All That Glitters is Not
Gold: Digging Beneath                                                                    Anthony ...
374                             Anthony Danna and Oscar H. Gandy, Jr.

observe at other levels of society. Developing a   ...
All That Glitters is Not Gold                                   375

data (1998, p. 44). Decision tree models are         ...
376                                Anthony Danna and Oscar H. Gandy, Jr.

Collaborative filtering systems. A new customer,...
All That Glitters is Not Gold                                  377

firm’s share of its customers’ business rather than   ...
378                                Anthony Danna and Oscar H. Gandy, Jr.

content a customer receives may appear to be    ...
All That Glitters is Not Gold                                        379

sible to parse out inaccurate data. Third, there...
380                              Anthony Danna and Oscar H. Gandy, Jr.

the separation of consumers into two or more      ...
All That Glitters is Not Gold                                   381

sophisticated form of this economic phenom-          ...
382                              Anthony Danna and Oscar H. Gandy, Jr.

population that could not otherwise afford to     ...
All That Glitters is Not Gold                                   383

supplied with second-rate goods and services. We     ...
384                             Anthony Danna and Oscar H. Gandy, Jr.

commitment to what we have come to recognize       ...
All That Glitters is Not Gold                                      385

being mined to generate information about cus-    ...
386                                Anthony Danna and Oscar H. Gandy, Jr.

   Mining’, Communications of the ACM 43(8),    ...
Upcoming SlideShare
Loading in …5
×

All That Glitters is Not Gold: Digging Beneath the Surface of ...

906 views
828 views

Published on

0 Comments
0 Likes
Statistics
Notes
  • Be the first to comment

  • Be the first to like this

No Downloads
Views
Total views
906
On SlideShare
0
From Embeds
0
Number of Embeds
1
Actions
Shares
0
Downloads
6
Comments
0
Likes
0
Embeds 0
No embeds

No notes for slide

All That Glitters is Not Gold: Digging Beneath the Surface of ...

