Business models to offer customized output in electronic commerce

C. Scheer et al. / Business models to offer customized output in electronic commerce 165Complexityhighlimitednonbuilding a house yachtcustomerdrivenindividual journey health care products/servicesoutputconsulting redecoration products/servicesmoving house servicesmeal in a restaurant services around a carinterior decoration products repair patchcustomercentricoutputpackage tour tailor-made clothescar with standard equipmentstandard softwaretrain journeypetrolcommodityseller-drivenoutputfastfoodindividual colour of a carhaircutbasic foodindividual music cdnonlimitedhighIndividualityFig. 2. Classification of customized output from a customer’s perspective [24].– Limited complexity: the features can be chosenfrom a pre-defined selection which offers at leastone feature (e.g. mutual dependent specificationof colour and interior).– High complexity: there are no restrictions for thedesign of features. The customer can determinethe features (e.g. the construction of the car canbe designed).We are now able to transfer the parameters and thegranular gradation into a matrix (cf. Fig. 2).Furthermore we differentiate between three classesof output to classify the different degrees of individualityand complexity. These classes focus on the releasingmoment of manufacturing which can be customerdrivenand/or seller-driven:– Seller-driven output: it is manufactured andstandardized independently from individual customer’sneed. The production process is sellerdriven.– Customer-centric output: it offers a number ofpre-defined options. The customer can customizethe output within these options. The productionprocess is both seller and customer-driven.– Customer-driven output: it allows the customer anindividual design of the output. The productionprocess is customer-driven.See Fig. 2 for a compiled classification of personalizedoutput from a customer’s perspective. We are nowable to measure the felt adaptation in three classes ofoutput by the parameters individuality and complexity.3. Business models to offer customized output inelectronic commerceThe requirements of individualized output have theirsources in customer-driven markets and the correspondingbusiness models. Specifically business modelsfor Electronic Commerce apply modern informationtechnology to shape the seller’s process and theinterface to the customer’s process (cf. Fig. 3).Business model and its interface to the customer’sprocess are different depending on the degree of personalizedoutput. The reason for this can be found inthe temporal consideration of customer’s specificationwhich determines both seller’s process and customer’sprocess. On the basis of the characteristics of personalizedoutput (cf. Section 2) we will therefore discussmentioned differences in the following.The seller-driven output can be completely controlledby the seller and is manufactured independentlyfrom the customer’s needs. The seller’s processesand the organizational structures can be designed in aseller-driven environment. The model of mass productionrealizes the seller-driven output. It leads to standardizedoutput concerning design and distribution [9].Mass production pursues the principle of Henry Ford:“You can have any color car you want as long as it’sblack” [20]. The production of variants can also beused to realize seller-driven output with limited personalization.Here the customer gets products or servicesin different variations of features which are set by themanufacturerand cover average individual needs. Each

166 C. Scheer et al. / Business models to offer customized output in electronic commerceSeller's processand businessmodelMarketingSalesProductionDistributionAfter SalesInteractions onelectronic marketsInformation Agreement TransactionCommunicationinterfaceInformation TechnologyIT/physicalCustomer'sprocessInformationSpecification Order Payment UsageFig. 3. Customer process and seller process and the communication interface between both processes in Electronic Commerce.variation is made for a small group of customers. Thiscan lead to a high number of variants which won’t fitexactly the customer’s needs [19].In the customer’s process the customer selects a completelyseller-driven output. This can be explained bythe example of an used car purchase. Here the customercan not individually manipulate the output, apart fromthe negotiation of the price. The configuration space iszero since neither features or values can be assigned.Therefore there is no need for adaptation. In ElectronicCommerce the seller-driven output is specified usingcatalogs in which all available products and/or servicesand further information are categorized.A customer-centric output will be realized in a processwhich is customer and seller oriented. At thebeginning of the value chain the business processesand the organizational structure are driven by manufacturer’sinterests. This changes at the order penetrationpoint, also called freeze point. At this pointthe seller integrates the customer’s specification withthe production process. In general, the specificationsof the customer are integrated as late as possible. ”-Value chain customization begins with the downstreamactivities, closest to the marketplace, and may thenspread upstream. Standardization, in contrast, beginsupstream, with