EVALUATION OF METHODS FOR CONVERTING REQUEST

FOR QUOTATION DATA INTO ORDINAL PREFERENCE DATA

Estimating Product Preference in Online Shopping System

Toshiyuki Ono

Systems Development Laboratory Hitachi, Ltd., 292 Yoshida-cho, Totsuka, Yokohama, Japan

Hirofumi Matsuo

Graduate School of Business Administration, Kobe University, 2-1 Rokkodai, Nada, Kobe,Japan

Norihisa Komoda

Graduate School of Information Sceince and Technology, Osaka University, 2-1 Yamada-Oka, Suita, Osaka,Japan

Keywords: Product Preference, Online shopping, Multiattribute utility function, Conjoint analysis

Abstract: Obtaining timely information on consumer preference is critical for the success of marketing and operations

management. In a previous paper we proposed a method of estimating consumer preference by using their

history of browsing among possible configurations of personal computer in an online shopping environment.

It consisted of three steps: (1) collecting data on each consumer’s browsing history regarding quotations and

purchase requests, (2) converting requests for quotations and purchase order data into ordinal preference

data, and (3) estimating consumer preference for product attributes by applying a multiattribute utility

function. The underlying assumption with this method was that a product configuration that was quoted

later would be preferred to those quoted earlier. Another assumption was that how many times a product

configuration was quoted would not affect estimates for product preference as long as this was quoted at

least once. Although these assumptions are critical in estimating consumer preference, their validity has not

been examined. In this paper, we evaluate the validity of such hypotheses regarding the relationships

between consumer preference and the sequence and frequency of quoted product configurations, and

propose six methods of estimating consumer preference. We show through experiments that, for about 60%

of examinees, all the proposed methods could approximate consumer preference obtained by conjoint

analysis, and that the six methods have almost equal accuracy. We therefore concluded that any of the six

methods could be used equally well for estimating consumer preference in a timely fashion.

1 INTRODUCTION

Obtaining timely information on consumer

preference is critical for the success of marketing

and operations management. Personal computer

(PC) manufactures carefully control their inventories

of components because their profit margins are

rapidly declining (Kurawarwala and Matsuo, 1996),

and the underage and overage costs of inventory

may become prohibitively expensive. To keep such

inventory related costs under control, these

companies need to track shifts of consumer

preference in a timely fashion.

Many methods of estimating consumer

preference have been developed. One category of

such methods is to estimate consumer behaviour

such as preference and price sensitivity by

constructing a marketing model based on buying

information collected by tools such as a POS (point

of sales) terminal (Andrews and Manrai, 1999;

Bucklin and Gupta, 1999; Cooper, 1993). Another is

conjoint analysis (Green et al., 2001; Lilien et al.,

1992), which is used to estimate the consumer

preference for each of product attributes. Input data

of conjoint analysis is collected by questionnaire that

Ono T., Matsuo H. and Komoda N. (2005).

EVALUATION OF METHODS FOR CONVERTING REQUEST FOR QUOTATION DATA INTO ORDINAL PREFERENCE DATA - Estimating Product

Preference in Online Shopping System.

In Proceedings of the Seventh International Conference on Enterprise Information Systems, pages 24-31

DOI: 10.5220/0002545200240031

 SciTePress

survey how consumers make trade-offs between

attributes collects.

Methods that use POS data, however, fail to

collect information on the preferences of the

shoppers who do not make purchases, while

methods that use elaborate questionnaires are

expensive and time-consuming even if questionnaire

survey can be conducted by online environment such

as Internet (Miller and Dickson, 2001). When we are

dealing with short-life-cycle products like PCs, we

need an economical and fast means of tracking shifts

in consumer needs in response to new product

launches, price changes, and competitors’ moves.

The widespread use of Internet has allowed

companies to offer product information and prices in

real time, and take purchase orders through online

shopping systems (Kalakota and Whinston, 1997).

Ono and Matsuo (2000) focused on the browsing

data of consumers who did not make purchases. The

proposed method used data on consumer browsing

history involving available product configurations in

an online shopping environment. The information

could easily and inexpensively be collected by the

seller, reflecting individual consumer preference as

well as her/his chosen set.

The proposed method consisted of three steps:

(1) collecting data on individual consumers’

browsing histories for quotations and purchase

requests, (2) converting requests for quotations and

purchase order data into ordinal preference data, and

(3) estimating consumer preference on product

attributes by applying a multiattribute utility

function (Green and Krieger, 1993).