  1. 1. All That Glitters is Not Gold: Digging Beneath Anthony Danna the Surface of Data Mining Oscar H. Gandy, Jr. ABSTRACT. This article develops a more compre- sphere. We suggest more ethically sensitive alterna- hensive understanding of data mining by examining tives to the unfettered use of data mining. the application of this technology in the marketplace. In addition to exploring the technological issues that KEY WORDS: analytics, customer relationship man- arise from the use of these applications, we address agement, data mining, marketing discrimination, per- some of the social concerns that are too often ignored. sonalization, price discrimination, privacy, profiles, As more firms shift more of their business activi- public sphere ties to the Web, increasingly more information about consumers and potential customers is being captured in Web server logs. Sophisticated analytic and data Introduction mining software tools enable firms to use the data contained in these logs to develop and implement a This article examines issues related to social complex relationship management strategy. Although policy that arise as the result of convergent devel- this new trend in marketing strategy is based on the old idea of relating to customers as individuals, opments in e-business technology and corporate customer relationship management actually rests on marketing strategies. As more firms shift many segmenting consumers into groups based on profiles of their business activities to the World Wide developed through a firm’s data mining activities. Web (the Web), increasingly more information Individuals whose profiles suggest that they are likely about consumers and potential customers is to provide a high lifetime value to the firm are served being captured in Web server logs. Sophisticated content that will vary from that which is served to analytic and data mining software tools enable consumers with less attractive profiles. firms to use the data contained in these logs to Social costs may be imposed on society when develop and implement a complex relationship objectively rational business decisions involving data management strategy. Although this new trend mining and consumer profiles are made. The ensuing in marketing practice is based on the old idea discussion examines the ways in which data mining of relating to customers as individuals, customer and the use of consumer profiles may exclude classes of consumers from full participation in the market- relationship management actually rests on seg- place, and may limit their access to information essen- menting consumers into groups based on profiles tial to their full participation as citizens in the public developed through a firm’s data mining activities. Individuals whose profiles suggest that they are likely to provide a high lifetime value to the firm Anthony Danna is a recent graduate of the masters program will be provided opportunities that will differ at the Annenberg School for Communication at the from those that are offered to consumers with University of Pennsylvania. Oscar H. Gandy, Jr. is the Herbert I. Schiller Term less attractive profiles. Professor at the Annenberg School. He is the author of Although there are some observers who invite The Panoptic Sort, Communication and Race, and a careful assessment of the costs and benefits that an engagement with the ethics of identification published data mining represents for the corporation, only in the Notre Dame Journal of Law, Ethics & Public very limited attention is being paid to the dis- Policy. tribution of costs and benefits that we might Journal of Business Ethics 40: 373–386, 2002. © 2002 Kluwer Academic Publishers. Printed in the Netherlands.
  2. 2. 374 Anthony Danna and Oscar H. Gandy, Jr. observe at other levels of society. Developing a that includes a user’s Internet connection speed, more comprehensive understanding of the social software platform, and Internet service provider impact of this marketing technology is the focus (ISP) address. Data from these off-line and online of this article. touch points can also be mined with third-party We begin with an examination of the ways in demographic and psychographic databases3 when which data mining technologies are applied in aggregated into a data warehouse. Data mining the market to support corporate marketing strate- algorithms can be run in these warehouses to gies. Technological and application-related issues discover hidden patterns and trends that are used are taken up before we introduce a discussion of to create consumer profiles. the social concerns that emanate from the Data mining is increasingly being seen as an application these technologies in the public essential business process. Firms awash in data are and private sectors. We conclude with several desperately trying to capitalize on it. Over half recommendations for mitigating the negative of Fortune’s top 1000 companies planned on social impacts of data mining. using data mining technologies in 2001 to help determine their marketing strategy, a substantial increase since 1999 when under a quarter of How personalization programs work these firms used data mining as a knowledge dis- covery technique (LeBeau, 2000). The following Data mining technology is employed in a variety describes the primary methods of data mining of analytic and customer relationship manage- used by firms for knowledge discovery. ment programs that are sold directly to firms or offered through an application service provider. Neural networks and decision trees. Artificial neural In their promotional literature, the software com- networks are designed to model human brain panies that sell these programs emphasize the functioning through the use of mathematics. In need that both business-to-business and business- order to be applied as a data mining technique, to-consumer enterprises have to build better and neural network processing elements must first be more profitable relationships with their customers trained to discover patterns and relationships by in a customer-centric economy. using a sample of data. The network is tested against a second set of data to validate the predictive model it has generated. How well the Analytics network performs in predicting values in the validation set is used as an indicator of how well Analytic software allows marketers to comb the network will predict outcomes with new through data collected from multiple customer data. Because neural network technology has the touch points to find patterns that can be used to capability of learning, it does not require inten- segment their customer base. Call center, product sive programming instructions to sort through registration, and point-of-sale transaction gener- data. Neural networks have been employed in ated data are typical of the off-line touch point communication research to predict television data used in this type of analysis. The Web is extreme viewers and nonviewers (Paik and another touch point that creates vast amounts of Marzban, 1995) and in the financial services data that firms are including in their mining industry to develop credit-scoring criteria and to activities. Web-generated data includes informa- predict bankruptcy. tion collected from forms, transactions, as well Like neural networks, data mining through the as from clickstream records. Clickstream data1 use of decision tree algorithms discerns patterns allows for path analysis, shopping cart analysis,2 in the data without being directed. According the analysis of entry and exit points, and the to Linoff, “decision trees work like a game of analysis of search terms or key words entered by 20 Questions,” by automatically segmenting data a visitor. Through the use of cookies, firms can into groups based on the model generated when add technographic information to their database the algorithms were run on a sample of the