fundamental design, and then progressivelyembraces fabrication, assembly, and distribution”[9] Starting at the order penetration point, the outputwill be adapted within a range of pre-defined options(i.e. values and features) to fit customer’s needs.Another way of customer orientation is to extend thestandardized product or service with additional valueaddingservices [20]. The concept of mass customizationcan be used to implement the customer-centricmanufacturing of output “with enough variety and customizationthat nearly everyone finds exactly what theywant” [20]. Finally, mass customization offers the customera number of pre-defined values. They can beused to define the also pre-defined features of the output[19]. Individuality can also be created with additionalservices, a specific degree of delivery serviceand a kind of product image. Decisively the customerchooses the options which are relevant for his satisfaction.The resulting complexity for the manufacturer canbe reduced by the mass production of modular output,by new concepts of production, usage of informationtechnology, supply networks and additional points oforder penetration [19,20].Specifying the customer-oriented output the customercan influence the output concerning its featuresand values. The customer evaluates and selects offeredoptions in order to specify his custom output. Usingthe example of a car configuration we can depict thecustomer’s process. The customer specifies differentfeatures (e.g. the colour) within pre-defined values(e.g. colour red, green, blue) and can thereby configurehis individual car. The integration of the customerprocess can be realized by configurators which offer allavailable options and record the customer’s decision.A customer-driven output will be realized with thedegree of individuality and/or complexity determinedby the customer. The organizational structure must bedesigned order specific to combine required resourcesand functions. The trigger of all activities is the customer’sorder [24].

C. Scheer et al. / Business models to offer customized output in electronic commerce 167IntermediaryPublicAdministrationCompany ACompany A1Company BEmployeeCustomerInternet technology, internet services and communication rulesphysical flow / traditional value chaindigital flow / virtual value chainFig. 4. Internet business model to create the customer-driven output [24].Implementation of customer-driven output has beendiscussed in literature in different concepts. One categoryof concepts suggests internet-based business modelsin which different suppliers are aggregated in orderto co-produce custom output [16,22,24]. The conceptof Scheer and Loos for example describes an orderspecificcooperation of suppliers which is initializedand coordinated by an intermediary (cf. Fig. 4). Theintermediary records the customer’s specification andsplits his requirements in different orders. This isnecessary because of the high degree of individualityand/or complexity of the output which can not be realizedby one supplier. The suppliers are mostly integratedby information technology in their value chainand supply chain relationships. They produce parts ofthe output (concerning the order) within their resourcesand core competencies. All parts are integrated at theend of the value chain by the intermediary and offeredas one single output to the customer [24].The customer has more influence to specify the outputin contrast to the process of customer-centric output.To realize high individuality and/or high complexityin the output the seller offers some leeway in thespecification which is not restricted to pre-defined featuresand values. At the beginning of the specificationprocess the customer can configure the output withinthe pre-defined options of a customer-centric output.If the customer is not content within the offered rangeof features and values, the process will be extended.The extended configuration process provides the functionalityto compose an individual model of output byadding and deleting features from a repository, addingindividual features and creating individual values.In summary we can state that seller-driven output andcustomer-oriented output are successfully transferredto business models in Electronic Commerce. The theoreticalbase is available which principally consists ofknowledge in doing traditional commerce and the enrichmentin doing business with information technology.Questions can be found in the implementation ofthe customer-oriented output in Electronic Commerce.At this point we see a range of research fields starting atthe organization of virtual supply chain and value chainrelationships up to the specification interfaces for output.A serious question in our point of view concernsthis specification interface in Electronic Commerce.4. Specification of customer-oriented output inelectronic commerceCustomer’s integration has a special significance inElectronic Commerce as human interaction needs to bereproduced comparable to traditional brick and mortarsales using information technology. In case of sellerdrivenand customer-centric output,this will not imposea problem, as the options are set fixed by the seller andcustomer’s choices are constricted.In this context the seller can use his knowledge aboutoutput, features and values to directly control the retrievaland specification process of the customer in advance.Examples can be found at ordering processeswithin online bookstores or computer configurations atonline-distributors.To offer customer-driven output, an additional focuson customer integration is required that can be achievedby individual determination of features and/or valuesthemselves. At this point, the specification processcannot be controlled entirely by the seller anymore.The customer needs to be offered more possibilitiesto influence design within his or her model of output.In literature several tools to realize the customer’s designin the specification (using information technology)have been proposed [6,8,13–15,17,25]. They depict