The proposed method assumed that that a

product configuration quoted later would be

preferred to those quoted earlier. It also assumed that

how many times a product configuration was quoted

would not affect estimates of product preference as

long as it was quoted at least once. However, some

consumers might prefer a configuration that was

quoted earlier or one that was selected more

frequently. In this paper, we evaluate the validity of

several conceivable hypotheses regarding the

relationships between product preferences and the

sequences and frequency of quoted product

configurations. We are not concerned with the wider

issue affecting preferences such as choice of store,

quality of on-line store and prior experience

(Marsden et al., 1999).

The rest of this paper is organized as follows.

Section 2 describes the procedure for estimating

product preference. Section 3 describes proposed

methods of converting requests for quotation data

into ordinal preference data. Section 4 reports our

experimental evaluation of the proposed methods

through experiments. Section 5 concludes the paper.

2 ESTIMATING PRODUCT

PREFERENCE

2.1 Definition of Product Preference

The product attributes of the PC we consider in this

paper are its total price and the performance levels

of its components such as the CPU, random access

memory and hard disk drive. Product attributes have

various levels. The random access memory level, for

example, is measured by its storage capacity (e.g.,

256 and 512 MB).

When consumers purchase a product, they decide

whether some attributes are more important than

others, and what levels for all attributes are required

or preferred. A particular configuration that a

consumer purchases is regarded as the configuration

that has the largest total preference calculated from

the preference for each product attribute. The

product preference of each consumer can be

expressed as an additive model of the multiattribute

utility function (Green and Krieger, 1993):

）1（　　)(),,,(

∑

kkK

xuxxxU L

where x

is the level of attribute k, U(x

, x

,…,x

) is

the product preference for attribute levels equal to x

2, …,

and x

, and u

) is the preference for attribute

k at level x

As preference differs from one consumer to

another, the attribute preferences are estimated for

each consumer. When the preferences of all

attributes are estimated, we can estimate the

preference for a product that is represented by the

combination of attributes.

2.2 Procedure for Estimating Product

Preference

Ono and Matsuo’s method proposed (2000) consists

of three steps. This subsection describes each step in

detail.

(1) Collecting requests for quotations and

purchase order data

The consumer’s series of requests for quotations

and orders are collected using an online shopping

system. Figure 1 shows a web screen for such a

system. The consumer’s operational procedure using

online shopping system is as follows.

EVALUATION OF METHODS FOR CONVERTING REQUEST FOR QUOTATION DATA INTO ORDINAL

PREFERENCE DATA: Estimating Product Preference in Online Shopping System

Figure 1: Screen for online shopping service.

Table 1: Example consumer series of requests for

quotations and purchase order data.

# Consumer ID Time Request CPU Memory …

1 123.45.6.7 14:09:06 Quote 900 MHz 256 MB …

2 123.45.6.7 14:09:59 Quote 1.1 GHz 512 MB …

3 123.45.6.7 14:11:02 Quote 900 MHz 256 MB …

4 123.45.6.7 14:11:35 Quote 1.1 GHz 256 MB …

5 123.45.6.7 14:12:20 Quote 900 MHz 512 MB …

6 123.45.6.7 14:13:20 Order 900 MHz 512 MB …

(a) She/he selects her/his desired product

configuration from a menu of component

alternatives.

(b) The consumer sends a quotation request for

the selected product configuration.

selected product configuration.

(d) If the consumer decides to purchase the

product with the configuration, then she/he

sends her/his order to purchase it.

Otherwise, she/he quits or repeats (a) to (c).

The online-shopping server system collects and

stores a series of quotation requests (b) and purchase

order information (d), if any.

Table 1 has an example of a consumer’s series of

requests for quotes and purchase order. Here, she/he

first asks for a quote for a configuration comprised

of a 900-MHz CPU and 256-MB memory. Then,

she/he asks for a quote for a configuration

comprised of a 1.1-GHz CPU and 512-MB memory.

After five rounds of such requests for quotes, she/he

finally places a purchase order for a configuration

comprised of a 900-MHz CPU and 512-MB memory.

In this example, the consumer can be identified

by her/his source IP address to access the Internet.

We could also use a preliminarily registered

consumer ID or a “cookie” (Lemay et al., 1996),

which is Internet technology to read and write in a

file on the consumer’s PC.