  3. 3. All That Glitters is Not Gold 375 data (1998, p. 44). Decision tree models are Collaborative filtering systems typically use infor- commonly used to segment customers into mation collected in a registration process for “statistically significant” groups that are used as analysis. Observational personalization systems a point of reference to make predictions (Vaneko analyze clickstream data and dynamically serve and Russo, 1999). personalized content based on that analysis (Mulvenna et al., 2000). This data is then fed to Market basket analysis and clustering. Both neural a recommendation engine where it is compared networks and decision trees require that one to profiles of previous visitors in order to provide knows where to look in the data for patterns, as the current user with content that is predicted to a sample of data is used as a training device. The match that user’s preferences. Users can be use of market basket analysis and clustering tech- anonymous or identifiable in this process. The niques does not require any knowledge about following scenarios illustrate how personalization relationships in the data; knowledge is discovered systems that rely on analytics and data mining when these techniques are applied to the data. may be used in marketing applications. Market basket analysis tools sift through data to let retailers know what products are being pur- Manual decision rule systems. Bob and Alice are chased together. Clustering tools group records both customers of Main Street Bank. Like most together based on similarity/dissimilarity scores financial services companies, Main Street Bank is applied to each data point in the individual interested in reducing its churn rate. By applying record. Linoff notes that “clustering is typically its information about Bob to the predictive one of the first techniques applied; the segments models it generated through its data mining found in clusters often prove useful” (1998, activities, the bank is able to identify Bob as a p. 44). customer who has a high lifetime value ranking but is at risk of leaving the bank for another financial services provider. Alice, on the other Campaign management and personalization hand, fits the profile of a low lifetime value customer and is also assumed to be unlikely to Firms find the segments found in clusters prove leave the bank. to be most useful when they are integrated into Since the bank is concerned with reducing a marketing strategy. Campaign management churn, it offers Bob an interest bearing checking and personalization solutions incorporate or use account and a reduced loan rate as an enticement elements of analytic programs to allow marketers to stay; these offers are Main Street’s effort to to engage in a targeted one-to-one dialogue strengthen its relationship with Bob. In commu- with their customers. For the most part, Web nicating these changes to Bob, the bank empha- personalization occurs in one of three ways: sizes the value it places on having him as a manual decision rule systems, collaborative customer. Because Alice falls into a less lucrative filtering systems, and observational personaliza- category of customer, she is not served the same tion systems. These three categories are not offers as Bob. In fact, because Main Street Bank’s mutually exclusive as programs can combine profile of Alice suggests that she is unlikely to elements of each. Manual decision rule systems leave the bank, Alice is served with a notice that serve personalized content based on static user the fee she is assessed for using a teller inside the profiles. Static user profiles contain information bank will be increasing. In communicating these collected about a consumer over the course of changes to Alice, the bank emphasizes the value that consumer’s relationship with the firm. The it places on having her as a customer and reminds profiles are static in that they are not altered her that she can continue to use the ATM at no as a result of the consumer’s Web activities. additional charge. Thus we see that Bob, who is Collaborative filtering systems serve personalized perceived to be more valuable to the bank is content based on an analysis of information rewarded with lower prices for the services he provided by the consumer via a Web interface. uses, while Alice’s fees are likely to rise.
  4. 4. 376 Anthony Danna and Oscar H. Gandy, Jr. Collaborative filtering systems. A new customer, actually wanted to buy. A bit frustrated, Carol Ted, visits Groove.com, an e-commerce site that leaves the site in search of the software she needs. sells CDs. The site’s homepage encourages new At about the same time, Steve entered the site customers to register before making a purchase and selected the same computer and printer. in exchange for a discount on shipping. Ted is Although he chose the same products as Carol, enticed by the shipping offer and chooses to Steve did not receive the same offer for dis- register with Groove.com before making his counted educational software. He entered the site purchase. He completes an online registration from a different portal than that used by Carol; form where he is asked to rank his musical he had a different clickstream pattern from hers, preferences and rate different artists. After he and he used different terms in his keyword is finished registering he browses the site for search. Steve’s navigational pattern resulted in his music. As a registered user, a correlation engine being assigned to a different profile. Steve fit best that uses his registration data to predict what his into the “college student” profile and as a result, preferences might be determines the content Ted he was offered a discount on a statistical software is served as he navigates through the site. Ted’s package. In fact, Steve is an English major. Like profile is updated each time he views informa- Carol, Steve’s projected needs did not accurately tion about an artist, or downloads a sample. In match his real needs. addition, each time he makes a purchase he As these scenarios suggest, all systems will err provides the correlation engine with new data to some degree when they attempt to predict that can be used to narrow the range of options individual interests and needs. The consequences he will be presented with on his next visit. that flow from the accumulation of such errors are at the base of the concerns we will discuss Observational personalization systems. Carol is inter- below. ested in purchasing a new computer and she visits TechStation.com, an electronics e-tailer. Carol is a first-time visitor to this site. After entering Perceived benefits of personalization and a few keywords to search the site and after campaign management browsing through several of the pages she selects the model she is interested in. Carol adds a Customer relationship management printer to her virtual shopping cart and continues browsing. The observational personalization The software companies that market personal- system used by the electronics store compares her ization products that use data mining techniques point of entry to the site, the keywords she used for knowledge discovery, speak to their poten- in her initial search, her clickstream within the tial clients in a language that make the benefits corporate site, and the contents of her shopping of these systems unmistakable. The promotional cart to the navigational patterns of existing materials these firms use on their Web sites customers already in firm’s database. Through employ the language of customer relationship this comparison, the system fits Carol into the management. Market share and the development “young mother” profile that it developed by of market share by acquiring new customers was mining the Web navigation logs generated by once the primary driver of marketing strategy. previous visitors and existing customers. However, customer relationship management Accordingly, the recommendation engine offers appears to be the philosophy that will drive mar- Carol a discounted educational software package keting strategies in the 21st century. Customer before she checks out. relationship management focuses not on share of Carol was, in fact, not a young mother, but market, but on share of customer. Marketing a middle-aged divorcee. She purchased the strategists have been able to demonstrate that a computer and printer she was interested in, but firm’s profitability can increase substantially by did not find the time management software she focusing marketing resources on increasing a
