setsof base
variables in order to provide betterclustering, as judged by
experts.
agrams (right) examples. (Point radius proportional to the
outlet sales
ing and Consumer Services 13 (2006) 231247 235inion was made by
means of outlet paired compar-ns. The experts were requested to ll
a questionnaireere pairs of outlets were compared and
evaluatedcording to a scale of ordinal dissimilarities (from 1 ry
similar outlets to 9 distinct outlets).The comparison was meant to
be generic, althoughme aspects as location, management performance
ande as well as clients characterization were emphasized.e
resulting symmetrical dissimilarity matrix wastained by consensus
among the several expertsolved. This procedure is termed a priori
as theperts opinion only regards the dissimilarities matrixed as
clustering base. Clusters where then obtaineding the hierarchical
Wards method resulting in thendrogram presented in Fig. 4. It
should be noted thather hierarchical methods as centroid and
grouperage linkage were tried and similar dendrogramsre
obtained.From the observation of the fusion index values andsion
index variations versus number of clusters chart,esented in the
same gure, six clusters were adopted.
-
ARTICLE IN PRESS
tly applied to the dissimilarity matrix and fusion indexes
chart.
RetailingMore complex stopping rules for determining thenumber
of clusters were considered (see very good textsin Everitt et al.,
2001; Gordon, 1999). Although
Fig. 4. Dendrogram of the Ward hierarchical method direc
A.B. Mendes, M.G.M.S. Cardoso / Journal of236conicting results
were obtained, in general, the sixclusters cut was supported.In
order to interpret the obtained clusters some
further exploration of the dissimilarity data wasrequired.
MDSMultidimensional Scaling non-metricanalysis, using the ALSCAL
algorithm by Takane,Young and Leeuw (Cox and Cox, 2000) was
performed.A solution with four dimensions was found whichaccounts
for an RSQ of 96% (Kruskal stress value of7.8%). Fig. 5 illustrates
the positioning of chain outletsclusters in the extracted MDS
dimensions, along withlabels based in the clusters
characterization.Regression procedures using several hundreds
of
target variables were performed considering the MDSdimensions as
explanatory variables. Results enabledthe identication of variables
depicted in Fig. 6.From this gure, dimension 1 is related to
outlet
dimension and car parking facilities and inverselyrelated with
outlet visibility and sales per outlet area.Dimension 2 is related
to influence area dimension andpercentage of households with
children or working inprimary or secondary sectors.MDS dimension 3
is related with the number of elder
residents and preferential costumers in the inuencearea, and
dimension 4 is associated with the percentageof occasional clients
and complex trip (passage) clients.These results helped to support
the clustersprFi
Fig
proand Consumer Services 13 (2006) 231247oling, which may be
summarized as follows (seegs. 5 and 6):
Small outlets: A very homogenous cluster of eightoutlets
characterized by very negative values indimension 1 and 3.
According to the characterization
. 5. Outlets in the space of the MDS extracted dimensions
and
le labels.
-
ARTICLE IN PRESS
location and outlet attributes
influence area attributes
clients attributes
number of salesoperatorsworkingplacesin shop
parkingplaces
salesturnover for square meter of sales area
mysteryshopping
evaluation of walking trip
outletvisibility
mysteryshoppingevaluation of car parking facilities near the
outlet
0
0.2
0.4
0.6
0.8MDS dimension 1
MDS dimension 2
MDS dimension 3
MDS dimension 4
percentageof residents working in primary or secondarysector
percentageof resident femalechildren less than 5 years old
Influencearea by Voronoidiagramspercentage
ofhouseholdswith children orgrandchildren less than 7 years
old
number of residentwomen with more than 65 years old
number of residentfamilies with 1-2 persons
0
0.2
0.4
0.6
0.8MDS dimension 1
MDS dimension 2
MDS dimension 3
MDS dimension 4
percentageof passage respondents
preferentialrespondents
percentageofrespondentsdeclaringoccasionalpurchases
percentageofrespondentsliving in a 5 minuteswalkingdistance
0
0.2
0.4
0.6
0.8MDS dimension 1
MDS dimension 2
MDS dimension 3
MDS dimension 4
Fig. 6. MDS dimensions characterization based on standardized
absolute value regression coefcients (black marks represent
attributes negatively
related with MDS dimension).