168 C. Scheer et al. / Business models to offer customized output in electronic commercesetCustomOption 2setOption 1customized model ofthe outputsetOption 1 getOptions 1CustomerConfiguratorgetOptions 1generic model of theoutputgetCustomOptions 31standard configuration process information flow2extended configuration process3suggested consultation interface in the extended configuration processFig. 5. Extended configuration process of customer-driven output.an extended configuration process. On a closer lookfrom customer’s perspective, it seems, that customersare offered powerful tools to specify the model of output.On the other hand we ascertain that there is noassisting guidance to precede or to go along with thedesign process. In a logical sequence there should bea consultation process which offers probable optionsbefore entering the design process. This can avoid allthose specification cases where the customer wants toquit the process because he does not find what he wantsin a complex configuration process. To face this problem,we propose an consultation interface within theextended configuration process.Figure 5 clarifies the cohesions. Within the traditionalconfiguration process for customer-centric output,customers are first offered a set of pre-defined options(features and/or values) (getOptions) and are thenasked for their choice (setOption). Main task of theconfigurator is to create a customized model of outputwith reference to options in the generic model. Thespecification process for customer-driven output differsin such a way (see Section 3), that additionally tools areoffered, which allow for more influence on the designprocess (setCustomOption). From our point of view,there needs to be an additional instance for advice (get-CustomOptions) to ensure, that the customer is able tocomplete the configuration successfully. The task ofthe configurator here is to create a customized modelwith the possibility of additional leeway in specificationwith reference to constraints in the generic model.The consultation interface in the extended configurationprocess for specification of customer driven outputshall therefore be described in the following. Existingapproaches in the field of configurators only offer additionallystatic informational resources in order to assistthe customer in the complete specification process. Itwould be more promising though, if the helpdesk of theinformation system would be modeled on the archetypeof the human interaction partner, so that the system getsenabled to give actual individual user support in eachstep of the configuration process.

C. Scheer et al. / Business models to offer customized output in electronic commerce 1695. Common approaches for consultation interfacein configuration tasksIn order to realize individual user support within theextended configuration process, we will present a consultationinterface, which creates a proposition for anydesired option value based on the individual configurationgoals and past configurations. The objective hereis to complete the configuration process in terms of thecustomer and to keep the customer in the process.At first it is fundamental to understand customer’sgoals of configuration. Therefore a measure needs to bedefined characterizing the preferences of the user. Onthe base of the measure, preferences can be analysedand used for user individual support. Basically fourapproaches are to be considered within this context (cf.Fig. 6).The horizontal dimension in Fig. 6 describes how tocollect information about the customer. A customermodel concerning configuration goals, preferences andresources can either be built by gathering informationvia direct inquiries – the user provides information directlyby himself – or via passive observance basedon behavior analysis. The knowledge about the customerthen needs to be matched with know-how frompast configurations (so called experience model) andthe product model in order to be able to propose likelyoption values. Know-how about past configurationscan either be in the form of the specific past configurationruns themselves (memory-based experience) orin the form of an abstract model (model-based experience)based on the collected data. As a result, thereexist a variety of approaches, each with different attributesand therefore also different suitability for usewith configurations for customer-driven output.Customer active data collection: A registration processis used to query the customer about his preferencesand goals, so that the collected information canafterwards be used as basis for individual support.