  5. 5. All That Glitters is Not Gold 377 firm’s share of its customers’ business rather than by marketing analysts. Most programs come with increasing its number of customers (Peppers and pre-installed reporting tools that can be easily Rogers, 1993). customized. PrimeResponse summarizes this One of the basic tenets behind customer rela- theme best by noting that its product minimizes tionship management is the Pareto Principle, the “dependency on the IT department and puts notion that eighty percent of any firm’s profit is power back into the hands of your marketing derived from twenty percent of its customers. staff ” (PrimeResponse, 2001). Engaging in a dialogue with that twenty-percent in order to ascertain what their needs are and General to specific needs customization. Dialoging offering goods and services to meet those needs with customers is how the “loop” in the mar- are said to be what customer relationship man- keting process is closed. A firm starts with some agement is all about. Data mining technologies understanding of what its aggregate customers’ have allowed firms to discover and predict whom needs are. This is usually the result of market their most profitable customers will be by ana- research. The next stage in the loop involves lyzing customer information aggregated from developing a product or service to meet those previously disparate databases. The Web has needs. Communicating that product or service created a forum for firms to engage in a one-to- offering to customers follows that process. The one dialogue with particular segments of their loop is closed when a firm gets feedback from customer base in order to ascertain what the its customers and uses that feedback to refine its needs of those segments are. product or service offering. Feedback can come To better understand how segmenting for the in the form of a direct answer to a question con- purposes of customer relationship management cerning needs or it can be indirect and processed might be actualized, we analyzed the promotion in such a way that it is used to discover needs. materials posted on the Web sites of forty ana- Analytic and personalization software programs lytical, personalization, and e-commerce software allow firms to dialogue with their customers and vendors. The sample was generated from trade refine their product or service offerings in a one- press articles on customer relationship manage- to-one fashion. By using analytical and person- ment and from Hoover’s online directory. Our alization software, e-businesses can determine analysis of the software companies very hetero- what an individual customer’s needs are – based geneous Web sites revealed five recurring themes on the profile the customer fits into – and that appear in their promotional material. Using provide that customer with a product or service similar buzzwords and phrases, these themes that meets those needs. Peppers and Rogers speak to the essential elements of customer rela- (1993) call this type of feedback collaboration. In tionship management and the ways in which e- their view, the firm and its customers collaborate businesses can use Web to make the strategy work to meet each other’s needs. to their advantage. 360° view of customers. Integrating data from the Ease of use. The technology that makes analyt- Web and other customer touch points can give ical and personalization software work can firms a more holistic view of their customers. involve very complex processes including the use This information is used to segment the customer of algorithms, online analytical processing, and base into communities of customers with similar neural networks. This complexity can be intim- characteristics (Peppers and Rogers, 1997). The idating for, and beyond the area of expertise of more information a firm has about its individual staff in the marketing department who are ulti- customers, the easier it will be to create profiles mately responsible for implementing these tech- and place individuals into them. As discussed nologies. The software firms, for the most part, above, these customer profiles, the product of have designed their programs with this in mind. analytic programs, are essential tools in the per- Interfaces have been designed specifically for use sonalization process. Although the personalized
  6. 6. 378 Anthony Danna and Oscar H. Gandy, Jr. content a customer receives may appear to be Market conditions and competition unique, it is specific rather than unique because it is based on these profiles. None of the per- The competition between software companies sonalized content served to a customer on a Web that market analytic and personalization software site is truly unique to that individual. It is specific is fierce. In order to differentiate themselves in to the group that the customer is determined to the marketplace, companies that offer analytic belong to (Newell, 2000). programs have partnered with companies that offer personalization programs to provide firms Internet time analysis. Analytic and personalization with the most sophisticated technology available software allows firms to respond to their cus- to implement customer relationship management tomers in “Web-time.” The speed at which Web programs. However, a white paper published by log data can be analyzed and compared to stored the Patricia Seybold Group in 1999 predicted profiles to serve up personalized content gives that these best-of-breed partnerships would customers a seamless experience as they move dissolve in favor of integrated applications through a firm’s site. Speed is also an important (Harvey, 1999). That prediction, to some extent, variable in dialoging with customers. If a has come true (Gonsalves, 2001a; Gonsalves, customer does not respond to personalized 2001b). Software companies that specialize in content in the way predicted by the profile she Web analytics and personalization will likely is assigned to, new iterations of the calculations suffer the fate of net. run by the recommendation engine can be run As more bricks-and-mortar companies expand to instantaneously provide her with new content their Web presence and transform themselves (Greenberg, 2001). into bricks-and-clicks e-businesses software firms like SPSS that can supply a complete portfolio of Measuring return on investment. More software data mining, analytics, and Web analytics will companies emphasize measuring return on become market leaders. With the growth in investment (ROI) than any of the four themes knowledge discovery and personalization, pro- listed above. While ROI receives a lot of atten- fessional services firms have established consulting tion in the marketing of analytic and personal- practices to help firms capitalize on this new ization software, calculating the ROI for technology. e-business initiatives is complex and as a result there is little consensus as to how it should be done (Peppers and Rogers, 1997, pp. 384–387; Technological and application issues Greenberg, 2001, pp. 265–266). Regardless of how a firm chooses to calculate ROI however, a Like any new technology, the software programs customer’s lifetime value measurement is always discussed herein are not error proof. First, part of the equation (Newell, 2000, p. 58). applying data mining algorithms to data from According to Newell, a customer’s lifetime value disparate databases is not a simple task. Error is is “perfect for calculating ROI for CRM likely when data needs to be extracted from programs because everything aimed at strength- disparate databases, loaded into a data warehouse, ening the customer relationship has the objective and cleaned prior to being mined (Young, 2000). of increasing the customer’s profitability over This is especially the case when data from a Web time” (2000, p. 58). Regardless of how a firm traffic log is integrated with other data sources mines its databases to segment its customers, an (Tillett, 2000; Fink and Kobsa, 2000). Second, estimation of lifetime value is part of each profile if there is no way to separate good data from bad created. The idea that some customers are worth data, erroneous data that finds its way into a data more than others is the foundation for customer warehouse is just as likely to be subjected to relationship management. data mining algorithms as is more accurate data. There are myriad sources for error in data col- lection, and in many cases it is virtually impos-
