Vincent P. Magnini, Earl D. Honeycutt, Jr. and - Sage Publications

http://cqx.sagepub.com

Quarterly Cornell Hotel and Restaurant Administration

DOI: 10.1177/0010880403442009 2003; 44; 94 Cornell Hotel and Restaurant Administration Quarterly

Vincent P. Magnini, Earl D. Honeycutt, Jr. and Sharon K. Hodge Data Mining for Hotel Firms: Use and Limitations

http://cqx.sagepub.com The online version of this article can be found at:

Published by:

http://www.sagepublications.com

On behalf of: The Center for Hospitality Research of Cornell University

can be found at:Cornell Hotel and Restaurant Administration Quarterly Additional services and information for

http://cqx.sagepub.com/cgi/alerts Email Alerts:

http://cqx.sagepub.com/subscriptions Subscriptions:

http://www.sagepub.com/journalsReprints.navReprints:

http://www.sagepub.com/journalsPermissions.navPermissions:

at SAGE Publications on December 2, 2009 http://cqx.sagepub.comDownloaded from

http://www.chr.cornell.edu

http://cqx.sagepub.com/cgi/alerts

http://cqx.sagepub.com/subscriptions

http://www.sagepub.com/journalsReprints.nav

http://www.sagepub.com/journalsPermissions.nav


FOCUS ON RESEARCH DATA MINING

94 Cornell Hotel and Restaurant Administration Quarterly APRIL 2003

© 2003, CORNELL UNIVERSITY

Data Mining forHotel Firms:Use and Limitations

Data mining can provide a window into customers’ behavior—if it’s handled correctly.

In the hotel industry, knowing your guests—where theyare from, how much they spend, and when and on whatthey spend it—can help you formulate marketing strate-

gies and maximize profits. Fueled by the proliferation of cen-tralized reservation and property-management systems, hotelcorporations accumulate large amounts of consumer data. Thisinformation can be organized and integrated in databases thatcan then be tapped to guide marketing decisions. However,identifying important variables and relationships located inthese consumer-information systems can be a daunting task.The relatively new process known as data mining can be in-strumental in overcoming such obstacles.1 From stores of in-

BY VINCENT P. MAGNINI, EARL D. HONEYCUTT, JR., ANDSHARON K. HODGE

formation, data-mining technology extracts meaningful pat-terns and builds predictive customer-behavior models that aidin decision making.2

Data mining is a largely automated process that uses sta-tistical analyses to sift through massive data sets to detect use-ful, non-obvious, and previously unknown patterns or datatrends.3 The emphasis is on the computer-based explorationof previously uncharted relationships (i.e., using “machinelearning” methods that typically require only limited humaninvolvement).4 Without data mining, valuable marketing in-

1 For a discussion of the use of compiled data, see: Robert K. Griffin, “DataWarehousing: The Latest Strategic Weapon for the Lodging Industry?,”Cornell Hotel and Restaurant Administration Quarterly, Vol. 39, No. 4(August 1998), pp. 28–35. For a discussion of the use of guest-historydata, see: Paula A. Francese and Leo M. Renaghan, “Data-base Marketing:Building Customer Profiles,” Cornell Hotel and Restaurant AdministrationQuarterly, Vol. 31, No. 1 (May 1990), pp. 60–63.

2 See: A. Kamrani, W. Rong, and R. Gonzalez, “A Genetic AlgorithmMethodology for Data Mining and Intelligent Knowledge Acquisition,”Computers & Industrial Engineering, Vol. 40, No. 4 (2001), pp. 361–377.

3 W. Frawley, C. Piatetsky-Shapiro, and C. Matheus, “Knowledge Discov-ery in Databases: An Overview,” AI Magazine, Fall 1992, pp. 213–228.

4 P.R. Peacock, “Data Mining in Marketing: Part 1,” Marketing Manage-ment, Winter 1998, pp. 9–18.



APRIL 2003 Cornell Hotel and Restaurant Administration Quarterly 95

DATA MINING FOCUS ON RESEARCH

sights about customers’ characteristics and pur-chase patterns may remain largely untapped.5 Byuncovering such previously unknown relation-ships, managers have the potential to develop awinning marketing strategy that increases theirhotel’s bottom line.

Hotel managers understand the importanceof adapting to the changing business environmentnot only to remain competitive, but merely tosurvive. As a result, technology has become a largeand growing expense for many hotel corpora-tions. Under such a technology framework, datamining is a valuable competitive tool beingadopted by hotel corporations in an effort to cre-ate customer value. However, given the impor-tance and complexity of data mining, senior ho-tel managers report a low level of understandingabout data mining’s capabilities, how it works,and what value this technology contributes.6 Thepurpose of this paper is to educate hotel manag-ers about the benefits and application of datamining on the properties they oversee.

