Enabling e-transactions with multi-attribute preference...

Available online at www.sciencedirect.com

European Journal of Operational Research 186 (2008) 748–765

www.elsevier.com/locate/ejor

Decision Support

Enabling e-transactions with multi-attribute preference models

John C. Butler a, James S. Dyer b,*, Jianmin Jia c, Kerem Tomak d

a Department of Accounting and MIS, Fisher College of Business, The Ohio State University,

Columbus, OH 43210, United Statesb Department of Management Science and Information Systems, The Graduate School of Business,

University of Texas at Austin, Austin, TX 78712, United Statesc Department of Marketing, Faculty of Business Administration, The Chinese University of Hong Kong, Shatin, NT, Hong Kong

d Yahoo! Inc., Search and Marketplace Division, 2811 Mission College, Santa Clara, CA 95054, United States

Received 10 May 2006; accepted 30 January 2007Available online 21 March 2007

Abstract

This paper describes potential applications of multi-attribute preference models (MAPM) in e-commerce and offerssome guidelines for their implementation. MAPM are methodologies for modeling complex preferences that depend onmore than one attribute or criterion, and include multi-attribute utility theory, conjoint analysis, and the Analytic Hier-archy Process. There are numerous examples of applications in e-commerce that would benefit from the acquisition ofinformation regarding the preferences of a consumer, a customer, an advice seeker, or a decision maker. Here, the focusis on applications of MAPM models in B2C and B2B websites, where preferences of consumers are assessed for the pur-pose of identifying products or services that closely match their needs.

In this paper, we provide an overview of decision aids with the MAPM approach, emphasizing how the MAPM struc-ture of an individual’s preferences may be assessed. This discussion is illustrated with examples of the use of alternativeMAPM assessment approaches that are incorporated in existing websites. We then discuss how MAPM applicationsshould be tailored for success in these environments.� 2007 Elsevier B.V. All rights reserved.

Keywords: Multi-criteria analysis; Decision analysis

1. Introduction

The purpose of this paper is to describe thepotential applications of multi-attribute preferencemodels (MAPM) in e-commerce and to offer someguidelines for their implementation. MAPM are

0377-2217/$ - see front matter � 2007 Elsevier B.V. All rights reserved

doi:10.1016/j.ejor.2007.01.051

* Corresponding author.E-mail address: [email protected] (J.S. Dyer).

methodologies for modeling complex preferencesthat depend on more than one attribute or criterion,and include multi-attribute utility theory (Keeneyand Raiffa, 1976; Dyer and Sarin, 1979), conjointanalysis (Green et al., 2001), and the Analytic Hier-archy Process (Saaty, 1980; Forman and Gass,2001). In this paper we focus on the approaches thatfeature additive, compensatory value functions butmany of the ideas apply to outranking approacheslike ELECTRE (e.g., Roy, 1996) as discussed in

.

mailto:[email protected]

J.C. Butler et al. / European Journal of Operational Research 186 (2008) 748–765 749

Stewart and Losa (2003). There are numerousexamples of applications in e-commerce that wouldbenefit from the acquisition of information regard-ing the preferences of a consumer, a customer, anadvice seeker, or a decision maker.

A cottage industry has emerged where consumersare given ‘‘free’’ on-line decision aids to supportdecisions ranging from voting in the last presidentialelection in the United States to selecting a pet. Themain objective of some of these business-to-con-sumer (B2C) tools is to collect valuable consumerdata from real decision makers. Many of these toolsattempt to emulate MAPM, but often do so in an adhoc manner and with varying degrees of success. Webelieve that the informed use of MAPM would leadto decision support systems that make more valu-able contributions and recommendations, leadingto more site visits from users, and potentiallygreater profits. We also believe that these decisionsupport tools can provide additional benefit asadd-ons to existing web-based service providers,especially in the business-to-business (B2B) domain.

Like Edwards and Fasolo (2001) we start with thepremise that the proper use of MAPM is a normativeconcept. We further assume that website developerstake a prescriptive view and attempt to help consum-ers use an MAPM to accurately capture their prefer-ences. This assumption is reasonable for websiteswhose primary value proposition is unbiased advice.However, we recognize that some of these servicesmay be biased towards a particular vendor or prod-uct, and also discuss ways in which consumer prefer-ences could be manipulated via a web-based system.

In this paper, we will provide an overview ofdecision aids that offer MAPM-like support,emphasizing how the MAPM structure of an indi-vidual’s preferences may be assessed. This discus-sion will be illustrated with examples of the use ofalternative MAPM assessment approaches that areincorporated in existing e-enabled websites. We willthen discuss how MAPM applications should be tai-lored for success in these environments.

2. Web-based decision aids

As the amount of data that needs to be processedthrough a web interface increases, companies relyon more automated decision aids. One way of auto-mating the decision making task is to add ‘‘intelli-gence’’ to the decision support tools used foranalyzing the information generated in B2C andB2B electronic commerce applications.

2.1. Decision aids, intelligent agents, and shopbots

The topic of computerized decision support sys-tems has been the subject of a great deal of work forseveral decades now, although the focus of these earlysystems was mainly on decision aids for managersand executives. An early example of a ‘‘man–machineinteractive’’ decision support tool for selectingamong alternatives based on trade-offs among multi-ple attributes is provided by Dyer (1973).

However, the past decade has seen an explosivegrowth in the use of electronic commerce, primarilythrough the World Wide Web, and has expandedthe number of potential users of these decision sup-port systems to include both personal shoppers andprofessional purchasing agents. Within this new envi-ronment, there is an obvious motivation for market-ers to enhance the shopping experience for consumersin order to increase sales and profitability.

As a result, decision aids have been created for anumber of websites to support these on-line shop-pers and professional buyers. These decision aidsmay be classified as examples of intelligent agents(West et al., 1999), or more specifically as recom-mendation systems or shopbots. Typically they assistbuyers in searching a website or across several web-sites for products and services that offer low costsand meet other desired performance objectives.

2.2. Approaches to modeling preferences

There are many different approaches to modelingpreferences, and several different strategies havebeen incorporated into e-commerce sites. One ofthe most common approaches to obtaining someknowledge about a consumer’s preferences is toinfer that information based on her past purchasehistory. The objective of the analysis of this datais to use collaborative filtering to identify sub-groups among the population of consumers withsimilar purchasing patterns (and implicitly, withsimilar preferences). Based on this classification,the consumer can be presented with alternatives thatare commonly purchased by others within her sub-group. Perhaps the most well known example of thistype of a decision aid for consumers is the serviceprovided by www.Amazon.com.

The advantage of collaborative filtering is that itis a relatively painless way for the consumer to beprovided with alternatives that may be consistentwith her preferences. On the other hand, this meth-odology does not create an explicit model of prefer-

http://www.autobytel.com

750 J.C. Butler et al. / European Journal of Operational Research 186 (2008) 748–765

ences that can be modified by the consumer. As aresult, collaborative filtering may not be useful inthe evaluation of new or novel products and serviceswith no purchase histories.

Also, a preference model based on collaborativefiltering is not actively developed in cooperationbetween the consumer and the decision aid. In somecases, this collaboration between the consumer andthe agent in the construction of a representation ofpreferences may increase the consumer’s commit-ment to the process, and her ultimate satisfactionwith the choice that is made (Kahn and Huffman,1998).

In contrast, recommendation agents or shopbotsthat seek specific information regarding the con-sumer’s preference use content filtering. This infor-mation is obtained by identifying the attributesthat are relevant for a choice and assessing theimportance of these attributes. One possibility is touse collaborative filtering to identify attributes andobjectives that the consumer may care about as afirst step. This would be particularly useful for aninexperienced consumer. Next, some informationmay be obtained regarding how varying the levelsof performance on individual attributes translatesinto satisfaction with the product or service. Theresult is a multi-attribute preference model, orMAPM, that will be the focus of this paper.