(2) Converting requests for quotes and purchase

order data into ordinal preference data

Based on collected requests for quotes and

purchase order data, we can attempt to rank

configurations that have been requested for quotes

by each consumer in the order of preference. The

collected data contains purchasers’ data and non-

purchasers’ data. The product configuration that is

ordered to purchase should be ranked the highest in

terms of preference. Product configurations that are

not quoted should be ranked lower in the order of

preference. However, the question is how

configurations that are quoted and then not ordered

should be ranked in each consumer’s order of

preference. We propose several rules for converting

requests for quotation data into ordinal preference

data in Section 3, and evaluate these in Section 4.

(3) Estimating product preference

Three models for measuring a consumer’s

multiattribute utility function are described in Green

and Krieger (1993). The vector model assumes that

the preference is linearly related to product attribute

levels. The ideal-point model assumes that

preference is inversely related to the weighted

squared distance between the level of an attribute

and the individual’s ideal level of the attribute. The

part-worth model assumes a function representing

the discrete part-worth levels for each attribute.

Because the performance levels of components

such as CPUs and random access memories are

discrete, we use the following part-worth model:

）2（　　　

∑

mkl

where U

is a consumer’s product preference for

configuration m, d

mkl

is an indicator that takes on

value of one when attribute k is at level l for

configuration m, and zero otherwise, and λ

are

part-worth for attribute k at level l. The ordinal

preference data and the configurations can be

regarded as sample values of U

and d

mkl

in Eq. (2),

and the estimate of part-worth, λ

is calculated using

linear regression analysis.

ICEIS 2005 - SOFTWARE AGENTS AND INTERNET COMPUTING

3 PROPOSED METHODS OF

CONVERTING REQUESTS FOR

QUOTATION DATA INTO

ORDINAL PREFERENCE DATA

3.1 Hypotheses on relationships

between consumer preference and

sequences and frequency of

requests for quotations

In this subsection, we propose the following

hypotheses on the relationships between the

consumer preference for configurations and the

sequence and frequency of configurations that

appear in a consumer’s series of requests for

quotation, and propose methods of converting

requests for quotation data into ordinal preference

data. In Section 4, we attempt to verify the validity

of these hypotheses.

(1) Ranking the configurations based on the

sequence of quoted product configurations

Hypothesis 1a: Consumers prefer

configurations quoted later.

- Configurations quoted later rank higher in

the order of preference.

Hypothesis 1b: Consumers prefer

configurations quoted earlier.

- Configurations quoted earlier rank higher in

the order of preference.

Hypothesis 1c: Consumer preference is not

reflected in the order of the sequence of

requests for quotes.

- All configurations requested for quotes rank

in the same order of preference.

(2) Ranking configurations based on frequency of

appearance in sequence

Some consumers quote the same configuration

many times. We propose the following hypotheses

on the relationships between the ordinal preference

for configurations and the frequency of appearance

in the sequence.

Hypothesis 2a: Consumers prefer

configurations more frequently requested

for quotes.

- Configurations requested for more quotes

rank higher in the order of preference.

Hypothesis 2b: Consumer preference is not

reflected in the frequency of requests for

quotes.

- All the configurations requested for quotes

rank in the same order of preference.

3.2 Algorithm for deriving the ordinal

preference of configurations

We propose the following algorithm to assign a

positive integer to each configuration based on a

sequence of configurations where a consumer

requested a quotation. An integral number is

assigned to each configuration that reflects the order

of preference for the configuration. This integral

number for each configuration is referred to below

as the preference number. A preference number of

one is assigned to the most preferred configuration.

The smaller the preference number, the more

preferred the configuration is.

Steps 1-7 are applied to each consumer’s data.

Step 1: Set the preference number of n =1.

Step 2: If the collected data includes the

product configuration that the consumer

placed order to purchase, then a preference

number of n=1 is assigned to the

configuration that is ordered to purchase.

Then, set n: =n+1, and proceed to Step 5.

Otherwise, proceed to Step 3.

Step 3: Of the unprocessed request data in the

consumer’s series of requests for quotations,

(i) If Hypothesis 1a or 1c is applied, then

select the latest configuration in the

consumer’s series of requests for quotations.

(ii) If Hypothesis 1b is applied, then select

the earliest configuration in the consumer’s

series of requests for quotation.

Step 4: Consider one of the following three

cases for the selected configuration,

(i) If the configuration selected in Step 3 has

already been assigned a preference number,

then proceed to Step 5.

(ii) If the configuration selected in Step 3 has

not been assigned a preference number,

then preference number n is assigned to the

configuration

(ii-a) If Hypothesis 1a or 1b is applied,

then set n: =n+1 and proceed to Step 5.

(ii-b) If Hypothesis 1c is applied, then

proceed to Step 5.