  7. 7. All That Glitters is Not Gold 379 sible to parse out inaccurate data. Third, there is customer is possible, such a view can never be the case of missing data. Econometricians and complete. Transaction-generated data provides a statisticians have developed a sizable literature historical snap-shot. Its predictions of the future that addresses various methodologies for fitting are based on the past, and we know the past is models involving missing data; however, the often an unreliable guide to the future. Predictive more important questions in this area revolve models rarely take human serendipity into around the sources of missing data. Are partic- account, and it is virtually impossible to predict ular classes or populations of people more likely the circumstances that will shape an individual’s to have missing data associated with their records? choices in the future. If so, how do the statistical methodologies employed to fill in these missing data points affect the data mining-derived knowledge generated Social concerns about these groups? With substantial room for error, recommen- When presented to firms, the principles of dation engines will have difficulty coming up customer relationship management and the with recommendations that their targets will software tools that allow it to be implemented find appropriate. An error in data cleaning, for in an e-business setting appear to be a rational example, could easily alter the information stored means to a profitable end. Yet, there are always about a particular customer and ultimately affect some social costs that are imposed on society how that customer will be segmented. What when business decisions based on data mining a recommendation engine determines is the and consumer profiles are made. The ensuing dis- “right” content for a consumer will not neces- cussion will focus on the ways in which data sarily be considered “right” in the eyes of that mining activities and the use of consumer profiles consumer. Berry and Linoff note that predictions systematically excludes classes of individuals from based on data mining activities are nearly always full participation in the marketplace and the wrong “at the level of individual consumers” public sphere. (2000, p. 20). They argue that the benefits We suggest that price and marketing discrim- derived from the small percentage of predictions ination result from the profiles generated by data that are “right” outweigh the costs of not having mining practices. The sorting and allocation made any predictions at all. This small percentage of information and opportunity that are the can generate a notable increase in sales or click consequence of data mining can be thought of rates for a firm. When measured in an action- in terms of an invitation. For some categories able outcome like click-through or purchase of persons, price discrimination is an invitation dollars these predictions can then be, on average, to leave quietly. For many of these same persons, “right” for the firm. Using actionable outcomes “Weblining” and marketing discrimination as measurement tools, firms can gauge the appro- ensures that invitations are rarely if ever addressed priateness of a recommendation that is based on to them. As we will discuss, these outcomes raise a prediction. If a recommendation generates a fundamental concerns about fairness or distrib- sale for instance, it is right for the firm. If it does utive justice (Hausman and McPherson, 1996; not, it is wrong. This information is then added Hochschild, 1981; Roemer, 1996). to the consumer’s profile and is used to calculate new recommendations. Given the room for error in data, it is unlikely Price discrimination – an invitation for exclusion that a recommendation based on a prediction will precisely capture the needs of an individual Most economists would agree that price dis- consumer. Data stored in a record will never be crimination occurs when the same good or able to truly represent a complex autonomous service is sold to different consumers at different individual. Although the personalization software prices. In a classic work on the subject, Stigler companies claim that a 360° view of the (1966) observes that discrimination necessitates
  8. 8. 380 Anthony Danna and Oscar H. Gandy, Jr. the separation of consumers into two or more the most common practitioners of second-degree classes whose valuations of the good or service price discrimination. Although all seats on an differ. Stigler notes that this segmentation airliner move passengers from point A to point “requires the product sold to the various classes B, airlines version their fundamentally equivalent differ in time, place, or appearance to keep buyers products through fare restrictions. For example, from shifting” (1966, p. 209). In addition to price Southwest Airlines lists seven different fares,5 each sensitivity, the time, place, or appearance of a with different restrictions for travel between product will determine a consumer’s valuation of Baltimore and Chicago. Southwest has identified that product and assigns that consumer into a different fare classes and passengers self-select class. From our discussion of data mining prac- their fare class according to their valuation of the tices, it is easy to see how these practices can be restrictions. Incidentally, Southwest also practices used to classify consumers based on an estimated a form of third-degree price discrimination by or predicted valuation. Stigler defines price dis- offering children, infant, and senior citizen fares. crimination as “the sale of two or more similar This particular form of price discrimination does goods at prices which are in different ratios to not generally raise ethical concerns about fairness marginal cost” (1966, p. 209). Using this defin- because those who receive these travel discounts ition, either the price, the good or service, or a are presumed to have limited resources, and combination of both could vary across consumer might not otherwise be able to travel. classifications. Windowing in the film industry provides In order to sell similar goods or services at another example of second-degree price dis- prices that differ in their ratio to marginal cost crimination where market power, consumer firms must have, according to Varian (1989), sorting, and arbitrage all come into play (Owen some degree of market power, the ability to and Wildman, 1992). Windowing allows the pro- sort consumers, and the ability to prevent arbi- ducers of programming to sell different versions trage.4 Market power can come in the form of of the same product to consumers in order to monopoly or oligopoly, either within an industry extract maximum profit. Film distributors stagger or for a product or service itself. Sorting the release of a film through different channels consumers by valuation can happen in several (theater, video/DVD, pay-per-view, cable, broad- different ways. Sorting consumers into classes cast, syndication) in order to differentiate or through data mining is the most sophisticated version their product. A producer has market method of classifying consumers by their valua- power in that it has a monopoly on the market tion; however, price discrimination will generate for a particular film. The staggering of release marketplace disparities even when consumers through different channels invites consumers to self-select their class. Arbitrage is not an issue sort themselves into categories defined by the for service providers like banks, but it is a major distributors (theater patron, renter, etc.) based on issue for firms that sell information products and their ability and willingness to pay. By staggering other goods that can easily be resold. release and employing copyright protections, Economists have formally defined three types producers have the ability to prevent arbitrage. of price discrimination. The most common form It is clear in this example, as it is more generally is third-degree price discrimination. In cases of that sellers always benefit from price discrimina- third-degree price discrimination, firms exploit tion (Meurer, 2001, p. 91). The questions the differences in price sensitivity they have iden- becomes one of determining which, if any tified in the marketplace. Student or senior groups of consumers benefit. Rawlsian egalitar- citizen discounts realized by self-selection are ians might express some concern about this examples of third-degree price discrimination. particular form of price discrimination because In second-degree price discrimination, firms those with limited income are least likely to be are able to further exploit the differences in price provided any group-specific discounts (Baker, sensitivity and extract more surpluses from con- 2002). sumers by versioning their product. Airlines are First-degree price discrimination is the most