A.B. Mendes, M.G.M.S. Cardoso / Journal of Retailing and
Consumer Services 13 (2006) 231247 237
-
4.
tiore
19anmmCa
coraco(seav
am20th
ARTICLE IN PRESSRetailSupervized learning methods rely on
enormousounts of data for internal validation (Hand et al.,01;
Berry and Linoff, 1997). Since, in this applicationeaux and Gayet
(2001) and Chou et al. (2000).Alternative target variables were
considered for treenstruction: the sales turnover for several years
and thetio of sales turnover over the sales area that is a verymmon
outlet performance measure in the literaturee for example Birkin et
al., 2002). All the remainingailable variables were considered as
predictors.chassication and Regression Trees (Breiman et al.,84).
Regression trees simultaneously cluster outletsd forecast the
outlet turnover based on the targetean values in the tree leafs.
Recent decision treearketing applications can be found for instance
inrdoso and Moutinho (2003), Cooley (2002), Mi-Cl2. A posteriori
experts knowledge integration
In the second approach, experts knowledge integra-n is made a
posteriori, in the evaluation of alternativesults provided by a
regression tree method: CARTdimension 3 and low-to-medium values in
4. There-fore, these are outlets with big Voronoi inuenceareas, and
consequently low levels of competition,and especially low levels of
preferential costumers butalso mean to low levels of passage
clients.of these dimensions, these are small outlets with
smallinuence areas indicating high levels of competition.
Neighbourhood outlets: Constitute a ve outlet clusterprimarily
characterized by low values in dimension 4related to percentage of
passage clients and low valuesin dimension 2 related to inuence
area dimension. Sothis cluster has the smallest inuence areas
andpercentage of households with children and also veryfew
occasional and passage clients.
High-potential outlets: Four outlets with low-to-medium values
in dimension 2 and high values indimension 3. In consequence
corresponds to outletswith small-to-medium inuence areas but
highpercentages of preferential costumers justifying
thehigh-potential label.
Small high-potential outlets: A three-outlet clusterwith very
high values in MDS dimension 3 and verylow values in dimension 1.
This is a very high-potentialcluster with many preferential
costumers and percen-tages of households with children but very
small salesarea.
Big outlets: Two very big stores as the high values ofdimension
1 conrm. Both have negative values indimension 2 indicating high
levels of competition andsmall inuence areas, and high values in
dimension 4indicating also high percentages of passage clients.
Low-potential outlets: Two medium size outlets withhigh values
in dimension 2, very law values in
A.B. Mendes, M.G.M.S. Cardoso / Journal of238e number of cases
is very low compared to the numbershould be noted that the biggest
outlets are excludedfrom this cluster as were split in the rst tree
node. Andso, these are not necessarily the biggest inuence areasfor
the existent outlets.The last split variable was calculated as the
percentage
of inquire respondents, which claimed to spend at least75% of
the mensal expenses in food on the outlet andthe rest in a
hypermarket. As we found that these loyalcostumers had residences
near the outlet, this cluster waslatpocaHigh-potential outlets are
characterized by all theter splits and big influence areas by the
Voronoilygon methods and correspond to bigger sales. Itoutlets
sales area and also to the biggest outlet annualsales.Low-potential
outlets have small-to-medium dimen-sion, have small number of
children in the influencearea and very low sales turnover.In Fig. 7
the best tree is presented. This tree wasevaluated by experts who
agreed that all the splits madesense and the clusters in the
terminal nodes were alsoreasonable. The corresponding leave-one-out
estimatefor the percentage of explained variance, referred to
the2002 turnover, is 86%. However, it should be noted thatbetter
values were found for trees that were evaluated aspoorer by the
experts.Clusters proling may be directly derived from the
tree, which was the most appreciate characteristic of
thisapproach as it greatly facilitates expert cluster valida-tion.
Thus, clusters were named according to thepropositional rules
associated with the correspondingleaf nodes (see Fig. 7):
Big outlets correspond to the bigger values for thebigger sales
area corresponds to a leaf with meansmaller annual
turnover).Leave-one-out error estimates were presented to
theexperts for tree selection decision support.of attributes,
special attention must be given to themodel stability issue. In
fact, several different splittingvariables may lead to similar
precision index values or,alternatively, the adoption of only
slightly differentbranching rules may signicantly decrease
precision.Moreover, the fact that there are so few observationsturn
the use of test set error estimates impossible. Inorder to overcome
these limitations:
Several trees with different parameterizations andalternative
quasi-tied splitting attributes were grown.
Only trees with meaningful decision rules wereconsidered by
experts as alternative models. Inparticular, splitting variables
were rejected whencounterintuitive decision rules emerged (e.g. if
a
ing and Consumer Services 13 (2006) 231247lled neighbourhood
outlets and the other transit outlets.
-
ARTICLE IN PRESS
A
Retail
>