2.3. Considerations of context

The efficacy of the use of MAPM as a Web-baseddecision aid will depend on several aspects of the con-text of the decision task. These aspects include theconsumer’s familiarity and level of expertise regard-ing the product or service, and the significance ofthe decision in terms of cost and commitment.

In order to construct and utilize a recommenda-tion agent featuring the MAPM approach, the usermust be able to express preferences for different lev-els of performance on attributes both within thesame attribute and across attributes. Individualswith low levels of knowledge regarding a productclass may have a difficult time making the necessarytrade-offs in terms of product attributes, and there-fore be very uncomfortable with both the processand the results of the ranking that is offered. Atanother extreme, an expert may have prior knowl-edge regarding available products and services,and prefer to search quickly for detailed informa-tion on a small number of alternatives without thedistraction of responding to preference assessment

questions. For similar reasons, consumer purchasesthat are considered routine, or that require rela-tively low commitments of resources, may not beappropriate environments for the intrusion of adecision aid.

Therefore, decision aids based on MAPMapproaches to approximating consumer preferencesmay be more appropriate for a relatively inexperi-enced B2C consumer who at least has sufficientknowledge to describe desired levels of performanceon product attributes, and for infrequently pur-chased durable products such as cameras, laptopcomputers, and automobiles. MAPM approachesmay be even more useful in B2B purchasing envi-ronments, where the buyer’s specification of prefer-ences in terms of attributes, weights of importance,and desirable performance levels may provide usefulinformation to potential sellers, and vice versa. Theparticipants in a B2B relationship may also berequired to justify their choices which is consistentwith having a well specified MAPM.

As a result, some websites offer the consumer thechoice of two or more of these types of decisionaids. These sites allow her to select the decision sup-port that is most appropriate for the decisioncontext.

3. A framework of MAPM

For the types of problems we will consider in thispaper, for example the selection of the best responseto a request for quote (RFQ) by a buyer at a B2Bsite, there are numerous measures of the desirabilityof the alternatives. Often these measures are conflict-ing and require trade-offs by the decision maker; thecheapest quotes may be for lower quality products.Under ideal circumstances, a MAPM approach willprovide a faithful representation of the decisionmaker’s preferences, and can be used in an auto-mated decision support system to simplify the tasksof identifying and comparing desirable alternatives.

Multi-attribute utility theory provides the theo-retical basis for a MAPM, and so we will begin withan informal review of the basic concepts of this the-ory as it is commonly applied in the e-commerceenvironment. As we shall discuss, e-commerceapplications of conjoint analysis and of the AHPmay be viewed as alternate approaches to the assess-ment of multi-attribute utility models.

A multi-attribute utility model of preference isdeveloped by defining attribute utility or value func-tions on performance measures of the alternatives,


and by aggregating performance on multiple criteriainto a single score. As far as we are aware, all of theweb-based applications of MAPM use the simplestaggregation rule, the additive multi-attribute model.Further, these applications also explicitly or implic-itly treat the consumer choice decision as though itoccurs in the context of certain outcomes or attri-bute values, that is, these applications typically donot provide information on attribute levels definedby probability distributions. Formally, there are dis-tinctions among the assumptions that justify the useof the additive multi-attribute model of preferenceand among the appropriate assessment methodolo-gies for the domains of certain versus risky choices.This distinction is often emphasized in the literatureby using the term value function for the case of cer-tainty and utility function for the case of risk, andwe will adopt this convention for emphasis.

Let vi(xi) be a single-attribute value function overthe attribute measure xi. This single-attribute valuefunction provides a functional relationship betweenthe observed attribute levels of performance andtheir perceived values or utilities. The simple addi-tive multi-attribute value function may be written as

vðx1; x2; . . . ; xnÞ ¼Xn

i¼1

viðxiÞ: ð1Þ

The additive model provides an accurate representa-tion of the preferences of individuals only if theirtrade-offs among any two of these criteria are notaffected by common outcomes on the remainingn � 2 criteria (Keeney and Raiffa, 1976). This addi-tive preference independence condition is an impor-tant one, and may not be recognized by designers ofweb-based decision aids.

Further, this additive value function provides abasis for ranking alternatives, but it does not mea-sure preference differences and it should not be usedfor risky choices. More formally, (2) is an additiveconjoint preference structure (Krantz et al., 1971).

To simplify the assessment of the additive prefer-ence model, an even stronger independence condi-tion is required, known as difference independence(Dyer and Sarin, 1979). Loosely speaking, this con-dition requires that common performance levels onthe other attributes should not affect the individual’spreferences for increases or decreases in the perfor-mance levels on any one attribute. This conditionallows each individual single-attribute value func-tion vi(xi) to be assessed without explicitly consider-ing the values of the other attributes, a common

practice in many e-commerce and traditional appli-cations. When this condition is met, the additivevalue function is also called a measurable valuefunction, and the scores of a measurable value func-tion provide the basis for comparisons of preferencedifferences among alternatives rather than simplerankings of the alternatives.

This assessment is often simplified by scalingeach single-attribute value function vi(xi) from 0 to1, and by adjusting for the scaling of these normal-ized value functions using weights wi for each attri-bute. With this choice of scaling, the additive multi-attribute value function can be written as

vðx1; x2; . . . ; xnÞ ¼Xn

i¼1

wi�viðxiÞ; ð2Þ

where �við�Þ is a single-attribute value function overattribute i scaled from 0 to 1, wi > 0 is the weightfor measure i, and

Pni¼1wi ¼ 1.

This scaling of the single-attribute value func-tions can be accomplished by assigning a score of0 to the worst level of the attribute x0

i and a scoreof 1 to the best level x�i . Then, other levels of perfor-mance between these two extremes might beassigned scores by direct rating, by ‘‘part worthscores’’ from conjoint analysis, or using otherassessment approaches. As an alternative, a simplemathematical relationship may be assumed orassessed (Farquhar and Keller, 1989; Kirkwood,1996).

There are several approaches for assessingweights for the additive multi-attribute value model,but they all should be consistent with the resultsobtained from trade-off judgments (Keeney andRaiffa, 1976). To illustrate the idea, we consider atrade-off between price and resolution for buying adigital camera, assuming that the performance lev-els of the other features are the same. Specifically,for two digital camera alternatives A = (€200,1600 · 1200 pixels) and B = (___, 1920 · 1600 pix-els), the decision maker is asked to specify a pricefor B such that the two alternatives are equallyattractive.

Suppose the decision maker would be indifferentbetween A and B if the price of B were set at €300.From (2), we have

w1�v1ð€200Þ þ w2�v2ð1600� 1200 pixelsÞ¼ w1�v1ð€300Þ þ w2�v2ð1920� 1600 pixelsÞ

where �v1 and �v2 are value functions for the price andresolution attributes respectively, and w1 and w2 are


corresponding weights. We may conduct similartrade-offs for other pairs of attributes to obtain sev-eral equations like the above. If all the single-attri-bute value functions have been determined, theseindifference equations can be solved for the weights.

As the name implies, conjoint analysis was devel-oped based on the use of statistical techniques toapproximate the additive conjoint preference struc-ture (2) (Luce and Tukey, 1964). A typical proce-dure in many applications of conjoint analysis isto assess the rank-order or overall value for alterna-tives with different profiles of attribute levels, andthen to use the holistic judgment information toestimate discrete levels of single-attribute valuefunctions by regressions, hierarchical Bayes models,or linear programming. When used in this manner,conjoint analysis is a decomposition method, andit does not need to evoke weights through directassessment. This holistic assessment approach istypically used with hundreds or even thousands ofconsumers in consumer preference research regard-ing product designs and service offerings (Greenet al., 2001).