Data Mining vs. Statistical ModelingData mining differs from traditional statisticalmodeling in a variety of ways. Data mining fo-cuses on machine-driven model building, whilestatistical modeling stresses theory-driven hy-pothesis testing. Data-mining techniques buildmodels, whereas classical statistical tools are su-pervised by a trained researcher who possesses apreconceived notion of what to examine. Withstatistical a priori analysis, relevant associationsmay be overlooked. By building dependency hy-potheses instead of merely verifying them,though, data-mining techniques reveal importantlinks. For example, Marriott Vacation Club In-ternational reduced the volume of direct mail itneeded to reach target sales levels by correlatingresponse rates to specific vacation offerings andspecific customer characteristics.7

Data mining also offers enormous gains interms of performance, speed of use, and userfriendliness.8 While data miners must understandstatistical principles, highly specialized statisticalknowledge is not necessary to study, understand,and improve decision-making processes. Datamining helps managers to spot trends morequickly.

Because researchers may ignore the assump-tions and limitations of a theoretical model,traditional statistical analyses in customer-satisfaction research are often biased. Satisfactionresearch includes measures of the importance thatcustomers place on product and services at-tributes. Typically, these measures are highly cor-related, which can dramatically bias the statisti-cal values that determine attributes’ importancerankings. Also, statistical analyses usually assumethat relationships between independent and de-pendent variables are linear—which is often notthe case. Therefore, violation of these assump-tions can result in biased and misleading statisti-cal outcomes. Data-mining techniques (e.g., neu-ral networks) overcome these limitations andoutperform traditional statistical analyses in caseswhere such assumptions do not apply.9

Another considerable advantage over tradi-tional statistical models is data mining’s abilityto easily handle large and complex datasets.10

Data-mining techniques are not hamperedby large numbers of predictive variables, andthat feature makes data mining useful for select-ing variables, that is, identifying those withina set that are most relevant. The ability to handlelarge numbers of variables also makes datamining more realistic than statistical models inrepresenting the complexity of a typical businessenvironment.

While many analytical techniques can be clas-sified as data-mining tools, opinion has not coa-lesced regarding exactly which techniques shouldbe considered part of the data-mining tool kit.The tools listed in the accompanying sidebar (onthe next page) almost certainly belong, however.

5 M. Shaw, C. Subramaniam, G. Tan, and M. Welge,“Knowledge Management and Data Mining for Market-ing,” Decision Support Systems, May 2001, pp. 127–137.

6 C.S. Dev and M.D. Olsen, “Marketing Challengesfor the Next Decade,” Cornell Hotel and Restaurant Ad-ministration Quarterly, Vol. 41, No. 1 (February 2000),pp. 41–47.

7 Peacock, op. cit.

8 Le Bret, op. cit.

9 Ibid.

10 Peacock, op. cit.





Looking at that toolkit, decision trees, asso-ciation rules, case-based learning tools, neuralnetworks, and genetic algorithms are categorizedas machine-learning methods, while the otherscan be thought of as machine-assisted aids to sup-port human learning.11

With data-mining techniques, levels of a priorispecification can vary. In some cases, certain in-dependent variables and dependent variables maybe specified for examination, while predictor vari-ables in other cases may be uncovered only bythe data-mining tool. The point remains, though,that in comparison with traditional statisticalmethods, data-mining techniques invariably aremore data driven than they are user driven.

We have observed that some hotel corpora-tions are attempting to harness the power of in-formation by investing in data-mining technol-ogy that exploits consumer information. HiltonCorporation uses E.piphany E.4 software at itsBeverly Hills headquarters, for instance,12 andStarwood Corporation recently invested in UnicaCorp’s Affinium software.13 Such data-miningtechnology allows hotel corporations to predictconsumer-behavior trends, which are potentiallyuseful for marketing applications. For example,Starwood’s marketing staff can run reports andanalysis on customer and occupancy data storedin a data warehouse that combines customer andtransaction information from all company prop-erties. Such information indicates where custom-ers who visit a specific hotel live. If the data re-veal that the Sheraton Fisherman’s Wharf in SanFrancisco experiences a surge in visitors from FortLauderdale in April, for instance, hotel market-ers can increase promotional efforts in Fort Lau-derdale during the late winter months.14 Exhibit1 lists examples of how information gleaned fromdata mining can be used in a hotel corporation’smarketing activities.

A Data-mining Toolkit

• Association rules: Information from customer-purchase histories isused to formulate probabilistic rules for subsequent purchases.

• Case-based reasoning: Sets of attributes from new problems are com-pared with attribute sets from previously encountered problems (calledcases) to find one or more boilerplate examples that provided goodoutcomes or solutions.

• Decision trees: Automatically constructed from data, these yield asequence of step-wise rules; good for identifying important predictorvariables, non-linear relationships, and interactions among variables.

• Descriptive statistics: Averages, variation, counts, percentages, cross-tabs, simple correlation; used at the beginning of the data-miningprocess to depict structure and identify potential problems in data.

• Genetic algorithms: Use procedures modeled on evolutionary biology(e.g., selection, mutation, survival of the fittest) to solve predictionand classification problems or develop sets of decision rules.