  9. 9. All That Glitters is Not Gold 381 sophisticated form of this economic phenom- ings. According to Rogers, the bank “has nudged enon in that it requires firms to perfectly exploit more than 60 percent of its customers [that were the differences in price sensitivity between paying on a fee-for-service basis] into flat-fee consumers. The seller charges the buyer the packages” because customers with flat-fee highest price the buyer is willing to pay for a packages tend to stay loyal to the bank (2001, good or a service. Shapiro and Varian note, p. 1). “it is awfully hard to determine what is the The Royal Bank of Canada was not concerned maximum price someone will pay for your about retaining the loyalty of the 40 percent of product or service” (1999, p. 39). Enter knowl- its fee-for-service customers it did not nudge. edge discovery through data mining. Suddenly This is an example of what Peppers and Rogers this awfully hard process becomes easier for (1997) call “firing” the customer; in this case the one-to-one marketers who have to make deci- bank made certain that a segment of its customer sions about the differential pricing of products base had a disincentive to stay. We would suggest and services in order to increase the firm’s share that the formal definition of price discrimination of the high lifetime value customers’ business. Stigler (1966) offers is broad enough to account If those customers who have a predicted high for the actions the Royal Bank of Canada took lifetime value are the ones a firm needs to keep, to “fire” its customers. In an e-business setting, then those with a predicted low lifetime value enticing certain customers to stay and others to are the ones a firm needs to get rid of or other- leave can be accomplished through personalized wise convert to a more profitable status. Many content. Personalized content can take many firms come to the conclusion that low margin forms, including that of the price tag associated customers are not worth the effort necessary to with a particular product or service. turn them into high margin customers. The We have seen how data mining lends itself to easiest thing to do is to entice those customers customer relationship management. With its to leave (Newell, 2000, p. 42). This is often emphasis on increasing the firm’s share of pre- achieved through price discrimination. dicted high lifetime value customers’ business, Peppers and Rogers (1997) have recom- customer relationship management necessarily mended placing customers into a three-tier lends itself to price discrimination. Price dis- hierarchy, based on a calculation of potential crimination and the segmenting of consumers for value: Most Valuable Customers, Most Growable the purposes of exclusion need not only take Customers, and Below-Zeros. According to place under the rubric of customer relationship Peppers and Rogers, Below-Zeros represent management. “the flip side of the Pareto Principle – the In the emerging market for digital informa- bottom 20 percent who yield 80 percent of tion products and services, price discrimination losses, headaches, collection calls, etc.” (1997, will become increasingly popular as technology p. 416). allows for versioning and differential pricing. In The financial services industry is skilled in their guidebook to survival in the networked the art of price discrimination. This skill is the economy, Shapiro and Varian (1999), emphasize result of data mining technologies that help to the need to version content and price in the segment customers (Peppard, 2000). Rogers production of information goods in order to (2001) describes efforts by the Royal Bank maximize profitability. Cohen (2000) contends of Canada to take a more customer-centric that price discrimination of this type would seri- approach to management by tiering its customer ously restrict access to high quality products, base in order to better channel communication especially for low-income consumers who would and services. First, the bank mined its databases be priced out of the market or have no choice and developed an algorithm to model the lifetime but to settle for products of lesser quality. value of its customers and to estimate the Whereas price discrimination in information “growability” of certain segments. With this data markets is often justified in terms of its potential in hand, the bank set out to differentiate its offer- for increasing access to the segments of the
  10. 10. 382 Anthony Danna and Oscar H. Gandy, Jr. population that could not otherwise afford to ferentiate services and their associated fees. While purchase information goods, the evidence seems Weblining can encompass price discrimination, to suggest that quite the opposite result occurs. it is a more general term used to describe cases Those with more substantial resources are where classes of consumers are excluded from actually provided discounts or subsidies in order the marketplace. Invitations are systematically deliver their attention to advertisers who value withheld. Like its bricks-and-mortar world them more (Baker, 1994, pp 66–69). counterpart, redlining, Weblining involves the categorical discrimination of groups based on characteristics of their neighborhoods rather Weblining and marketing discrimination – no than on information about specific individuals invitation at all (Hernandez et al., 2001). In the financial services industry, consumers Discriminatory pricing strategies are only one profiled above some risk criterion level are way firms can exclude certain classes of con- unlikely to learn about lending programs and sumers from the marketplace. Profiles can also be other credit offers. Lambert (1999) describes this used to determine if a class of goods and services process through his examination of the ways in that are offered to them in the first place. In an which traditional, direct marketing practices are e-commerce setting, it is commonplace for con- used to deliver credit and loan financing offers sumers to receive differential access to goods and to desirable borrowers. A decision about who services as the result of collaborative filtering receives information about lending programs is or observational personalization techniques. made in a context where risk “is no longer Consumers who fit into a particular profile may defined in terms of default, but as the failure not be offered certain goods as readily as those to be significantly profitable” (Lambert, 1999, who fit into other profiles. Discrimination in p. 2185). Although lenders cannot by law access and service based on a constructed profile prohibit consumers who do not receive direct has consequences for people in physical spaces solicitation from submitting an application, it is like neighborhoods as well as in administrated highly unlikely that many such applications spaces like Web sites. would be forthcoming. Take the case of Kozmo.com, the Internet- The fact that the law is most concerned about based home delivery service that closed its doors such discrimination only when the victims are in April 2001. Kozmo was accused of geograph- members of protected groups does not erase the ical redlining by residents in several of the cities fact that many consider the practices to be unfair in which it offered door-to-door delivery of because people are not being treated as individ- entertainment and food products. Although uals capable of making a rational choice in their it had distribution centers located in predomi- own interest. nantly African-American neighborhoods in both In the case of redlining, a particular irony Washington, D.C. and in New York City, Kozmo emerges that helps to make this point. An indi- did not make its services available to the residents vidual who is denied credit because of the neigh- of these neighborhoods. Its executives claimed borhood in which she lives may be unable to the company had simply made a rational business succeed with a claim of discrimination under decision and used neighborhood Internet usage civil rights law because she is not a member of as the basis for defining its service area. A more a protected group. She is a victim, not because extensive evaluation of Kosmo’s business practices of her race, but because of the race of the people suggested that there had been a pattern of that live in, and help to determine the profile of discrimination in each of the cities in which her neighborhood.6 Kozmo.com operated (Zaret and Meeks, 2000). The societal impact of unfair discrimination Marcia Stepanek (2000) of Business Week by commercial firms varies with the nature of the coined the term “Weblining” to describe how goods and services involved, as well as with the banks segment and rank their customers to dif- populations that are denied quality, and over-