However, the conjoint analysis models imple-mented in the Internet environment to assess buyerpreferences use choice-based conjoint models, orhybrid models that require subjects to directly esti-mate (‘‘self-explicate’’) both the desirability andthe importance of attributes and levels. These meth-odologies also use direct assessments of preferenceintensities, or preference differences to more closelymirror the assessment methodologies appropriatefor the measurable value function (2).

Finally, the AHP methodology may be given afirm theoretical foundation through an interpreta-tion of its assessment questions as judgmentsregarding ratios of preference differences. Viewedfrom this perspective, and with proper scaling ofthe single-attribute value functions, the AHP is analternative approach to assessing the measurablevalue function (2) (Dyer, 1990).

4. Examples of MAPM in e-commerce

There are a number of examples of applications ofMAPM in e-commerce. In this section, we highlightsome of these examples, and discuss the assessmentmethods that they employ to approximate modelsof the users’ preferences. This discussion is summa-rized in Table 1. Although this discussion is writtenin the present tense for ease of exposition, most ofour observations were made in February 2002 and

we realize that some of the web sites surveyed inTable 1 may no longer exist in the form that wedescribe. In some cases, it seems that persons notfamiliar with MAPM intuitively used related con-cepts and features in the design of these sites. Oftenthey could benefit from some simple changes in theassessment protocols. In other cases, the site descrip-tions explicitly indicate that formal MAPMapproaches are used in their implementation.

4.1. Examples of MAPM models for B2C decision

support

Consumers are often interested in evaluating anumber of alternatives prior to finalizing a purchasedecision. In recent years, many websites have beencreated that provide information to consumersregarding specific classes of products. Haubl andTrifts (2000) report that these online shopping assis-tants may have strong favorable effects on both thequality and efficiency of purchase decisions, allowingshoppers to make better decisions with less effort.

In many cases, the consumers must specify thealternatives they wish to consider, and the site sim-ply provides product specifications along with linksto potential dealers and stores. Examples of suchsites would include Auto-by-Tel (www.autoby-tel.com) and MySimon.com (www.mysimon.com).

Other sites now offer enhanced decision supportto assist the consumer in the initial screening of alter-natives, and in the in-depth comparison of thisreduced set of alternatives, the consideration set,based on multiple criteria. Dealtime (www.deal-time.com) offers the ability to compare productsand services on a variety of attributes. For example,the site provides a comparison of long distance tele-phone services based on the user’s telephoneexchange. An initial screen provides a complete listof the alternative service providers in the relevantarea, ranked by a single attribute (e.g., annual sav-ings). The user is allowed to choose among severalother attributes to re-sort these rankings. The useris then allowed to select a subset of the availablealternatives, and a second screen provides a side-by-side comparison of this consideration set on anexpanded number of attributes, and with additionaldetails. The process of reducing the number of alter-natives considered is often called winnowing (e.g.,Edwards and Fasolo, 2001).

CallVision (www.callvision.com) provides a nat-ural extension of the capabilities of the Dealtime siteby providing technologies to assist a consumer

http://www.mysimon.com


http://www.dealtime.com

http://www.callvision.com


http://www.money.cnn.com

Table 1Summary of selected websites

Applications Provide/request Steps

B2C

Auto-By-Tel www.autobytel.com Product information Select from a set of alternatives, receive matching products andsellersMySimon.com www.mysimon.com

Dealtime www.dealtime.com Product information andcomparisons

Sort the alternatives by attributes, and receive detailedinformation on consideration set

CallVision (I-Optima SST)www.callvision.com

Product categories andattributes

Input information about current product characteristics,receive ranking of the plans

Best Place to Live (MoneyMagazine) www.money.cnn.com

Recommendation of best cityto live in

Provide weights for seven pre-specified attributes and receiveranking of cities

Active Decisionswww.activebuyersguide.com (nolonger a standalone site)

Product categories andattributes

Choose the important attributes, indicate desirability of levelsof performance, determine importance weights, receive productrankings

My Product Advisorwww.myproductadvisor.com

Online Insightwww.onlineinsight.com

Match consumers withproducts based on productfeatures

Provide importance ratings of the attributes, make choicesrevealing trade-offs, receive product list

B2B

Perfect Source www.perfect.com Match suppliers to buyersbased on vendor and productfeatures

Provide importance weighting of requirements, receivecandidate list

SAS www.sas.com Perform weighted supplierranking to identify the bestsuppliers

Choose important criteria and use a dynamic weightingapproach to rank the suppliers

Asparity Decision Solutionswww.plansmartchoice.com

Match employees to healthcare product plans


Colmart www.colmart.com Match buyers and sellers ofpaper products



considering multiple services or products based on aMAPM evaluation of alternatives. For example, aconsumer using Call Vision’s I-Optima SST to eval-uate wireless telephone services would be presentedwith the following product attributes (or objectives):price, quality, coverage, and data services. The‘‘weights of importance’’ for the MAPM modelare estimated through comparisons of the impor-tance of the other three attributes with the impor-tance of price, using percentages.

The site uses a novel approach based on compar-isons to the consumer’s current service provider toestimate the scores of the alternatives on the criteria.The consumer is asked to identify her current ser-vice provider, and to rate its performance on thefour attributes above. In addition, the consumer isasked to input her current monthly bill, and infor-mation regarding such usage issues as the percent-age of long distance calls, roaming calls, and anyrequired features such as Internet access. The resultis a ranking of the plans offered in the consumer’s

area, based on these inputs. As before, the consumercan choose one or more alternatives from thisranked list for a side-by-side comparison that offersadditional details.

Active decisions (www.activebuyersguide.com)provides a more sophisticated example of a multi-dimensional search based on MAPM concepts. Asshown in Fig. 1, a user considering the purchaseof a camera is able to search for a product basedon how she anticipates using the product or basedon the product attributes. Novices may prefer tothink in terms of usage – e.g., I like 8 · 10 prints –while experts may prefer to think in terms of theattributes – e.g., I need at least 4 megapixels (Mitch-ell and Dacin, 1996). The current implementation ofActives Buyers Guide does not allow the user toprovide weights for the attributes, but the user cansort on any attribute. Previous implementations ofActive Buyers Guide and other products offeredby Active decisions do allow the specification ofweights as discussed in Section 5.

http://www.activebuyersguide.com

http://www.autobytel.com


http://www.dealtime.com


http://www.money.cnn.com

http://www.activebuyersguide.com

http://www.myproductadvisor.com

http://www.onlineinsight.com

http://www.perfect.com

http://www.sas.com

http://www.plansmartchoice.com

http://www.colmart.com

Fig. 1. Usage or attribute focus at Active Buyers Guide.


Another website that does address the importanceof attributes is My Product Advisor (www.mypro-ductadvisor.com). As shown in Fig. 2, My ProductAdvisor initially asks the user to think about howshe will use the camera, in addition to how impor-tant each aspect of usage is relative to the others.The user is given a list of ‘‘other’’ attributes thatshe might include in the analysis, and has the optionof specifying value function scores for different levelsof some of the attributes to approximate the valuefunctions of these attributes. For example, it maybe reasonable to assume that the value function forcost is linear, particularly after a consumer has lim-ited the range of costs to consider. However, thevalue function for resolution may need to be assessedas it is likely to be concave for all but the most expertusers. Finally, she can specify the attribute weightsfor her selected attributes, as shown in Fig. 3.

These responses provide an approximation to theweights and single-attribute value function scores of

the additive MAPM model (2). Given this informa-tion, a list of cameras is generated and ranked usingthe underlying MAPM model. Once again, addi-tional information regarding each camera is avail-able by clicking on the appropriate icon.