• Neural networks: Applications that mimic the processes of the humanbrain; capable of learning from examples (large training sets of data)to discover patterns in data; can combine information from many pre-dictors and work well even with correlated variables, non-linearrelationships, and missing data.

• Query tools: Provide summary measures such as counts, totals,and averages.

• Regression-type models: Ordinary least-squares regression, logistic re-gression, discriminant analysis; used mostly for confirmation of modelsbuilt by “machine-learning” techniques.

• Visualization tools: Histograms, box plots, scatter diagrams; usefulfor condensing large amounts of data into a concise, comprehensiblepicture.—V.P.M., E.D.H., and S.K.H.

11 Ibid.

12 L. Stevens, “CRM Analytics—CRM by the Slice—Running Analytics Is Expensive, So Companies Are Focus-ing on Areas with Customers,” Internetweek, April 9, 2001,pp. 35–38.

13 G. Tischelle and J. Maselli, “Hotels Turn to IT to StemLosses,” Informationweek, December 17, 2001, pp. 31–32.

14 Ibid.





Harrah’s Data-mining Success StoryIn 1997 Harrah’s hotels and casinos introduceda trademarked loyalty-card program, “Total Re-wards,” which tracks customers’ purchasing ac-tivities and provides rewards that encouragespending at Harrah’s properties. Rather thanbuild glitzy properties with eye-popping attrac-tions, Harrah’s pursued a customer-service-oriented strategy centered around data-miningtechniques. Harrah’s used an information systemcalled WINet to link all its properties, allowingthe firm to collect and share customer informa-tion company wide. The process effectivelychanged the corporate culture from an every-property-for-itself mentality to a collaborative,customer-focused enterprise.15

The WINet system connects and consolidatescustomer information from all of the company’stransaction, slot-machine, hotel-management,and reservation systems. Key pieces of informa-tion—gender, age, place of residence, and typesof casino games played—help predict which cus-tomers are most likely to become frequent users.Based on this information, Harrah’s designs mar-keting strategies to retain those customers. Cus-tomers’ purchasing and gaming patterns aretracked, too, so that the company can target itscustomers with the most-appropriate incentives.For example, customers who reside outside thelocal area receive complimentary hotel rooms ortransportation, while drive-in customers receivefood, entertainment, or cash incentives.16

Data-mining techniques help to reveal datapatterns and relationships that can be used todevelop strong models for predicting the poten-tial value of each customer. Given that retaininga customer is less costly than attracting a newone, building strong relationships with valuedexisting customers can boost profits. Having in-formation regarding such things as the customer’sbirthday, anniversary, and favorite foods anddrinks allows a hotel to provide excellent, tailoredcustomer service that cements brand loyalty.Harrah’s discovered that the 30 percent of itscustomers who spent between $100 and $500

EXHIBIT 1

Examples of the uses of data-mining informationin hotel marketing • Create direct-mail campaigns

• Plan seasonal promotions

• Plan the timing and placement of ad campaigns

• Create personalized advertisements

• Define which market segments are growing most rapidly

• Determine the number of rooms to reserve for wholesale customersand business travelers.

per visit accounted for 80 percent of companyrevenues and generated nearly 100 percent ofprofits. In the first two years of its rewards pro-gram, Harrah’s saw a $100-million increase inrevenue from customers who visited more thanone property.17 Currently, Harrah’s ranks first inthe industry in profit growth.18

Because the WINet system can consistentlyidentify which customers will be most valuableover the long term, data mining is also useful fordetermining when to avoid offering incentivesto customers who are not lucrative. Harrah’s es-timates that it has saved some $20 million bywithdrawing incentives from customers who arenot likely to return.19

Despite Harrah’s success, some remain skep-tical of data mining’s customer benefits and long-term financial payoffs. As an example, SusanDobscha, co-author of “Preventing the Prema-ture Death of Relationship Marketing,” adviseshotels that giant central databases “are not wherecustomers want a relationship forged. A customerwould probably prefer a lower price over, say,having their beverage choice anticipated.”20

Another important caveat regarding data min-ing is that any relationship discovered must bevalid to benefit a company’s performance. WhenBritish Columbia Telecom tried to reward 100of its best customers by inviting them to a

15 M. Levinson, “Harrah’s Knows What You Did LastNight,” Darwin Magazine, May 2001.

16 J.A. Nickell, “Welcome to Harrah’s,” Business 2.0, April2002.

17 Ibid.

18 Levinson, op. cit.

19 Ibid.

20 “Mining Hotel Data,” Data Warehouse Report, October20, 1998.





Vancouver Grizzlies basketball game, for instance,it selected customers from the database compris-ing frequent 900-number users. After sendinginvitations to the printer, the marketing staff re-alized that those 900-number users included alarge number of sex-line enthusiasts. The com-pany avoided a serious gaffe by refining the cri-teria to create a list of truly loyal guests.21

Data-mining Applications for theHotel IndustryThe tasks performed by data mining can begrouped into the following five categories.