– Model-based experience (Method 1): The predictionis based on the assignment of the customer to apre-defined user class. The classification dependson information provided by the customer [27]. Theproblem of initially not having any information toapply is avoided, as the predictions are already setbefore the actual configuration starts.– Memory-based experience (Method 2): The providedinformation is compared to informationabout other users which were collected in the past.The profile matching the configuration goal of thecurrent customer best is searched. In a secondstep, the corresponding configuration run containingall the option values is used for predictions.Therefore it is necessary to associate completedconfigurations with the information provided atthe registration process.Customer passive data collection: User’s behaviorsis assessed based on information provided by the userat the time he or she requests additional help. Henceit causes problems if the user asks for support at thevery beginning of the configuration process, since fewinformation about his or her configuration goals areavailable.– Model-based experience (Method 3): A modelis created based on past configuration runs andis compared to the behavior of the user, who requesteda prediction. Different approaches basedon different data structures can be applied: aweighted tree structure could be used [12] as wellas a model consisting of strict association rules [5].– Memory-based experience (Method 4): The behaviorof the customer is compared to individualpast configuration runs of other users. Predictionsare based on those option values of the past configurationrun matching best the information providedby the user so far [7].None of the presented approaches can solve the problemsimposed by configurations for customer-drivenoutput by it one: If active information retrieval is chosen,the option values are already specified before theactual configuration even starts. Thereby the predictionengine works completely independent from decisionswithin the configuration process. With passiveinformation gathering it is not possible to create anylikely predictions in the first steps of the process, sinceno information about the user is known yet. The fourpossible methods can be estimated as follows:– Method 1: The probability of creating a likely predictionwhich matches user’s configuration goalsis very little if clustering methods are used becauseof the large number of possible option valuesin cases of customer-driven Output. As allinformation was supplied by the customer himselfin advance, there is apparently the need to drawconclusions for every single option value based onthat higher level information provided within theregistration. This again cannot be possible, as theconcept of customer driven output asks for a highdegree of user involvement. The gathered infor-

170 C. Scheer et al. / Business models to offer customized output in electronic commerceModelStrict Association RulesSupport/ ConfidenceDatabase#010 => #14540/100Run 1 {#010,#026,#044,#145,#260...}#031 => #13640/100Run 2 {#007,#031,#075,#136,#203...}Algorithm#075 => #20340/100Run 3 {#010,#034,#075,#145,#203...}Run 4 {#012,#031,#066,#136,#207...}#136 => #03140/100Run 5 {#015,#037,#056,#126,#203...}#145 => #01040/100#203 => #07550/66.6Fig. 6. Simplified depiction of model creation.mation based on general information necessarilycannot provide that much insight into the preferencesregarding the choice between certain optionvalues.– Method 2: Comparable to the previously regardedapproach, this method is also better used for configurationswith a lesser degree of user involvement.Here, detailed option values need to be derivedfrom general information, which in practicecannot succeed especially if the amount of optionvalues is large. Additionally it is difficult to setpriorities among past configuration runs, whosequeries fully match the provided information, butcame to different configuration results.– Method 3: Likely predictions can be created comparingthe provided information within the processso far to a model based on past configuration runs.Four conditions have to be met: (1) The databasethat the model is based on needs to have a certainminimal size, (2) the user needs to have completedthe first steps within the configuration process inorder to have provided some information, (3) theuser needs to act in a way which at least partiallyresembles the behavior that users in past configurationruns have shown and (4) a suitable algorithm,which is able to extract cohesions amongoption values, has to be used in order to create themodel [26].