Online Insight’s Solution Suite (www.onlinein-sight.com) provides another example of aconsumer-facing software product that allows con-sumers to express their personal preferences andtrade-offs for products and services. Based on theelicitation of these preferences, the software thenmatches the consumer with the appropriate prod-ucts. Each recommended product is supported witha graphical representation showing how it ranksagainst the needs and preferences expressed by theconsumer. The stated objective of this feedback isto increase consumer confidence and trust, and toincrease consumer loyalty.

An example of the dialogue used by the OnlineInsight solution may be useful. Suppose that a com-





Fig. 2. Relative importance of camera usage features at My Product Advisor.

Fig. 3. Specifying attribute weights at My Product Advisor.


pany has several options in a particular product linethat may be differentiated by the features: brand,

color, top speed, warranty, and price. The collabo-rative conversation with a consumer begins by


obtaining a rudimentary importance rating of thesefeatures on a 1–5 scale, with 5 being the most impor-tant. In some cases, the system will assume ordinalpreference information concerning such features asprice (cheaper is better), speed (faster is better),and warranty (longer is better). For categorical fea-tures, such as brand and color, the consumer is alsoasked to make preference ratings using the same5-point scale.

Next, the consumer is asked to help refine theunderlying model of her preferences by asking suchquestions as the following: Which of the followingproducts would you purchase?

Product A
Product B
Brand
Brand X Brand Y Price 2000 1800 Color Green Blue
Based on a series of questions such as this, theunderlying preference model is ‘‘tweaked’’ usingconjoint analysis to assess the single-attribute valuefunctions.

There are other examples of recommendationsystems that are not intended to induce a purchase,per se. For example, Money Magazine offers a sup-port tool to help users choose a city to reside. Theuser can weight seven pre-specified criteria, includ-ing the possibility of allocating no weight, and theseweights are used to rank cities. The criteria includeaspects of weather, crime, culture and job growth.

4.2. Examples of MAPM models for B2B decision

support

There are numerous examples of commercial B2Bsoftware and websites that offer users the opportu-nity to match products and services based on multi-ple criteria. In many cases, the details of how theseproducts function are proprietary, but some of themappear to be using the basic concepts of MAPM. Wewill first provide examples of software products thatare available to implement such ideas, and then dis-cuss a few representative examples of websites thatoffer MAPM matching capabilities. More examplesof the use of MAPM in B2B applications can befound in Teich et al. (2004). We focus on situationswhere the goal is to find a stand-alone alternativebut note that MAPM can easily be extended to deci-sion contexts involving portfolios of several alterna-tives (Kleinmuntz, 2007).

4.2.1. MAPM in B2B software

Perfect Source (www.perfect.com) is a modularset of software products that provide an end-to-end strategic sourcing system. The module to sup-port buyer purchasing includes the ability to intro-duce several ‘‘requirements’’ or criteria (e.g.,shipping, return policy, and restocking fee), and todetermine the ‘‘importance weighting’’ of theserequirements using intuitively appealing slider bars.Suppliers use a complementary module allowingthem to respond in terms of multiple requirementsas well. The resulting system is called a weightedscorecard, which scores and ranks bids and thenoptimizes them into suggested multi-source awardpackages line-item by line-item.

SAS� (www.sas.com) offers a Supplier Relation-ship Management software suite that is designed toassist procurement professionals in managing theirstrategic sourcing efforts. The suite is composed offour related modules, with the first being a data basethat supports the other decision support tools. TheSpend Analysis module allows the user to rank sup-pliers according to the criteria most important tothe company. The criteria that are often used insuch applications include supplier type, quality ofsupply, lead times, purchase price, and financial via-bility. The software uses a ‘‘dynamic weighting’’ ofthe chosen criteria to add ‘‘balance and flexibility’’to the process of ranking the suppliers.

In addition, the Procurement Scorecard moduleprovides the ability to monitor the selected suppliersand the entire procurement organization against aspecific set of ‘‘key performance indicators’’. Thisscorecard will alert the appropriate personnel whenprocurement objectives are in danger of not beingmet. Finally, the Sourcing Strategy module includesa wizard-driven application that guides businessusers through the process of identifying measurableobjectives and defining business rules to constrainthe model, which then determines the suppliers touse and calculates the amount to be spent with eachof them to achieve the stated objectives.

As Milgrom (2004) notes, electronic markets willbe utilized only when they offer value relative to tra-ditional supply channels. These value-added servicescould include reduced transaction costs or enhancedinformation. Milgrom (2004) asserts that support-ing the evaluation of ‘‘true multi-dimensional cus-tom offers in response to complex, weighted,concurrent, multi-dimensional RFQs’’ allows a‘‘better match between buyers and sellers . . . helpingtailor the non-price aspects of the transaction . . .’’.


http://www.sas.com


4.2.2. MAPM in B2B applications

Asparity Decision Solutions (www.plansmart-choice.com) has developed a decision support appli-cation to assist end users in sorting through theoptions and choosing a health plan that best fitstheir needs. The preference module asks the userto perform three steps. First, she clicks on the attri-butes or features of a health plan that are importantto her, and, second, she rates the level of importanceof the chosen attributes. Third, she must respond toa series of trade-off questions. The preference mod-ule then creates a unique preference function foreach user, matches it with the available health plans,and presents the results in rank order. We use theterm ‘‘preference function’’ because it is not clearif a utility or value function or some other prefer-ence representation is being utilized by the web site.Summaries of this preference information are avail-able to employers to aid them in negotiating newoptions with health care providers.

The objective of the Colmart.com Internet trad-ing site (www.colmart.com) is to provide a forumfor buyers and suppliers to buy and sell paper prod-ucts. Much like the Asparity Decision Solution, thebuyer is allowed to set ‘‘priorities’’ or weights onproduct attributes from a dropdown list, using ascale from 1 to 5. The buyer then specifies ‘‘idealvalues’’ for each of the attributes, and an approxi-mation to a single attribute value function is usedto score each alternative based on its distance fromthe specified value. The alternatives are thenselected based on a weighted sum of these distancemeasures, which are a proxy for an additive multi-attribute value function. As an alternative, the usercan enter the range of acceptable levels on eachattribute, and the software will find all of the sup-plier alternatives that match the profile. Keskinocaket al. (2001) provides additional details regardingthese matching processes.

This process is an example of a multi-attributereverse auction (or procurement auction) (e.g.,Bichler and Kalagnanam, 2005; Parkes and Kala-gnanam, 2005; Teich et al., 2006). These approachesallow potential suppliers to bid on multiple attri-butes, in some cases in an iterative fashion, anduse the potential buyers’ MAPM to score the bids.There must be some trust on the part of the buyersthat they will not be exploited by providing theirMAPM and typically the potential sellers are givenlimited feedback in the form of changes in buyer bidevaluations. Early experimental work suggests thatbuyers are better off using these systems and seller

profits are not significantly degraded (Chen-Ritzoet al., 2005; Bichler, 2000).

Online negotiation is another possible applica-tion of MAPM in online settings. As outlined inKersten and Noronha (1999) these negotiation sup-port systems (NSS) can be used to improve the out-comes of both parties by directing the processtowards pareto-dominating alternatives based onthe specified preferences. NSS also require the nego-tiators to trust the system enough to providedetailed preference information.

5. Some guidelines for building MAPM applications

The examples surveyed above provide some indi-cation of the widespread adoption of MAPM as abasis for modeling consumer preferences in theB2C and B2B environments. This discussion alsofeatures a variety of assessment protocols, anddegrees of sophistication in the MAPM modelingefforts. In some cases, these differences may reflectconscious decisions by the website designer in trad-ing off accuracy of the MAPM representation versusthe cognitive burden imposed on the user. In othercases, these differences may reflect a lack of aware-ness of some basic concepts of MAPM that wouldenhance the users’ experience with the website.