(1) Classification arranges customers intopre-defined segments that allow the sizeand structure of market groups to bemonitored. Also, predictive models canbe built to classify activities. An illustra-tion of such a model is one that predictswhich segment’s usage rate will experiencethe largest decrease when a particularpromotion expires. Classification usesthe information contained in sets of pre-dictor variables, such as demographicand lifestyle data, to assign customers tosegments.

(2) Clustering groups customers based ondomain knowledge and the database, butdoes not rely on predetermined groupdefinitions. This function is beneficialbecause it aids hoteliers in understand-ing who are their customers. For example,clustering may reveal a subgroup withina predetermined segment with homog-enous purchasing behavior (e.g., a sub-group of holiday shoppers within thetransient segment) that can be targetedeffectively through a specific ad campaign.(The idea is that the members of the sub-group will increase their number of staysor become more loyal.) On the otherhand, clustering may indicate that previ-ously determined segments are not parsi-monious and should be consolidated toincrease advertising efficiency. Informa-tion such as demographic characteristics,

lifestyle descriptors, and actual productpurchases are typically used in clustering.

(3) Deviation detection uncovers dataanomalies, such as a sudden increase inpurchases by a customer. Information ofthis type can prove useful if a hotel cor-poration wants to thank a guest for heror his recent increase in spending or offera promotion in appreciation. Marketingmanagers may also attempt to draw cor-relations between surges in deviationswith uncontrollable business-environmentfactors that are not represented in the da-tabase (e.g., a sharp increase in gasolineprices).

(4) Association entails the detection of con-nections between records, driven by as-sociation and sequence discovery. For ex-ample, a possible detected association maybe that a particular segment’s averagelength of stay increases after a specificadvertising campaign. Another associa-tion task could be employed in an effortto determine why a specific promotionwas successful in one market, but inef-fective elsewhere. Specific information re-garding customer-purchase histories isnecessary to formulate probabilistic rulespertaining to subsequent purchases.

(5) Forecasting predicts the future value ofcontinuous variables based on patternsand trends within the data. For instance,the forecasting function can be used topredict the future size of market segments.With forecasting one can also use datatrends to project which hotel amenitiesare of growing importance to consumersand will be key drivers of the consumer’sfuture perception of value.

In the hotel industry, the most commonsources of data are CRSs and PMSs. Some hotelcorporations also use information that resides inguest-loyalty-program databases. Hilton, for in-stance, analyzes data contained within its trade-marked Hilton HHonors database.22 Another

21 S. Press, “Fool’s Gold?,” Sales and Marketing Manage-ment, June 1998, pp. 58–61.

22 L. Stevens, “IT Sharpens Data-mining Focus—Insteadof Building Data-mining Applications with No Clear Goal,Companies Are Setting Priorities Up Front to MaximizeROI,” Internetweek, August 6, 2001, pp. 29–30.





Data mining can be a powerful andvaluable marketing tool. However,simply investing in data-mining tech-nology may not guarantee success.

potentially important source of data is the infor-mation provided by guest-satisfaction surveys.

Guidelines for Effective Data MiningWhen properly employed, data mining is a pow-erful and valuable marketing tool. However, sim-ply investing in data-mining technology may notguarantee success. As presented below, sevenguidelines influence the effective management ofdata-mining technology.

Guideline #1: Match your IT priorities withan appropriate provider. There is high demand forand low supply of data-mining expertise as morecompanies realize the potential value of the in-formation residing within their databases. Tocapitalize on this demand, a number of second-tier research firms now provide data-mining ser-vices.23 However, providers offer a wide range ofskill levels. The most-skilled providers can turndata into useful information. Companies that ini-tially set clear priorities have a greater chance ofreaping maximum benefits from data-miningprojects than do firms that are unsure of theirgoals.24 Clear priorities include goals about whatthe firm would like to achieve through data min-ing and when it will be achieved. Without goalsand objectives the hotel corporation is uncertainabout what it is shopping for when seeking a dataminer. It is also important to communicate thesegoals to prospective providers. When selecting aprovider, ask the following six questions:

• Does the provider have experience setting uppredictive models with marketing applica-tions? Data mining has applications otherthan marketing. Data mining’s ability todetect patterns in data is used extensivelyin criminal justice and anti-terrorism ef-forts to anticipate illegal activity, for in-stance. Wall Street also employs data min-ing to predict moves in the financialmarkets. Large global corporations usedata mining to gain efficiencies in pur-chasing and production throughout theirnetworks. Therefore, it is not enough tohave a data-mining consultant, but onemust find a provider that has experience

in marketing. Building models to predictconsumer behavior is a form of data min-ing that requires specific expertise. Forexample, a data miner with marketing-applications experience would know toreplace a zip code with resident character-istics, such as median income.25

• Does the provider have experience in creat-ing models within the hospitality industry?Marketing applications of data mining areemployed across diverse industries. Build-ing predictive models for a grocery store,a furniture chain, an airline, or a hotel isdifferent in each case. It is beneficial tofind a provider that has experience in set-ting up models in the hotel industry. Sucha provider would more clearly understandhotel-guest-segmentation processes, forexample.