  11. 11. All That Glitters is Not Gold 383 supplied with second-rate goods and services. We who are thereby informed, and it is quite likely have suggested that access to information may be to work against the interests of those consumers particularly troublesome in that it may limit con- who have been excluded from the flow of sumers’ ability to make informed choices in the information. marketplace, or to participate effectively within We believe the seriousness of this problem is the public sphere. amplified when it is the government that dis- criminates in the supply of information. In his State of the State address in January Information in the public sphere 2001, California Governor Gray Davis announced the launch of California’s new e-gov- The Internet has often been hailed as an ernment portal, the “My California Homepage.” enhancement to the democratic public sphere. BroadVision provides the site with personaliza- By its very nature, the Internet is thought to tion technology while Broadbase Software is pro- expand access to a broad range of voices and per- viding the analytical tools necessary to evaluate spectives (Sparks, 2001). However, when it is visitor data and build profiles. California is not used as a tool to segment and divide it can have alone in implementing technology designed for a deleterious impact on the democratic process. the private sector in the delivery of governmental Segmenting consumers for the purposes of deliv- services and information through the Web. ering policy-related information creates and Delaware’s Web portal incorporates content exacerbates inequalities that can distort public management solutions from Eprise Corporation discussion and debate (Gandy, 2001). If infor- (2001), who stresses the importance of using mation about public policies is more accessible profiles to customize the delivery of Web to one class of citizens than to others, then the content. Profiling and segmentation can result quality of the dialogue between and among in some content being made extremely difficult the governed will most certainly be affected. for the average user to find. Indeed, in some The Web offers opportunities to lower barriers cases, this information may have been placed “off to access and engagement; segmentation only limits” to particular classes of citizens. creates new barriers. For example, increased concern about moni- Sunstein (2001) makes the case that filtering toring and managing access to Web-based gov- on the Web erodes opportunities to have shared ernment information has developed following experiences and to lessen the likelihood that we the events of September 11, 2001. Limitations on will be exposed to viewpoints and information access to formerly public information are likely that we may not seek out on our own. These, to be based on characterizations of users devel- he argues are essential to meaningful deliberation oped through data mining techniques (How Sept. and to the functioning of a democratic society. 11 changed America, 2002; Gerstein, 2001). Although he briefly discusses collaborative fil- As funding levels for intelligence agencies in tering techniques, Sunstein’s argument rests the United States and abroad are increased in primarily on the assumption that filtering deci- support of a new war against terrorism, it seems sions are made by individuals out of their own likely that many of these dollars will flow self interest. However, when a firm uses data to business providing analytics software, data mining techniques to create the profiles that are mining, and data warehousing resources (Gomes, then used to serve filtered information the effects 2002). Developments made in response to gov- Sunstein describes become considerably more ernment contracts are likely to enhance the capa- problematic. A profit-seeking firm will filter the bility of data mining, data warehousing, and information it supplies in its own economic business analytics systems currently being devel- interest, especially where those interests are tied oped for the commercial market. In addition, as to the interests of advertisers and investors (Baker, the cost of existing systems drops, the use of these 2002). This filtering will not necessarily serve the technologies will spread beyond the Fortune 1000 private or collective interests of the individuals to commercial enterprises of varying size and
  12. 12. 384 Anthony Danna and Oscar H. Gandy, Jr. commitment to what we have come to recognize have been compelled to issue public apologies as fair information practices (OECD, 1980). As when the discriminatory nature of their routine the use of these technologies become more business practices have been revealed in the press. ubiquitous the likelihood of their affecting Others, like DoubleClick have seen their stock commercial and civic life in the ways we have fall out of favor with investors when discrimina- discussed will become even greater. Smaller firms tory practices have been brought to light. are more likely to fall under the radar screen of There are some signs that organizations whose watchdogs and are less likely to belong to the very life depends upon proprietary technology industry groups that have been assigned regula- for rating and ranking consumers recognize the tory duties in the place of government. importance of informing consumers about the ways in which their life chances are determined in the marketplace. Fair Isaac, the leader in credit Conclusions and recommendations scoring, has recently developed a commercial product that would inform consumers about the Decision makers should not discount the costs components of their credit scores, and how they imposed on society when data mining and might be improved (Simon, 2002). Of course, we consumer profiles are used to identify and are not suggesting that the problems of market segment individuals into groups on the basis of discrimination that we have described can be estimates of value. Although progress toward overcome by selling access to information about establishing a reasonable expectation of privacy the means by which such discrimination is in transaction-generated information has been accomplished. stalled in response to the terrorist threat, early We are recommending that every decision attempts to formulate a policy on consumer pro- maker who bears any responsibility for imple- filing that would win the support of the business menting a marketing strategy based on data community have emphasized the importance of mining consider more than its impact on the “notice and choice.” bottom line. We are especially hopeful that At the very least, consumers should be decision makers would consider the Rawlsian informed of the ways in which information principles of special regard for those who are least about them will be used to determine the oppor- advantaged, rather then being guided, as they tunities, prices, and levels of service they can seem most often to be, by a utilitarian calculus expect to enjoy in their future relations with a that is blind to distribution (Hausman and firm. We note that a reliable test of the ethical McPherson, 1996; Roemer, 1996). status of any business practice is the extent to A Rawlsian perspective on social justice, which it can be exposed to the light of public guided by the bright light of publicity, would go review. We might understand this test as an appli- a long way toward revealing the true value in cation of the Kantian standard of “universal the base metals that data mining has uncovered acceptability” or the “Golden Rule” which so far. admonishes us not to do anything to another than we would not have them do to us (Spinello, 1997, p. 37). Notes Our examination of discussions of data mining 1 and segmentation techniques within the trade Clickstream data represents the “footpath” a Web site visitor creates while navigating through a site. press reveals a broad awareness among the users Data points are generated when site visitors click of these techniques, that the public is concerned, through the site, following links. For the typical e- indeed often outraged when they discover the business, clickstream data can grow to several terabytes ways in which they are graded, sorted, and in size. To put the size of these databases into per- excluded from opportunities that others enjoy spective, it should be noted that two terabytes of data (Berry, 1999). Because they have ignored this would be roughly equivalent to the amount of data basic principal of mutual respect, many firms stored in an academic research library. In addition to