In this section, we begin with a discussion ofsome recommendations that should be given seriousconsideration in most website implementations ofMAPM-like decision aids. Next, we acknowledgethat a clever implementation of a MAPM assess-ment technique might actually be used to manipu-late preferences, and identify some ways in whichthis might be done. On the one hand, a web designermight wish to take advantage of some of theseinsights to select how information is presented toa consumer to influence her choices, within ethicallyappropriate standards. On the other hand, the userof such websites may also want to be aware of thepotential for the manipulation of her expressed pref-erences, so that she may avoid being guided to amisstatement of her preferences, and led to alterna-tives that she would otherwise not prefer.

5.1. Avoiding common mistakes in MAPM website

design

Our review of the applications of MAPMapproaches to support consumer decisions in B2Cand B2B websites indicates that several mistakesare commonly made in their designs. Further, these



http://www.colmart.com


mistakes are relatively easy to overcome in this envi-ronment, and should be considered in the design ofany decision aid based on preference modeling.These mistakes are not unique to web-based imple-mentations of MAPM, and Keeney (2002) discussessome of these issues in a more general setting.

5.1.1. Using attributes rather than objectives

The identification of the appropriate objectivesfor an application of a MAPM is an important step,and requires an understanding of a subtle butimportant notion. That is, it is important to sepa-rate fundamental objectives from means to accom-plish objectives. A simple example, borrowed fromClemen and Reilly (2001) may help to illustrate thisconcept. For an individual, working fewer hoursmay seem to be an important objective. However,it may be the case that this is only a means ofaccomplishing something else that is really impor-tant to the individual, such as spending more timewith family. In this case, spending more time withfamily would be a fundamental objective, whileworking fewer hours would be a means objective.

This distinction is an important one that relatesto the design of e-commerce B2B and B2C websiteinterfaces. Most sources of information regardingproducts and services provide information relatedto the attributes of the product or service. Theseare typically things that can be objectively mea-sured, and relate in some manner to aesthetic oroperational features. However, there is no a priori

reason to believe that these product attributes willbe the appropriate measures for a consumer. Inmany cases, these product attributes may be rele-vant means objectives, but may need to be com-bined in some unique and personal manner by theconsumer to provide information related to a funda-mental objective. These product attributes may alsorelate to more than one fundamental objective.

A focus on product or service attributes in a webdesign is a natural mistake, and is also common intraditional ‘‘hard copy’’ reviews of products andservices. However, when a MAPM model is devel-oped based on these attributes, there are at leasttwo issues that may lead to poor approximationsof the user’s preferences.

First, the dialogue with the consumer that isrequired to develop a model of her preferencesrequires several value trade-offs, where comparisonsmay be required between hypothetical (or real)alternatives that differ on some attributes. Thesetrade-offs may be very difficult for the individual

to make when the attributes do not map directlyinto her fundamental objectives. For example, aconsumer considering the purchase of an automo-bile may be given attribute values for horsepowerand torque, which may be viewed as proxy attri-butes for performance. Unless the consumer is anautomotive engineer, it may be very difficult forher to thoughtfully make value trade-offs between,say, an additional 20 lb-ft of torque versus a reduc-tion of 10 hp.

Perhaps her fundamental objective is more clo-sely related to acceleration, and some measure ofperformance on this product attribute may be muchbetter suited to her concerns and values. Naturally,we must also recognize that the expertise of the userof a web system may have an impact on the selec-tion of the appropriate measures of performancefor the product or service. This suggests that aweb-based system that offers the consumer thechoice among several different product and serviceattributes would be preferable to one that implicitlyassumes a universal set of fundamental objectivesamong all users. It should also provide sufficientinformation regarding the available attributes sothat the user, with a click of the mouse, can obtainuseful information to clarify the meaning andpotential significance of the attribute.

However, this flexibility regarding the attributeshas its own potential for misuse, since a consumerinterested in automobile performance might betempted to select horsepower, torque, and 0–100 kmh acceleration as relevant attributes for herMAPM. Some careful thought might lead to therecognition that horsepower and torque (combinedwith weight), may be means objectives related tothe fundamental objective of acceleration, so includ-ing all three could lead to ‘‘double counting’’regarding this objective, and place too much weighton performance relative to the consumer’s truepreferences.

This latter point emphasizes the second concernregarding the use of attribute values in a MAPMmodel. As we discussed, most website applicationsof MAPM or MAPM-like approaches to recom-mending alternatives use an additive weightedmodel. The independence assumptions that arerequired to ensure that the additive form is appro-priate are typically not verified, and the attributesthat are used in the decision aid may not be compat-ible with these assumptions. In such a case, theseattributes may be relevant for the decision as meansobjectives, but they may need to be combined in a


non-linear fashion to reflect the preferences of theconsumer. Montibeller et al. (2005) discuss the com-plications of transforming a mental model of a deci-sion situation into a value hierarchy.

A related issue is that many websites specify theattributes, and implicitly the objectives, for the deci-sion maker. For example, Fig. 4 shows the sevenattributes used to rank cities to live in by MoneyMagazine. While it is clear that ‘‘Weather’’ is likelyto be an important consideration for most consum-ers, the website requires that these concerns be cap-tured in ‘‘Sunny Weather’’. A previousimplementation of the same website was used byone of the authors several years ago. The authorwas surprised by the recommendations as he pre-ferred warm, winterless, weather and was told heshould live in New England. After some investigat-ing he realized that this was due to the high weighthe specified for the attribute ‘‘Weather’’. The web-site defined ‘‘Weather’’ as having four seasonswhich was not consistent with the author’s prefer-ences for a single, warm season. After lowering theweight on the weather attribute – although theauthor felt that his definition of weather was veryimportant – the recommendations were improved.

Butler et al. (2006) develop an alternativeMAPM procedure where the attributes are com-bined to form mental models that predict the perfor-mance of alternatives on the fundamental objectivesof the decision maker. In the first step the decisionmaker specifies weights that reflect each attribute’s‘‘impact’’ on each of the objectives, and thereforecreates a predictive model for that objective. Inthe second step the decision maker specifies weights

Fig. 4. Selecting a place to live at Money Magazine.

on each major objective, and therefore creates apreference model for the overall objective. Thenthe weights assessed in steps one and two are com-bined to determine the appropriate weights on theattributes to be used as a basis for ranking alterna-tives and for choice. The predictive model weightsare statements of belief regarding the achievementof each objective while the objective model weightsare statements of preference, and we believe this dis-tinction should be explicitly recognized. Ignoringthis distinction and confounding judgments of beliefand judgments of preferences violates a fundamen-tal principle of classical decision theory.

The designer of a MAPM website should beaware of these issues, and carefully select the prod-uct attributes that will be made available to the user.In some cases, an interactive dialogue focused onidentifying the user’s fundamental objectives mightbe considered to guide the selection of the appropri-ate attributes. For example, Active Buyers Guideallows the user to specify either desired camerausage or desired attribute levels as shown inFig. 1. The first step at www.myproductadvisor.comis to specify the consumer’s objectives in terms ofhow a digital camera is to be used before askingabout attributes (Fig. 2). Later in the process theconsumer is asked about camera resolution as mea-sured by megapixels but this quantity is related towhat consumers actually value: photo clarityand photo quality. Finally, as illustrated in Fig. 5,MySimon shows actual images at a variety of mega-pixel choices so the user can see the implications ofchoosing a particular camera resolution. Wheneverpossible, the web designer should enable the userto understand how a product attribute relates toher objectives.

5.1.2. Making value trade-offs independent of

attribute range information

Many of the website applications present theconsumer with a list of attributes, and ask her torate their relative importance, perhaps on a scalefrom 1 to 5, without any information regardingthe ranges over which these attributes might actu-ally vary among the alternatives to be compared.The responses to these types of questions mayinclude significant errors, at best, or may be verymisleading when used in a MAPM model of prefer-ence. This issue is prevalent in self-explicated andhybrid implementations of conjoint analysis (Gib-son, 2001) and also in applications of the AHP(Dyer, 1990).