• Is the provider reputable? Because manysecond-tier companies provide miningservices, it is important to check the cre-dentials and reputation of the vendor.

• Does the provider offer the latest technologythat is appropriate? Because of the widerange of products available, it pays to doyour homework. It is crucial to invest inthe latest appropriate technology becauseit is extremely expensive and time con-suming to switch products after one isinstalled—in no small part becauseswitching products requires retrainingthe IT and marketing staff.

• Does the provider offer a product that hasvisual-exploration capabilities? Cutting-edge data-mining software has visual-exploration capabilities, which means

23 M. Brandel, “Spinning Data into Gold,” Computerworld,May 26, 2001, pp. 67–70.

24 Stevens, op. cit. 25 Brandel, op. cit.





that data patterns can be viewed as three-dimensional objects that can be rotated orzoomed for detailed analysis. In addition,pixel-oriented technology assigns colors todata values so that patterns and trends canbe examined. Visual exploration is an im-mense aid to managers and marketers be-cause it often serves as a preliminary toolin selecting the appropriate variables fordata-mining tasks.26

• Is the provider willing to provide a customcontract? Contract negotiations are a criti-cal step in initiating a successful data-mining program. The contract should beas precise as possible and should abstainfrom nebulous clauses discussing partner-ship.27 Moreover, the vendor’s standard

contract should not be used, because thestandard contract does not customarilyinclude specific performance standards orpenalty clauses if the vendor falls short ofrequirements. Worse, payment schedulesin standard contracts may favor the ven-dor.28 A custom contract should be writ-ten to include service-level measures anda termination clause. The buyer shouldbe particularly suspicious of so-calledchange-of-character clauses, which statethat the buyer may have to pay for anychanges in “functionality” throughout thelife of the contract. Change-of-characterclauses have caused many disputes becauseof the ambiguous nature of the term“functionality.”29

Guideline #2: Build segmentation and predic-tive models. Building appropriate segmentationand predictive models necessitates an extensiveknowledge of the hotel business. Exhibit 2 pro-vides examples of some of the many ways thathotel guests can be segmented.30 Transient ho-tels, convention hotels, extended-stay hotels, andresort properties all segment guests differently.Furthermore, guest segmentation is distinctivefor most hotel properties. Hilton’s and Marriott’sproperty-management systems segment and codemarkets at the property level, for instance, sinceeach location has its own particular segments. Agiven property may serve a set of corporate cli-ents, a group of government clients, and socialclients (e.g., weddings and reunions). The seg-ment categories contained in Exhibit 2 can bestrung into a large set of combinations. Further-more, a guest could potentially fit into severalcategories, which poses a challenge for currentdata-mining techniques.31 As a consequence,finding a provider that has experience creatingmodels in the hotel industry is a major benefit.Additionally, even if the provider has hotel expe-rience, it is critical that IT and marketing man-

EXHIBIT 2

Examples of hotel-guest segmentsGeographic

NationsStatesCountiesCities

Demographic

Age or life-cycle stageGenderIncome

Psychographic

Social classLife-stylePersonalityBehaviorOccasion of purchase decisionOccasion of useBenefits soughtUser status (e.g., potential, former, first time)Usage rateLoyalty statusBuyer-readiness stage

Source: P. Kotler, J. Bowen, and J. Makens, Marketing for Hospitality and Tourism,second edition (Upper Saddle River, NJ: Prentice-Hall, 1999).

26 M. Shaw et al., op. cit.

27 P. Lacity and R. Hirschheim, Beyond the InformationOutsourcing Bandwagon (New York: John Wiley & Sons,1995).

28 Ibid.

29 Ibid.

30 See: P. Kotler, J. Bowen, and J. Makens, Marketing forHospitality and Tourism, second edition (Upper Saddle River,NJ: Prentice-Hall, 1999).






Finding a data-base expert who has

experience creating models in the

hotel industry is a major benefit.

agers work closely with the provider to segmentthe market and build predictive data-miningmodels.

Once a data-mining model is built, confirma-tory testing must be conducted to assess its pre-dictive accuracy. For instance, a model designedto predict who will respond to a promotionshould be based on a prior offering in which it isknown who did or did not respond. After themodel is constructed, a “holdout” group from aprevious promotion can be analyzed to verify re-liability. If the holdout predictions do not repli-cate the results of the past promotion, then themodel may not be significantly predictive. Tofurther enhance accuracy, a score can be assignedto the model based on the level of agreementbetween the holdout group and the entire group.Subsequent refined models can then be tested andscored. Another standard approach to model vali-dation involves drawing two random samplesfrom the data. The first sample is used as a cali-bration sample to build the model, while the sec-ond is used as a holdout sample to evaluate themodel built from the calibration sample.32 Thevalidation process requires a knowledgeable ITprofessional, because when data subtleties thatarise only in the sample are used to build themodel, the model may be highly predictive ofthe sample but biased with regard to the popula-tion.33 This is called overfitting the data. To avoidcreating a biased model, the IT professional mustbe knowledgeable of the analytical procedure andpossess a basic understanding of the hotel seg-ment and promotional scenario from which thesample was extracted.