– Method 4: Theoretically the comparison with pastindividual configuration runs is a good approachto create likely predictions. In practice thoughpure memory based collaborative filtering and thecorresponding methods of optimization [30] by reducingthe size of the database are not adequate.The necessary search time in larger databases willserve as an additional reason of process abruptionitself when using the configurator. Methods aimingat reducing the database size then again cannotguarantee the quality of predictions when usedwith highly user individual configuration settings.Because of the mentioned pros and cons of each approachwe want to combine several methods in orderto create a support mechanism which is able to predictaccurately and dynamically regardless of the point intime, when the user requests a prediction in the configurationstep.6. Model-based consultation interface usingassociation rules in the extended configurationprocessOur approach consists of a model (see model-basedexperience in Section 5) which is based on transferringthe procedure to determine association rules tothe product configuration process. Additionally it usesclustering methods to solve the resulting problems. Bydoing so, our model can predict far more precisely on asmaller database than it would be possible using linearcorrelations to similar configuration runs. Predictionsare based on the cohesions extracted from all configurations,not just the exactly corresponding ones. Especiallyin configuration settings offering a large selectionof options it is still possible to make an accurate predic-

C. Scheer et al. / Business models to offer customized output in electronic commerce 171tion even though there has not been a single previousconfiguration run, which shows a strong correlation tocurrent user’s configuration goals.Association rules have the purpose of identifyingformerly unknown strong relations between objectswithin a database [2]. First, we want to equate transactionswith completed past configuration runs, whichare stored in a database. This enables to adapt thesupport/confidence framework for this purpose. Thedatabase shall be composed of completed configurationruns which each consist of several parameter values ofthe options.Thresholds are introduced to distinguish betweenrelevant and irrelevant relationships of objects in thedatabase. Therefore we need values to help us selectamong all existing relations:– Support-Thresholds describe the absolute frequencywith which two object values appear togetherwithin one stored configuration run measuredabove all configuration runs in the database.– Confidence-Thresholds measure the absolute frequencyof one object appearing in configurationruns in which also the other object appears. It ispossible, that objects exist, that appear in aboutevery relation. Therefore several relations includingthis one object get selected when exclusivelyconsidering the support threshold, although thereis only very little informational value within thoserelations. As a result there is the need to include asecond filtering method, which also takes into considerationhow relevant the information is withinthe given context to filter out relations that wereonly picked because of general commonness of atleast one of the objects.We need to consider the necessity of strictly constrainingthe results as we are only interested in theone best matching association rule for the specific predictiontask, and not a considerate amount of relationssuch as in basket case analysis. Also the query for predictionis highly specified – the option value range isknown–, which improves the efficiency of the search,as none of the rules that do not contain any value withinthe option value range need to be considered.The creation of the model based on past configurationruns which are now stored within the database isthe initial step. An algorithm is applied to the datato find common object relations. In the following theApriori Algorithm shall be used for that purpose, otherderivations such as Apriori TID or distributed algorithmssuch as the Count Distribution (CD), can be seenanalogous [3,4].The Apriori Algorithm [4] is an iterative procedure,which generates single-dimensional, single-level,boolean association rules. Apriori aims at filtering objectrelations that satisfy the pre-defined constraints inform of threshold values with the help of iterations ofsubsets. The algorithm starts out with the examinationof the commonness of itemsets with length 1, andthen continues based on the results item sets with moreelements.The algorithm can be characterized as follows:beginM1 ={ itemsets with length 1}for (k =2;Mk− 1! = {}; k ++) do// Calculation of support valuesbeginCk = apriori-gen( Mk-1 );//Candidate Generationfor all transaction t in DbeginCt = subset(Ck,t) ;// Candidates contained in tfor all Candidates c in Ct doc.supportcount= c.supportcount +1;endMk = {c in Ck I c.supportcount >=threshold support}endendThe absolute frequency of each item in the databasewith length 1 is calculated within the first iteration.Based on the object quantities with length K, potentiallycommon itemsets, so called candidates, of lengthK +1are generated and usually get stored in hash tablesfor algorithm optimization purposes. Then the frequencyof each candidate is calculated and comparedwith the thresholds: If the value is below the threshold,the candidate is sorted out. In order to constraintthe amount of candidates, information generated in theprevious iterations is included in the candidate generationprocess: the apriori attribute is used. This refers toa monotony characteristic, that every subset of a commonitemset, which is not empty, has to be common,too.In reverse, patterns of a defined length, whose supportis below the threshold, don’t have to be taken intofurther consideration, as no itemsets can be generated,which possibly could be common. Every K − 1 subsetof a common itemset with length K therefore has to becommon. The algorithm terminates when no furthercommon itemsets can be generated anymore.