Fig. 5. Mapping megapixels to photo quality at MySimon.


The weights in the additive value model (2) reflectthe relative importance of the ranges of attributeoutcomes in determining the overall value. This isdifferent from the notion of the intrinsic importanceof the attribute per se. Returning to the example ofthe automobile purchaser, most consumers wouldprobably agree that the most important objectivein purchasing an automobile would be to minimizethe cost of the vehicle that is chosen. However, oncethe consumer restricts herself to a consideration setof vehicles within the same general price class, theimportance of cost may diminish.

For example, if the consumer says that she hasdecided to purchase a mid-size, four-door sedan,and would like to have the assistance of an auto-mated decision aid in selecting among the alterna-tives, she may recognize that the cost of thesealternatives varies from €24000 to €28000, but theydiffer significantly in terms of such features as horse-power, safety ratings, and quality ratings. Withinthis subset of all automobiles that are available forpurchase, she may consider it much more importantto have an increase from the lowest to the highest

safety rating among this group, than to decreasethe cost of the purchase from €28000 to €24000.In such a case, the weight in the MAPM model onthe attribute of safety should be larger than theweight on cost (see Keeney, 2002 or Kirkwood,1992 for an elaboration regarding this importantpoint).

The swing weight method (von Winterfeldt andEdwards, 1986) provides an alternative to thetrade-off method discussed earlier, and requires thedecision-maker to consider the relative importanceof changes from the worst to the best levels on eachof the objectives. Fischer (1995) has investigated thisphenomenon, and offers suggestions for ensuringthat consumers are sensitive to the attribute ranges.Otherwise, there may be significant errors in theunderlying model.

We are not aware of a website that intentionallyfocuses the user on the range of attribute perfor-mance when assessing attribute weights. For exam-ple, there is no indication of range information atMy Product Advisor (Figs. 2 and 3) or Money Mag-azine (Fig. 4).


5.1.3. Making the process an unpleasant experience

The standard protocol for the elicitation of aMAPM model can require a number of tedioustrade-off questions within and across product attri-butes that may be hard to compare, poorly under-stood, and not obviously representative of any realalternatives. The process can be time-consuming,and frustrating for the user.

Ideally, a consumer will be offered a process thatis educational, informative, and fun. It is well knownthat decision makers will choose decision processesor heuristics based on recognition of the cognitivecost of the processes versus the perceived benefitsof greater decision accuracy (Payne et al., 1993). Aweb-based MAPM decision aid should be imple-mented with the recognition that the number ofquestions to be asked, and the cognitive burden thatis placed on the user should be commensurate withthe potential gains or losses associated with someerrors in the resulting recommendations.

Milgrom (2004) argues that for e-markets to suc-ceed, they must be scalable. As an example, onenecessary condition for scalability in the B2B envi-ronment is that suppliers are able to automate thescreening of RFQs so that they only receive thoserequests that they are most likely to be able to meet.Similarly, buyers need help sifting through a poten-tially huge number of offers.

An increase in the quality of information and thereduction of the transaction costs should be goals ofany recommendation system, search engine or intel-ligent agent. Many recommendation websites askthe user how many results to return. While this cer-tainly reduces the cognitive burden, there is nomechanism to insure that the user will choose to lookat the ‘‘right’’ number of alternatives. Other authorshave incorporated a measure of cognitive burdendirectly in the MAPM model (e.g., Montgomeryet al., 2004) forcing the decision maker to make apotentially difficult trade-off between some measureof mental processing load and the size of the consid-eration set. A MAPM allows a more straightforwardway to make the comparison of alternatives less dif-ficult: (1) by scoring the alternatives in an ‘‘intuitive’’unit to ease comparison across the alternatives pre-sented, and (2) by allowing the user to specify a min-imum ‘‘significant’’ difference to reduce the numberof alternatives presented. Both are related to arescaling of the MAPM scores of alternatives, as dis-cussed in Butler et al. (2001).

Sarin (1979) and Kirkwood and Sarin (1985) haveinvestigated how MAPM models can be approxi-

mated with relatively little cognitive burden on theuser, and how errors in the specification of thesemodels may impact the rankings of alternatives thatthey generate. Ravinder and Kleinmuntz (1991),Butler et al. (1997) and Jia et al. (1998) also provideresults that explore the implications of errors inassessments on the actual quality of decisions. Thisresearch may provide useful guidelines for determin-ing the numbers of trade-off questions that should beasked, or perhaps for identifying questions that willreduce the consideration set most efficiently.

The task of making the process pleasant seems tocontradict the other recommendations; it may bedifficult and unpleasant to think hard about prefer-ences and carefully identify objectives and attri-butes, and to make weight judgments. Here againthe contrast between a domain expert and noviceis important. An expert may even enjoy thinkinghard about her trade-offs and preferences. The costof the item and the importance of the decision mayalso influence the willingness of the consumer tothink hard about these issues. Perhaps the best solu-tion is to provide multiple mechanisms for userswith different abilities and desires. For example,Fig. 1 from Active Buyers Guide shows two inter-faces to identify evaluation attributes.

5.2. The construction of preferences

A great deal of research has confirmed what a lit-tle introspection by anyone would recognize: prefer-ences do not ‘‘exist’’ in the minds of individuals, butare constructed as the need for them arises (for dis-cussions, see Slovic, 1995 and Bettman et al., 1998).Further, the preferences that are constructed by anindividual may depend in a significant manner onthe context of the choice situation, even when thealternatives are identical (Loewenstein, 2001). Thisinsight offers two contrasting points of view thatrelate to the implementation of web-based MAPMdecision aids.

First, to assess an accurate model of an individ-ual’s preferences, the process of questioning shouldbe designed to minimize the impacts of well-knownbiases. For example, Tversky et al. (1998) docu-mented the existence of the prominence effect: Themore prominent (important) attribute will weighmore heavily in a choice between two alternativesthan in a trade-off question between the same alter-natives (Fischer et al., 1999).

We presented an example of a weight assessmentquestion when considering the price and resolution


of a digital camera in Section 2: A = (€200, 1600 ·1200 pixels) and B = (___, 1920 · 1600 pixels); thedecision maker is asked to specify a price for cameraB so that the two alternatives are equally attractive.Fischer et al. (1999) and Tversky et al. (1998) wouldcall this a matching task because the decision makeris asked to provide a response in terms of one of theattributes. Alternatively, a decision maker might beasked to choose between two fully specified alterna-tives in a choice task. For example, would you prefercamera A 0 = (€200, 1600 · 1200 pixels) or B 0 =(€300, 1920 · 1600 pixels).

Previous research suggests that in a matchingtask a decision maker is much more likely to placeless weight on the prominent attribute when com-pared to a choice task. The experimental evidenceis consistent with the conjecture that in the morequalitative choice task the prominent attribute isnot formally priced out in the mind of the decisionmaker, while the stimulus in the matching taskforces the decision maker to explicitly make thetrade-off (Fischer et al., 1999). Since choice andmatching questions may be used to assess theweights in a MAPM model, Keeney (2002) suggestsusing both and pointing out any implied inconsis-tencies in the responses to the decision maker toresolve this bias appropriately. Of course, this mayincrease the cognitive burden on the decision maker,especially if the approach is implemented in anautomated environment, and may diminish her will-ingness to use the decision aid. This is again an issueof effort versus accuracy that must be considered inthe design of a useful recommendation system.