Guideline #3: Collect data to support the mod-els. Accurate data collection is critical for success-ful data mining. The major obstacle to effectivedata mining, however, is inadequate data gather-ing and input.34 Data problems lead to a decreasein the value of any data warehouse, in additionto diminishing the value of proposed models.35

Problems with data are related to one or more ofat least three different shortcomings.

The first possible difficulty involves miss-ing or inaccurate data. For example, when oc-cupation information is available for only 15percent of a data set, it is difficult to create aprofile of customer occupations. Then again,it’s a problem if the data file contains occupa-tion information for 90 percent of the popu-lation, but the accuracy of the information is

poor. Hotel corporations can reduce inaccu-racy of this kind by asking guests for their cur-rent occupation.

A second obstacle is poorly coded data. Da-tabases must have standards regarding data for-mats, text case, and redundant codes.36 Al-though some software automatically formatsthe data properly, most do not. Problems thenoccur when data-input sources are added overan extended time and no one has ensured thatthe data entering the warehouse is properlyformatted. This would occur, for instance, if,when original data-mining technology was in-stalled, predictions were made based on the res-ervations system and the property-managementsystem, but then a subsequent decisionwas made to input data from guest-satisfactionsurveys. Problems would transpire when addi-tional data inputs are not standard or are codedimproperly. For example, some models re-quire continuous and ordinal data, while oth-ers demand categorical data fields or binaryconstructs.37

A third potential problem involves using hom-onyms (that is, putting the same label on two ormore different data elements) and synonyms (thatis, using two different labels for the same data

32 P.R. Peacock, “Data Mining in Marketing: Part 2,”Marketing Management, Spring 1998, pp. 15–25.


34 M. Smith, “Refining Raw Data,” Printing Impressions,Vol. 43, No. 9 (2001), pp. 36–37.


36 L. Stevens, op. cit.

37 Siragusa, op. cit.





Training is a key to the effectiveimplementation of data-miningsystems.

element).38 While it may seem tautological to ad-vise precluding this occurrence, the most commonculprit is a new user on the system. It is commonfor hotel and IT professionals to change companiesfrom time to time. Turnover causes coding prob-lems when new employees bring their old labelsand fail to use their new employer’s framework.

Guideline #4: Select the appropriate tools foranalysis and prediction. Numerous analytical toolscan be employed to transform data into usefulinformation. Some of the less-common analyti-

cal tools used by data-mining software includeregression models, factor analysis, cluster analy-sis, structural equation modeling, and self-organizing maps. On the other hand, the mostcommon statistical methods used in data-mining applications are decision trees, neuralnetworks, and genetic algorithms. As previouslymentioned, a decision tree is a rule-based modelconstructed of nodes (decision points) andbranches (connections between nodes) that reachnumerous outcomes based on traveling throughtwo or more nodes. A neural network is a non-linear predictive model that resembles a biologi-cal neural system and has the ability to learnthrough training. Last, a genetic algorithm is alearning-based model founded on the conceptof evolution. That is, partial solutions to a sce-nario compete with each other, and then the bestsolutions are used for further problem solving.39

Most of the statistical methods employ tech-niques that achieve a desired outcome. Likewise,each methodology has strengths and weaknesses,

and each is appropriate for a specific scenario.Therefore, the most effective results emanatefrom data miners who have the expertise to se-lect the most appropriate statistical method for agiven scenario and the hotel’s intended goals.40

For instance, a positive attribute of genetic algo-rithms is that they converge on an optimal solu-tion, but the method is most applicable to largedatabases since arriving at a valid outcome mayrequire many generations of competing solutions.Likewise, there are also pros and cons associatedwith neural networks. They are beneficial in ana-lyzing complex data because of their ability todiscover unusual trends, but monitoring accu-racy is difficult because many intricate relation-ships are handled invisibly by the methodology.41

Guideline #5: Demand timely output. Timeli-ness is critical in making marketing decisions.The length of time required to produce outputvaries widely among data-mining packages.Before Hilton Corporation upgraded its data-mining technology, for instance, the reports thatmanagers requested from IT would take three tosix weeks to arrive. “By the time they’d get thereport, it was often too late to act on it,” saidJoanne Flinn, vice president of leisure market-ing. With the new technology, managers receivereports in 30 minutes or less.42

Guideline #6: Refine the process. By its nature,data mining involves knowledge that evolves overtime. Never complete, data mining involves acontinuous cycle of inputs and outputs based onmodels that must be modified and refined asconditions change in the competitive environ-ment. Flexibility is needed to adapt the estab-lished models and processes to changes that oc-cur.43 Refinement consists of three actions:

(1) Chart progress toward initial goals. Usethe forecasting function of data miningto regularly set new goals.