172 C. Scheer et al. / Business models to offer customized output in electronic commerceTable 1Data collection in customer model and utilization for predictionHow to get the experience model? How to get the customer’s model?Customer active Customer passivedata collection data collectionModel-based experience Method 1 Method 2Memory-based experience Method 3 Method 4The overall procedure of generating strict associationrules is as follows:– At first candidates are generated by combiningobject combinations with length K − 1, as long asthey differ only in one element.– In a second step all candidates are checked againstthe confidence threshold, as not all candidates necessarilyneed to be strict association rules, thoughall strict association rules can be found among thecandidates in reverse. All candidates, whose confidencevalue transcends the threshold are acceptedas strict association rules in the form A ⇒ B andare included in the model.The result of this procedure is an model of experience,which consists of strict association rules takeninto account all past configuration runs within thedatabase (cf. Fig. 7).7. Consultation process in detailWhenever the customer needs additional help in formof a prediction, a pattern consisting of three steps basedon our previously developed model starts:1) Based on the query of the user, all associationrules within the model are picked, in which exactlyone object is within the range of the optionvalues to be predicted.2) Each relation then needs to satisfy certain requirements:The object in the picked relation, which isnot within the range of the option values a) maynot be zero and b) has to be previously picked bythe customer in the configuration process.3) At this point, there are precisely three possibilities:a) Exactly one relation is found. In this case nofurther evaluation is necessary and the resultcan be used as a basis for prediction.b) Several or even contradicting relations arefound. So it has be decided by a selectioncriterion such as the confidence value, whichrelation is most likely to appeal to the configurationgoals and preferences of the customer.As a result the relation with the highestconfidence value is chosen.c) No relation can be found. All relations havebeen filtered out in the previous steps, whicheither was caused by overly strict definitionsof the thresholds when creating the model ora small-sized database. The foremost mentionedproblem is less a technical than a strategicalproblem: The quality of the predictionsis dependent on the amount of strict associationrules in the model. The system administratortherefore can lower the thresholds to enforceprediction generation, while the qualityanalogously gets reduced. The limit concerningtechnical realization is a certain maximalcapacity of the database, which still allows forprompt response times.Still we face difficulties to make predictions when thecustomer requests support at the very beginning of theconfiguration procedure, when no or little informationabout his goals were previously gathered.As a solution we propose the usage of clusteringtechniques in combination with the previously developedmodel in order to be able to immediately generatepredictions. Precondition for this is the existenceof a registration process in which the customer revealshis top-level preferences,demographic information andvague design ideas about the attributes he wants hisproduct to have. Based on this information the systemis able to classify the user into a pre-defined category,whereas the amount of categories is dependant fromthe desired degree of detailing.For every user group there is not only one optionvalue pre-defined rather than a ranking which optionvalues correspond more closely to the specific user category.This is done in order to combine our previously developedmodel with the clustering technique to furtherimprove our prediction quality.We introduce a correlation coefficient which measuresthe degree of compliance between the value chosenvia the association framework and the in the regis-