More of the recommendation systems we sur-veyed have assessment procedures featuring choice

Fig. 6. Example of a combined choice/matching task fro

tasks as opposed to matching tasks. However therehave been some attempts to combine the two asillustrated in Fig. 6. A previous version of ActiveBuyers Guide asked consumers not only to state apreference (choice) for one of two options, but alsoto specify the strength of preference for one optionover the other. This type of question may help bal-ance the two types of assessment tasks.

Second, the observation that preferences may beconstructed suggests that a website may be designedto manipulate preferences to increase the likelihoodthat a particular product or service will be recom-mended for consideration. A simple example shouldmake this clear. As we have noted, a MAPM-basedwebsite will typically provide a list of objectives orattributes as the basis for the evaluation of a prod-uct or services. It is likely that many subjects mayaccept the set of objectives or attributes suggestedfor an evaluation, without expending the cognitiveeffort necessary to identify other attributes thatmight be more relevant to reflect their fundamentalobjectives. To the extent that preferences are actu-ally constructed during the process of preferenceelicitation, the set of attributes that are providedmay even influence the fundamental objectives thatthe user considers to be relevant for the evaluationof the product or service. Relatively inexperiencedconsumers lacking expertise regarding a productmay be particularly susceptible to manipulationregarding the appropriate attributes to use in theevaluation and selection of an alternative, as sug-gested in the discussion of the use of a website torecommend a city in which to reside (Fig. 4).

Efforts to suggest product attributes that arefavorable to a particular alternative are certainly

m an old implementation of Active Buyers Guide.


not new. Product advertisements in the traditionalmedia of newspapers and magazines will often offerside-by-side comparison matrices with rival brands,but with selected product criteria that clearly favorthe sponsoring product. Even the choice of whichalternative will serve as the basis of comparisoncan impact the assessed preferences. This suggeststhat ‘‘independent’’ shopbots and electronic agentsin the B2C environment should be required to revealany ties to particular products or services that theymay be used to evaluate.

Preferences may also be biased, or constructed,based on even more subtle manipulation of MAPMassessment methodologies. Stillwell et al. (1987),Weber et al. (1988) and Borcherding and von Win-terfeldt (1988) have shown that providing additionalattributes related to some fundamental objectivewill cause a user of the MAPM methodology toassign too much ‘‘weight’’ to the objective, asimplied by the numerical weights assessed for theindividual attributes. Returning to the automobileexample, an evaluation system that included a largenumber of attributes related to performance, suchas torque, horsepower, power-to-weight ratio, 0–100 kmh acceleration, 60–80 kmh acceleration, andso forth, would encourage users to construct prefer-ence models that would give relatively more‘‘weight’’ to the performance of the automobile.And, a consumer who used such a system couldbe influenced to conclude that performance is actu-ally more important as a fundamental objectivethan she may otherwise have believed.

These observations suggest that a cleverlydesigned website could be used to construct the pref-erences of consumers in such a way that they wouldbe influenced to choose more expensive vacations,say, or more sophisticated cameras or computers,simply through the dialogue used to elicit their pref-erences. This insight should come as no surprise, ofcourse, since skillful salespersons have been doingthe same thing for years. There is no reason tonaively assume that the results of an interaction withan automated decision aid produce more ‘‘objective’’product evaluations than one carried out with theaid of a salesperson.

An example of research that relates to this phe-nomenon is provided by Mandel and Johnson(1998), who show that the background of a websitecan suggest the importance of some of the productattributes, and alter user choices based on a singleexposure to the decision aid. This is clearly a ripearea for further research, as noted by Geoffrion

and Krishnan (2001), and also for a recognition thatthe age-old adage ‘‘buyer beware’’ will apply in thee-commerce environment as well.

6. Conclusions

It seems clear that MAPM will play an importantrole in the implementation of websites that provideshopping assistance to consumers in B2C environ-ments and to consumers in B2B environments. Wehave reviewed a number of current implementa-tions. When viewed from the perspective of MAPMand the lessons that have been learned based onresearch related to its application, many of the exist-ing websites fall short of offering the quality of deci-sion support that is feasible. There are someexceptions of course, and we have highlighted somesites that do represent insightful applications ofMAPM concepts. We emphasize again that ourobservations are based on the implementationsactive at these sites in 2002, and that some of themmay have been modified or deleted in the interim.

Many of the recommendations we have made arelikely to increase the cognitive burden of the con-sumer. It is important to verify that the additionaleffort is rewarded with additional benefit (Payneet al., 1993). For example, Butler et al. (2006) inves-tigate the trade-off between the number of requiredassessments associated with explicitly linking attri-butes to objectives and the resulting decision qual-ity. They find that for all but very small decisionproblems, it is worthwhile for the decision makerto expend the effort associated with thinking hardabout how the attributes impact her objectives. Sim-ilar work by Jia et al. (1998) suggests that whileweight approximations do perform relatively well,it is usually worth the effort to perform a formalweight assessment. Nonetheless, we do not advocateadding more and more assessment burden withoutan investigation of the associated benefits.

Some scholars have argued that individuals donot make decisions based on a thoughtful consider-ation of objectives and their relative importance,but rather use a variety of decision heuristics thatare strongly affected by the decision domain (e.g.,Loewenstein, 2001). We would not advocate thatevery consumer would (or should) seek the assis-tance of a website based on MAPM principles whenmaking routine purchases. Nevertheless, it is clearthat many consumers are using the web as an aidto shopping decisions, whether the actual purchase


is then transacted electronically or in a traditionalretail environment.

As we have noted, many B2C websites, such asActive Buyers Guide, offer the consumer multipleapproaches to identifying a most preferred alterna-tive, in recognition of the influence of different fac-tors on the consumer’s choice of a decision aid. Aswe have discussed, MAPM has a role to play in thisenvironment, but it may certainly be the case thatsome web-oriented shoppers will prefer product rec-ommendations based on collaborative filtering oralternate approaches. These alternative approachesare most likely to be preferred by inexperienced con-sumers who need help specifying the attributes, therelationships between the attributes, and their rela-tive importance.

The B2B environment should be an even morefertile ground, however, for MAPM applications,and some evidence in support of this observationis provided by the number of sophisticated softwareproducts and websites that we have reviewed above.The same scholars who have been skeptical of thepotential for rational decision support systems forconsumers offer an insight into why these applica-tions may be successful for B2B consumer support.Loewenstein (2001) argues that individuals do notactually make decisions based on rational theoriesof decision making, but may ‘‘make sense’’ of thosedecisions in retrospect by describing the process inthose terms.

In a B2B environment, the purchasing agent mayfeel the need to document the rationale for the deci-sions that are made, and to leave a ‘‘decision audittrail’’ in terms of the rational models that supportedmajor purchasing commitments. The attributes andtheir weights may reflect an explicit interpretation ofcompany policies, and perhaps reflect the consensusof several group discussions that are then delegatedto one person to implement. Within these environ-ments, the logic and appeal of MAPM would seemto make it a strong candidate for the documentationof purchasing actions.

References

Bettman, J.R., Luce, M.F., Payne, J.W., 1998. Constructiveconsumer choice processes. Journal of Consumer Research25, 187–217.

Bichler, M., 2000. An experimental analysis of multi-attributeauctions. Decision Support Systems 29, 249–268.

Bichler, M., Kalagnanam, J., 2005. Configurable offers andwinner determination in multi-attribute auctions. EuropeanJournal of Operational Research 160, 380–394.

Borcherding, K., von Winterfeldt, D., 1988. The effect of varyingvalue trees on multi-attribute evaluations. Acta Psychologica68, 153–170.

Butler, J.C., Jia, J., Dyer, J.S., 1997. Simulation techniques forthe sensitivity analysis of multi-criteria decision models.European Journal of Operational Research 103, 531–545.

Butler, J.C., Morrice, D., Mullarkey, P., 2001. A multipleattribute utility theory approach to ranking and selection.Management Science 47, 800–816.