(2) Compare and contrast the characteris-tics of the clustering output with theattributes of the classification output.When necessary, modify predictive

38 J. Chopoorian, R. Witherell, O. Khalil, M. Ahmed,“Mind Your Business by Mining Your Data,” S.A.M.Advanced Management Journal, Vol. 66, No. 2 (2001),pp. 45–51.

39 J. Hair, R. Anderson, R. Tatham, and W. Black, Multi-variate Data Analysis, Fifth Edition (Upper Saddle River,NJ: Prentice-Hall, 1998).

40 Siragusa, op.cit.

41 J. Hair et al., op. cit.

42 Stevens, op. cit.

43 R. Cline, op. cit.





Data-mining technology should beused in conjunction with other fore-casting and research techniques.

models based on changes in the size orstructure of customers’ market seg-ments. For instance, notable variancesin purchase patterns in one segmentand similarities in purchase patternsamong other segments may lead to re-finement of the segment, usually byadding a new criterion or dimension.This action can result in the discoveryof previously undetected segments withhomogenous characteristics.

(3) In addition to clustering and classifica-tion features of data mining, also usethe associations and deviation-detectionfunctions to assess the effects of promo-tions. Maintain a promotional historytable in the database to use as a learningtool for future campaigns and models.44

Guideline #7: Hire a well-trained staff and aknowledgeable IT manager. Information technol-ogy was initially viewed by the hotel industry asa back-office function that supports the financeand accounting areas.45 The industry has ad-vanced far beyond this view during the past de-cade. In two sessions sponsored by the Interna-tional Hotel and Restaurant Association(IH&RA), one in Singapore in 1997 and the sec-ond in Nice, France, in 1998, hotel-industry lead-ers pondered the role of technology. Among theconclusions reached were: “Going forward, tech-nology will be the most competitive weapon forany hospitality company. If hospitality organiza-tions want to compete successfully, they must doso by using technology to drive value to both thecustomer and to the firm.”46 However, imple-menting such recommendations at the propertylevel can be a difficult task.

Training is a key to effective implementationof data-mining systems. Productive data miningrequires two-fold proficiency among both ITmanagers and those who interpret the outputs.

The hotel’s IT managers must also be profi-cient with the data-mining system because thesystem requires continuous refinement. Just as

market segments, sources of data, and propertygoals change, so must predictive models andanalyses be modified and refined. It is an unsoundpolicy for the IT staff to be totally dependent onthe provider’s recommendations for refinementand alterations. Instead, the IT staff and data-mining provider should work together with theircommon goal being to maximize the technology’seffectiveness. The most-effective data-miningprojects occur when IT managers and providerscollaborate and share project information.

Second, adequate training must be pro-vided to all potential users of data-mining out-puts. At the corporate level this includes themarketing staff, operations managers, and thosedeveloping new properties. Users at the propertylevel include general managers, directors of salesand marketing, and the sales staff. Users mustbe instructed about the available reports andhow to properly interpret the information. Sincethe information is used for decision making, itis important for users to understand the bound-aries and limitations of the information.

Boundaries and LimitationsTechnology must serve managers’ purposes,rather than dictate processes.47 Along that line,data mining cannot capture all the informationrelating to what drives consumer behavior. Datamining is simply one of a number of researchmethods that help predict travelers’ demandtrends. Therefore, data-mining technologyshould be used in conjunction with other fore-casting and research techniques. With this inmind, managers should be aware of the follow-ing four limitations of data-mining technology.

44 Siragusa, op. cit.

45 Cline, op. cit.

46 M. Olsen and D. Connolly, “Antecedents of Technologi-cal Change in the Hospitality Industry,” Tourism Analysis,Vol. 4 (1999), p. 29.

47 C. Chudnow, “Knowledge-management Tools,” ComputerTechnology Review, Vol. 21, No. 11 (2001), pp. 28–29.





Limitation #1: Data mining analyzes only datacollected from existing customers. Data-mining soft-ware generates information by analyzing datapatterns derived from the company’s reservation,property-management, and guest-loyalty-program systems. Patterns thus detected canhelp predict the actions of current guests in thesystem and of those with similar needs andwants. Data-mining technology does not, how-ever, provide information about market segmentsnot found in the company’s databases. Moreover,a market segment that is currently small but ison the verge of experiencing substantial growthmay not be detected by data mining.

Another blind spot is the data in competitors’reservation systems. A key question in planninga marketing strategy in the hotel industry is: Whoare my competitors’ guests and where are theycoming from? Data-mining technology is unableto answer those questions.

Limitation #2: Databases used in the miningprocess are often hotel-brand specific. Just as datamining cannot analyze competitors’ markets, italso creates prediction models that are brand spe-cific. Thus, corporations that operate multiplebrands often must create a data warehouse and con-duct data mining for each brand. This is also truefor the franchisees that may have a portfoliocomprising, say, six Holiday Inns and four Marriotts.