C. Scheer et al. / Business models to offer customized output in electronic commerce 173Information provided inregistration processQueryResultsCorrelation* ConfidenceResultOption AOption Value #2Option BOption Value #6ClusteringCustomerClassificationGroup #16Correlationnumbers#136#1450,65 * 1000,45 * 10065,0045,00Information provided inconfiguration runQueryResultsOption AOption Value #12AR ModelOption BOption Value #34Strict Association RulesSupport/ ConfidenceOption COption Value #102#10 => #145#31 => #13650/10050/100Option DPredictionrequestedQuery:? => #130 - #165#75 => #203#136 => #03150/10050/100#145 => #01050/100#203 => #07550/66.6Prediction with higher value is chosenFig. 7. Model displaying the integration of AR und clustering techniques.tration process mentioned configuration goal. The finalselection therefore is no longer dependent exclusivelyon the confidence value of the relation, but on the productof correlation number and confidence value; if nomatching association rule could be found, the predictionis based solely on the advice out of the clusteringroutine. This newly calculated product is a compositethat tells about how likely the via association rule generatedvalue goes along with the user stated configurationgoals. This value therefore unites primary and secondaryinformation retrieval methods. By this it is possible,to intercept wrong predictions which otherwisewill appear especially in unusual configurations. Theexplicitly stated goals of the user featuring top levelinformation gets combined with detailed collaborativebased information about the most likely option values,to form a value, which combines best of both worlds:All possible prediction values derived via the associationrule mechanism get multiplicated with a number,stating how likely this prediction value is consideringthe user stated goals. The higher the total number, themore likely is that the value is an accurate prediction.When a prediction is requested, a query is issued.The model is searched for appropriate association rules,which include one of the previously selected optionvalues on the left side of the association rule as wellas an option value within the specified range on theright side of the rule. If the query was succesful, eachof the results get matched with a correlation numbermanually stored in the system for each customer groupdescribing how likely the results will match customer’sconfiguration goals (cf. Fig. 8).As stated before, customer-driven output does notrequire the existence of any pre-defined options, sothat it might be impossible to state an exact correlationnumber for the values of some options in advance. Inthese cases, a classification based on basic attributescan help to identify cohesions among the option val-

174 C. Scheer et al. / Business models to offer customized output in electronic commerceues. Possible methods for realization include patternmatching for design recognition tasks or RGB-ratiosfor characterization of color [28,29].The next step in our research activity involves theimplementation of the suggested model in the configurationprocess. This will help us to approve mentionedcomponents and their coactions in the consultation interface.8. SummaryStarting from the basics of user customizable productsand services in Electronic Commerce and relatedbusiness models, the paper describes the extended configurationprocess for specification of customer-drivenoutput. In addition to the approaches in literature, anew support component is suggested, that offers individualadvice within the specification process. Thiscomponent includes a prediction mechanism, which isable to generate likely predictions based on previousconfiguration runs. Thereby we use the strength of objectvalue relations as an indicator for the relevance ofan option value within that context. We combine thisinformation with the information provided by the userhimself in order to unify the explicitly stated configurationgoal with our prediction. As a result, we areable to make likely predictions which are less dependanton both the size of the database and the point intime, when the prediction is requested by the customerin the configuration step.References[1] C.B. Adair and B.A. Murray, Breakthrough Process Redesign:New Pathways to Customer Value, AMACOM, New York,1994.[2] R. Agrawal, T. Imielinski and A. Swami, Mining AssociationRules between Sets of Items in Large Databases, in: Proceedingsof the International Conference on Management of Data,Washington D.C., 1993.[3] R. Agrawal and J. Shafer, Parallel Mining of AssociationRules, IEEE Transactions On Knowledge And Data Engineering8(6) (1996), 962–969.[4] R. Agrawal and R. Srikant, Fast Algorithms for Mining AssociationRules, in: Proceedings of the 20th International Conferenceon Very Large Data Bases (VLDB), Kaufmann, M.,1994.