Butler, J.C., Dyer, J.S., Jia, J., 2006. Using attributes to predictobjectives in preference models. Decision Analysis 3, 100–116.

Chen-Ritzo, C., Harrison, T.P., Kwasnica, A.M., Thomas, D.J.,2005. Better, faster, cheaper: An experimental analysis of amulti-attribute reverse auction mechanism with restrictedinformation feedback. Management Science 51, 1753–1762.

Clemen, R.T., Reilly, T., 2001. Making Hard Decisions withDecisionTools. Duxbury Press.

Dyer, J.S., 1973. A time-sharing computer program for thesolution of the multiple criteria problem. ManagementScience 19, 1379–1383.

Dyer, J.S., 1990. Remarks on the analytic hierarchy process.Management Science 36, 249–258.

Dyer, J.S., Sarin, R.A., 1979. Measurable multi-attribute valuefunctions. Operations Research 22, 810–822.

Edwards, W., Fasolo, B., 2001. Decision Technology. AnnualReview of Psychology 52, 581–606.

Farquhar, P.H., Keller, L.R., 1989. Preference intensity mea-surement. Annals of Operations Research 19, 205–217.

Fischer, G.W., 1995. Range sensitivity of attribute weights inmulti-attribute value models. Organizational Behavior andHuman Decision Processes 62, 252–266.

Fischer, G.W., Carmon, Z., Ariely, D., Zauberman, G., 1999.Goal-based construction of preferences: Task goals and theprominence effect. Management Science 45, 1057–1075.

Forman, E., Gass, S.I., 2001. The analytic hierarchy process: Anexposition. Operations Research 49, 469–486.

Geoffrion, A.M., Krishnan, R., 2001. Prospects for operationsresearch in the e-business era. Interfaces 31, 6–36.

Gibson, L., 2001. What’s wrong with conjoint analysis. Market-ing Research 13, 16–19.

Green, P., Krieger, A.M., Wind, Y., 2001. Thirty years ofconjoint analysis: Reflections and prospects. Interfaces 31,S56–S73.

Haubl, G., Trifts, V., 2000. Consumer decision making in onlineshopping environments. Marketing Science 19, 4–21.

Jia, J., Fischer, G.W., Dyer, J.S., 1998. Attribute weightingmethods and decision quality in the presence of responseerror: A simulation study. Journal of Behavioral DecisionMaking 11, 85–105.

Kahn, B.E., Huffman, C., 1998. Mass customization. In: Dorf,R.C. (Ed.), The Handbook of Technology Management.CRC Press, Inc.

Keeney, R.L., 2002. Common mistakes in making value trade-offs. Operations Research 50, 935–945.

Keeney, R.L., Raiffa, H., 1976. Decision with Multiple Objec-tives: Preference and Value Trade-offs. Wiley, New York.

Kersten, G.E., Noronha, S.J., 1999. WWW-based negotiationsupport: Design, implementation, and use. Decision SupportSystems 25, 135–154.

Keskinocak, P., Goodwin, R., Rama, F.W., Akkiraju, R.,Murthy, R.S., 2001. Decision support for managing an


electronic supply chain. Electronic Commerce Research 1, 15–31.

Kirkwood, C.W., 1992. Estimating the impact of uncertainty on adeterministic multi-attribute evaluation. Management Science38, 819–826.

Kirkwood, C.W., 1996. Strategic Decision Making. DuxburyPress, Belmont, CA.

Kirkwood, C., Sarin, R., 1985. Ranking with partial information:A method and an application. Operations Research 33, 38–48.

Kleinmuntz, D.N., 2007. Resource Allocation Decisions. In:Edwards, W., Miles, R.F. Jr.Jr., von Winterfeldt, D. (Eds.),Advances in Decision Analysis. Cambridge University Press.

Krantz, D.H., Luce, R.D., Suppes, P., Tversky, A., 1971.Foundations of MeasurementAdditive and Polynomial Rep-resentations, Vol. I. Academic Press.

Loewenstein, G., 2001. The creative destruction of decisionresearch. Journal of Consumer Research 28, 499–505.

Luce, R.D., Tukey, J.W., 1964. Simultaneous conjoint measure-ment: A new type of fundamental measurement. Journal ofMathematical Psychology 1, 1–27.

Mandel, N., Johnson, E., 1998. When web pages influence choice:Effects of visual primes on experts and novices. Journal ofConsumer Research 29, 235–245.

Milgrom, P., 2004. An economist’s vision of the B-to-B market-place. Executive Whitepaper, available from: www.perfect.com.

Mitchell, A.A., Dacin, P.A., 1996. The assessment of alternativemeasures of consumer expertise. Journal of ConsumerResearch 23, 219–239.

Montgomery, A.L., Hosanagar, K., Krishnan, R., Clay, K.B.,2004. Designing a better shopbot. Management Science 50,189–206.

Montibeller, G., Ackermann, F., Belton, V., Ensslin, L., 2005.Reasoning Maps for Decision Aid: An Integrated Approachfor Problem Structuring and Multi-Criteria Evaluation.Working Paper 2005/11, Department of ManagementScience, University of Strathclyde, (available from: http://www.managementscience.org/research/ab0511.asp).

Parkes, D.C., Kalagnanam, J., 2005. Models for iterative multi-attribute procurement auctions. Management Science 51,435–451.

Payne, J.W., Bettman, J.R., Johnson, E.J., 1993. The AdaptiveDecision Maker. University Press, Cambridge.

Ravinder, H.V., Kleinmuntz, D.N., 1991. Random error inadditive decompositions of multi-attribute utility. Journal ofBehavioral Decision Making 4, 83–100.

Roy, B., 1996. Multicriteria Methodology for Decision Aiding.Kluwer Academic Publishers, Dordrecht.

Saaty, T.L., 1980. The Analytic Hierarchy Process. McGraw-Hill, New York.

Sarin, R., 1979. Ranking of multi-attribute alternatives with anapplication to coal power plant siting. Working Paper,Purdue University.

Slovic, P., 1995. The construction of preferences. AmericanPsychologist 50, 364–371.

Stewart, T.J., Losa, F.B., 2003. Towards reconciling outrankingand value measurement practice. European Journal of Oper-ational Research 145, 645–659.

Stillwell, W.G., von Winterfeldt, D., John, R.S., 1987. Compar-ing hierarchical and nonhierarchical weighting methods foreliciting multi-attribute value models. Management Science33, 442–450.

Teich, J.E., Wallenius, H., Wallenius, J., Koppius, O., 2004.Emerging Multi-Issue E-Auctions. European Journal ofOperational Research 159, 1–16.

Teich, J.E., Wallenius, H., Wallenius, J., Zaitsev, A., 2006. Amulti-attribute e-auction mechanism for procurement: Theo-retical foundations. European Journal of OperationalResearch 175, 90–100.

Tversky, A., Sattath, S., Slovic, P., 1998. Contingent weighting injudgment and choice. Psychological Review 95, 371–384.

von Winterfeldt, D., Edwards, W., 1986. Decision Analysis andBehavioral Research. Cambridge University Press.

Weber, M., Eisenfuhr, F., von Winterfeldt, D., 1988. The effectsof splitting attributes on weights in multi-attribute utilitymeasurement. Management Science 34, 431–445.

West, P., Ariely, D., Bellman, S., Bradlow, E., Huber, J.,Johnson, E., Kahn, B., Little, J., Schkade, D., 1999. Agents tothe rescue. Marketing Letters 10, 285–300.



http://www.managementscience.org/research/ab0511.asp

http://www.managementscience.org/research/ab0511.asp

Enabling e-transactions with multi-attribute preference...

Documents

Transcript of Enabling e-transactions with multi-attribute preference...