  13. 13. All That Glitters is Not Gold 385 being mined to generate information about cus- Politics: Communication in the Future of Democracy tomers, clickstream data is also analyzed to evaluate (Cambridge University Press, New York), pp. individual page traffic levels, gauge the effectiveness 141–159. of site design, and evaluate the popularity of content. Gerstein, J.: ‘Online Secrets. Internet Could Reveal 2 Shopping cart analysis uses clickstream data gener- Sensitive Information to Enemies’, ABCNews.com ated by e-commerce site visitors to investigate, among [On-line]. Available: http://abcnews.go.com/ other things, the point in the visitors’ footpaths where sectio. . . ngAmerica/WTC_011015_InternetSecrets purchases are made and to evaluate the point at which .html. shopping carts are abandoned by visitors. Gomes, L.: ‘Siebel Hopes Government Will Choose 3 Psychographic data includes information about its Software for the War on Terrorism’, Wall Street individual’s attitudes, behaviors, and beliefs. Journal Online [On-line]. Available: http://online. 4 Arbitrage is the term used to describe the resale of wsy.com/article_print/0,4287,SB101942665887847 a product in a market where price discrimination has 9920,00.htm. occurred. Arbitrage occurs when a consumer charged Gonsalves, A: 2001a, ‘Blue Martini Readies New a lower price for a product resells that product to a CRM, Commerce Apps’, TechWeb [On-line]. consumer who does not have access to the product Available: http://www.techweb.com/wire/story/ at that price. TWB20010209S0017. 5 See http://www.southwest.com. Accessed on Gonsalves, A: 2001b, ‘Personify Readies Customer 10/03/01. Analysis Product’, TechWeb [On-line]. Available: 6 Reference is made to the case of Cherry v Amoco http://www.techweb.com/wire/story/TWB20010 Oil Co., 490 F Supp 1026 (ND Ga 1980) in which 130S0018. a White woman who lived in a predominately Black Greenberg, P.: 2001, CRM at the Speed of Light: neighborhood was denied a gasoline credit card. The Capturing and Keeping Customers in Internet Real zip-code was one of several factors included in the Time (Osborne/McGraw-Hill, Berkeley, CA). rating system used by Amoco. Her zip code received Harvey, L.: 1999, ‘E.piphany E.4: Comprehensive the lowest of five ratings assigned by the company, but Advanced Customer Analysis Software for E- she was unable to demonstrate that racial animus was tailers’, Patricia Seybold Group Information Assets, the basis for the denial of credit. [On-line]. Available: http://www.epiphany.com. Hausman, D. and M. McPherson: 1996, Economic Analysis and Moral Philosophy (Cambridge References University Press, New York). Hernandez, G., K. Eddy and J. Muchmore: 2001: Baker, C. E.: 1994, Advertising and a Democratic Press ‘Insurance Weblining and Unfair Discrimination in (Princeton University Press, Princeton). Cyberspace’ SMU Law Review 54, 1953–1972. Baker, C. E.: 2002, Media, Markets, and Democracy Hochschild, J.: 1981, What’s Fair? (Harvard University (Cambridge University Press, New York). Press, Cambridge, MA). Berry, M. J. A and G. Linoff: 2000, Mastering Data ‘How Sept. 11 Changed America, and What it Costs Mining (Wiley, New York). our Liberty’, The Wall Street Journal, Online Berry, M.: 1999, ‘The Privacy Backlash’, Intelligent [On-line]. Available: http://online.wsj.com/ Enterprise (October 26), p. 20. article_print/0,4287,SB1015555457946917120,00. Cohen, J. E.: 2000, ‘Copyright and the Perfect htm. Curve’, Vanderbilt Law Review 56(6), 1799–1819. Lambert, T.: 1999, ‘Fair Marketing: Challenging Pre- Eprise Corporation: 2001, ‘Customizing Web Application Lending Practices’, Georgetown Law Information Delivery for Diverse Audiences’ [On- Journal 87, 2181–2224. line]. Available: http://www.eprise.com. LeBeau, C.: 2000, ‘Mountains to Mine’, American Fink, J. and A. Kobsa: 2000, ‘A Review and Analysis Demographics 22(8), 40. of Commercial User Modeling Servers for Linoff, G.: 1998, ‘Which Way to the Mine?’, AS/400 Personalization on the World Wide Web’, User Systems Management 26(1), 42–44. Modeling and User-Adapted Interaction 10, 209– Meurer, M.: 2001, ‘Copyright Law and Price 249. Discrimination’, Cordozo Law Review 23, 55– Gandy, O.: 2001, ‘Dividing Practices: Segmentation 148. and Targeting in the Emerging Public Sphere’, Mulvenna, M. D., S. S. Anand and A. G. Buchner: in W. Bennett and R. Entman (eds.), Mediated 2000, ‘Personalization on the Net Using Web
  14. 14. 386 Anthony Danna and Oscar H. Gandy, Jr. Mining’, Communications of the ACM 43(8), Wall Street Journal Online [On-line]. Available: 122–125. http://online.wsj.com/ar ticle_pr int/0,4287, Newell, F.: 2000, loyalty.com: Customer Relationship SB1016487907757224760,00.htm. Management in the New Era of Internet Marketing Sparks, C.: ‘The Internet and the Global Public (McGraw-Hill, New York). Sphere’, in W. Bennett and R. Entman (eds.), Organization for Economic Cooperation and Mediated Politics: Communication in the Future of Development: 1980, ‘Recommendations Concern- Democracy (Cambridge University Press, New ing and Guidelines Governing the Protection of York), pp. 75–95. Privacy and Transborder Flows of Personal Data’, Spinello, R.: 1997, Case Studies in Information and in M. Rotenberg (ed.), The Privacy Law Sourcebook Computer Ethics (Prentice Hall, NJ). 2000 (Electronic Privacy Information Center, Stepanek, M.: 2000, April 3, ‘Weblining’, Business Washington, DC), pp. 236–263. Week, pp. EB26–EB34. Owen, B. M. and S. S. Wildman: 1992, Video Stigler, G.J.: 1966, The Theory of Price (Macmillan, Economics (Harvard University Press, Cambridge, New York). MA). Sunstein, C. R.: 2001, Republic.com (Princeton Paik, H. and C. Marzban: 1995, ‘Predicting Television University Press, Princeton, NJ). Extreme Viewers and Nonviewers: A Neural Tillett, S.: 2001, ‘One-stop User Data Shop’, Network Analysis’, Human Communication Research InternetWeek, [On-line]. Available: http://www. 22(2), 284–306. internetwk.com/story/INW20010220S0003. Peppard, J.: 2000, ‘Customer Relationship Vaneko, J. J. and A. W. Russo: 1999, ‘Data Mining Management (CRM) in Financial Services’, and Modeling as a Marketing Activity’, Direct European Management Journal 18, 312–327. Marketing 62(5), 52–55. Peppers, D. and M. Rogers: 1993, Enterprise One to Varian, H. R.: 1989, ‘Price Discrimination’, in R. One: Tools for Competing in the Interactive Age Schmalensee and R. D. Willig (eds.), Handbook of (Currency, New York). Industrial Organization, Volume I (Elsevier Science Peppers, D and M. Rogers: 1997, The 1:1 Future: Publishing, Amsterdam), pp. 597–654. Building Relationships One Customer at a Time Young, D.: 2000, ‘The Information Store: CRM is (Currency, New York). Thrusting Data Mining Back to the Future’, PrimeResponse: 2001, ‘Solutions: E-tail & Retail’, Wireless Review 17(18), 42–50. [On-line]. Available: http://www.primeresponse. Zaret, E. and B. N. Meeks: 2000, April 4, com/solutions/retail.html. ‘Kozmo’s Digital Dividing Lines’, MSNBC Roemer, J.: 1996, Theories of Distributive Justice [On-line]. Available: http://www.msnbc.com/ (Harvard University Press, Cambridge, MA). news/373212.asp. Rogers, M.: 2001, ‘CRM Dividends for Royal Bank of Canada’, Inside 1to1, [On-line]. Available: http://www.marketing1to1.com. Annenberg School for Communication, Shapiro, C. and H. R. Varian: 1999, Information Rules: University of Pennsylvania, A Strategic Guide to the Network Economy (Harvard 3620 Walnut Street, Business School Press, Boston). Philadelphia, PA 19104-6220, Simon, R.: 2002, ‘Fair Isaac Plans Credit-score U.S.A. Help, But Watchdog Groups See Conflict’, The E-mail: ogandy@asc.upenn.edu

×