[5] Y. Chien and E. George, A Bayesian Model for CollaborativeFiltering, in: http://uncertainty99.microsoft.com/chien.htm(10.11.2002).[6] F. Feldkamp, M. Heinrich and K.D. Meyer-Gramann, SyDeR- System design for reusability, Artificial Intelligence for EngineeringDesign, Analysis and Manufacturing 12(4) (1998),373–382.[7] H. Inakoshi, S. Okamato, Y. Ohta and N. Yugami, EffectiveDecision Support for Product Configuration by Using CBR,in: The Fourth International Conference on Case-Based Reasoning,Vancouver/ Canada, 2001.[8] U. John, Konfiguration und Rekonfiguration mittelsConstraint-basierter Modellierung, Ph.D. Dissertation,Akademische Verlagsgesellschaft Aka, Berlin, 2002.[9] J. Lampel and H. Mintzberg, Customizing Customization,Sloan Management Review 38(1) (1996), 21–30.[10] K.J. Lancaster, Consumer Demand: A new Approach,Columbia University Press, New York, 1971.[11] K.J. Lancaster, Modern Consumer Theory, Edward Elgar,Hants, Vermont, 1991.[12] W. Lin, S. Alvarez and C. Ruiz, Efficient Adaptive-SupportAssociation Rule Mining for Recommender Systems, in:http://citeseer.nj.nec.com/483133.html (22.08.2002).[13] T. Männistö, H. Peltonen and R. Sulonen, View to ProductConfiguration Knowledge Modelling and Evolution, in: AAAIFall Symposium Series Configuration (Technical Report FS-96-03), Cambrigde, Massachusetts, 1996.[14] D.L. McGuinness and J.R. Wright, Conceptual Modelling forConfiguration: A Description Logic-based Approach, ArtificialIntelligence for Engineering Design, Analysis and Manufacturing12(4) (1998), 333–344.[15] S. Mesihovic and J. Malmqvist, Product Data Management(PDM) System Support for the Engineering ConfigurationProcess–A Position Paper–, in: Proceedings of 14th EuropeanConference on Artificial Intelligence (ECAI)–WorkshopConfiguration, Berlin, 2000.[16] H. Österle, Enterprise in the Information Age, in: BusinessNetworking: Shaping Enterprise Relationships on the Internet,H. sterle, E. Fleisch and R. Alt, eds, Springer, Berlin etal., 2000, pp. 17–54.[17] H. Peltonen, T. Männistö, K. Alho and R. Sulonen, ProductConfigurations–An Application for Prototype Object Approach,in: Proceedings of the Eighth European Conferenceon Object-Oriented Programming (ECOOP), Bolgona, Italy,1994.[18] A. Picot, Ein neuer Ansatz zur Gestaltung der Leistungstiefe,Zeitschrift für betriebswirtschaftliche Forschung (zfbf) 43(4)(1991), 336–357.[19] F.T. Piller, Mass Customization: ein wettbewerbsstrategischesKonzept im Informationszeitalter, Ph.D. Dissertation, Vol. 2,DUV/Gabler, Wiesbaden, 2001.[20] B.J.I. Pine, Mass Customization–The New Frontier in BusinessCompetition, Harvard Business School Press, Boston,Massachusetts, 1993.[21] B.J.I. Pine, B. Victor and A.C. Boynton, Making Mass Customizationwork, Harvard Business Review 71(5) (1993), 108–119.[22] J. Rayport, The Truth about Internet Business Models,in: http://www.strategy-business.com/pdf/099301.pdf(07.06.01).[23] R. Reichwald and B. Dietel, Produktionswirtschaft, in: Industriebetriebslehre:Entscheidungen im Industriebetrieb, E.Heinen, ed., Gabler, Wiesbaden, 1991, pp. 395–622.[24] C. Scheer and P. Loos, Concepts of Customer Orientation–Internet Business Model for Customer-driven Output, in: Proceedingsof the 10th European Conference on InformationSystems (ECIS), Information Systems and the Future of theDigital Economy, Gdansk, Poland, 2002.[25] C. Scheer and P. Loos, Kundenschnittstelle zur Spezifikationkundenorientierter Leistungen im Internetvertrieb– FachlicheAnforderungen und informationstechnische Implikationen, in:

C. Scheer et al. / Business models to offer customized output in electronic commerce 175E-Business–Integration industrieller ERP-Architekturen, P.Loos and N. Gronau, eds, Cuvillier, Göttingen, 2002, pp. 137–151.[26] L. Terveen and W. Hill, Beyond Recommender Systems:Helping People Help Each Other, in: HCI In The New Millennium,J. Carroll, ed., Addison Wesley, Reading, MA, 2001.[27] L. Ungar and D. Foster, Clustering Methods For CollaborativeFiltering, in: Proceedings of the Workshop on RecommendationSystems, Menlo Park, California, AAAI Press, 1998.[28] R. Veltkamp and M. Haagedorn, State-of-the-art in ShapeMatching, Technical Report, Utrecht University, Utrecht,1999.[29] A.K. Wright, Pattern Matching for Scheme, Report, Rice University,Houston, 1996.[30] K. Yu, X. Xu, J. Tao, M. Ester and H.-P. Kriegel, Instance SelectionTechniques for Memory-Based Collaborative Filtering,in: Second Siam International Conference on Data Mining,Arlington/Virginia, 2002.