Brand-specific marketing information is use-ful for the brand’s corporate office to plan mar-keting programs, which is largely what franchi-sees purchase. Conversely, brand-specificmarketing information may not be helpful if thehotel corporation that franchises numerousbrands wants to predict customer demand basedon a multiple-brand portfolio.

Limitation #3: Data mining may not segmenttravelers by psychographic traits. Segmenting con-sumers based on psychographic traits, such aspersonality and lifestyle, can be useful in the hotelindustry. This is because psychology and emo-tion play significant roles in the hotel guest’s de-cision process. That is, as seen in Exhibit 3, atraveler may select a destination for a variety ofpsychological reasons.48 One limitation of datamining is that common system inputs do notaccount for psychological factors that influencea traveler’s purchase decision.

A time-tested tool used in understanding hos-pitality demand trends is Stanley Plog’s psycho-graphic scale.49 Many key drivers of demandidentified by Plog, such as personality distribu-tion among travelers (e.g., dependables, ventur-ers, and centrics), are not common inputs intodata-mining systems. Hotels can acquire thisinformation from customer surveys.

Limitation #4: Data mining does not provideinformation about consumers’ thought processes.Itis important to engage consumers in research tobetter understand their thinking. Informationgenerated by data mining does not account forthe fact that approximately 80 percent of hu-man communication is nonverbal.50 Interviewsand focus groups are both useful methods forgathering information about the needs and wantsof hotel guests. The insight gained from thosetechniques is difficult to capture in the statisti-cal data-mining outputs. That is why it is im-portant to step back and ask what the hotelguest’s inherent needs are and what the productis really about. This involves conducting in-depthconversations with guests. At times, improvedinsight and perspective are gained from talkingwith three customers for two hours rather thanby surveying a thousand customers.51

EXHIBIT 3

Psychological determinants of demandEducation Relaxation

Escape Self-discovery

Family bonding Sexual opportunity

Prestige Social interaction

Source: P. Kotler, J. Bowen, and J. Makens, Marketing for Hospitality and Tourism,second edition (Upper Saddle River, NJ: Prentice-Hall, 1999).

48 Kotler et al., op cit.

49 S. Plog, “Why Destination Areas Rise and Fall in Popu-larity,” Cornell Hotel and Restaurant Administration Quar-terly, Vol. 42, No. 3 (June 2001), pp. 13–24.

50 G. Zaltman, “Rethinking Market Research: PuttingPeople Back In,” Journal of Marketing Research, Vol. 34, No.4 (1997), pp. 424–437.

51 K. Ohmae, The Borderless World (New York: McKinsey& Company, Inc., 1999).





Conclusion and ManagerialImplicationsData-mining technology can be a useful tool forhotel corporations that want to understand andpredict guest behavior. Based on informationderived from data mining, hotels can make well-informed marketing decisions—including whoshould be contacted, to whom to offer incen-tives (or not), and what type of relationship toestablish.

Data mining is currently used by a number ofindustries, including hotels, restaurants, automanufacturers, movie-rental chains, and coffeepurveyors. Firms adopt data mining to under-stand the data captured by scanner terminals,customer-survey responses, reservation records,and property-management transactions. Thisinformation can be melded into a single data setthat is mined for nuggets of information by data-mining experts who are familiar with the hotelindustry.

However, data mining is no guarantee of mar-keting success. Hotels must first ensure that ex-isting data are managed—and that requires in-vestments in hardware and software systems,

data-mining programs, communications equip-ment, and skilled personnel. Affiliated proper-ties must also understand that data mining canincrease business and profits for the entire com-pany and should not be viewed as a threat to onelocation. As seen in the Harrah’s example, imple-menting a data-mining system is a complex andtime-consuming process.

We advise hospitality managers to adopt adata-mining system and strategy if they have notdone so. Guidelines presented in this paper—including how to select and manage the data-mining provider—offer guidance for implement-ing a viable data-mining strategy. Since datamining is in its initial stages in the hotel indus-try, early adopters may be able to secure a fasterreturn on investment than will property manag-ers who lag in their decisions. Hotel corporationsmust also share data among properties anddivisions to gain a richer and broader knowledgeof the current customer base. Managementmust ensure that hotel employees use the data-management system to interact with customerseven though it is more time consuming than atransactional approach. �

Vincent P. Magnini (photoat far left) is a Ph.D. candi-date at Old Dominion Uni-versity ([email protected]).Earl D. Honeycutt, Jr.,

Ph.D. (middle photo), is aprofessor at Love School ofBusiness at Elon University([email protected]),where Sharon K. Hodge,

Ph.D. (photo at immediateleft), is an assistant profes-sor ([email protected]).



Documents

Vincent P. Magnini, Earl D. Honeycutt, Jr. and - Sage Publications