Step 5: If been a configuration remains that has

not been processed in the sequence of

requests for quotations, then proceed to

Step 3. Otherwise, proceed to Step 6.

Step 6: There are two cases for Hypotheses 2a

and 2b.

(i) If Hypothesis 2a is applied, then do the

following.

Apply the following Steps 6.1 to 6.7 for

each configuration that has been

requested for quotation many times.

EVALUATION OF METHODS FOR CONVERTING REQUEST FOR QUOTATION DATA INTO ORDINAL

PREFERENCE DATA: Estimating Product Preference in Online Shopping System

Step 6.1: Select the configuration that has

been requested for quotation many times.

Step 6.2: Count the frequency of quotations

for the selected configuration.

Step 6.3: Set i: =1

Step 6.4:

(a) If the selected configuration has a

preference number of 1, then preference

numbers that have already been assigned

to other configurations should be

increased by one.

(b) If the selected configuration has a

preference number of 2 and the

configuration that has been ordered by the

consumer has a preference number of 1,

then the preference numbers that have

already been assigned to other

configurations should be increased by one.

number for the selected configuration

should be reduced by one.

Step 6.5: Count up i = i+1

Step 6.6: If i is equal to the frequency of

quotation for the selected configuration,

then proceed to Step 7. Otherwise

proceed to Step 6.4

(ii) If Hypothesis 2b is applied, then proceed

to Step 7.

Step 7: Calculate the maximum value for the

preference numbers assigned to the

configurations in Steps 1 to 6. Assign the

maximum value plus 1, as the preference

number to the product configurations that

do not appear in the sequence of requests

for quotations and purchase order.

We evaluate six combinations of Hypotheses 1a-

1c and 2a-2b and these are listed in Table 2.

Table 3 lists the preference numbers for the six

methods, which were calculated from data on the

sequence of requests for quotations and purchase

order in Table 1.

Table 2: Methods of deriving ordinal preference for

configurations.

Quote-sequence

hypothesis

Quote-frequency

hypothesis

Method 1 Hypothesis 1a Hypothesis 2a

Method 2 Hypothesis 1a Hypothesis 2b

Method 3 Hypothesis 1b Hypothesis 2a

Method 4 Hypothesis 1b Hypothesis 2b

Method 5 Hypothesis 1c Hypothesis 2a

Method 6 Hypothesis 1c Hypothesis 2b

3.3 Estimating Part-worth

Based on the preference numbers derived in Section

3.2, we estimate the part-worth coefficients in Eq.

(2). Figure 2 has an example of a part-worth

estimate. The Method 2 column in Table 3 has

sample values for product preference U

in Eq. (2),

and corresponding configurations are represented by

mkl

. The estimate for part-worth, λ

, can be

calculated by using linear regression analysis. Part-

worth, λ

, expresses a consumer’s preference for

product attributes. Product preference U

for

configuration m can be estimated by the sum of the

part-worths of corresponding product attributes.

Preferenc e

number

・・・

＝

mkl

Product configuration

900 MH

1.1 GH

256 MB 512 MB

・・・

0101

・・・

1010

・・・

0110

・・・

1001

・・・

Part- worth

11(900 MHz)

12 (1.1 GHz)

21 (256 MB)

22 (512 MB)

・・・

Quote 2 →

Quote 3 →

Quote 4 →

Order 6 →

Other →

Figure 2: Example part-worth estimate.

Table 3: Examples of ordinal preference data

Consumer’s data on quotation

and purchase order

Ordinal preference data by proposal methods

Request CPU Memory … Method 1 Method 2 Method 3 Method 4 Method 5 Method 6

Quote 1 900 MHz 256 MB … 2 2

Quote 2 1.1 GHz 512 MB … 4 4 4 3 3 2

Quote 3 900 MHz 256 MB … 2 3 2 2

Quote 4 1.1 GHz 256 MB … 2 2 5 4 3 2

Quote 5 900 MHz 512 MB …

Order 6 900 MHz 512 MB … 1 1 1 1 1 1

Rest of above configuration 5 5 6 5 4 3

ICEIS 2005 - SOFTWARE AGENTS AND INTERNET COMPUTING

4 EVALUATION OF PROPOSED

METHODS

4.1 Experimental procedure

We designed an experiment and tested it on subjects,

who were university students majoring in science

and engineering, to evaluate the methods proposed

in the previous sections. The product attributes and

levels that we used are listed in Table 4. The total

price of a product configuration is determined by the

sum of the component prices in Table 4. We

constructed the experimental system as a Web site

on the Internet to mimics an actual online shopping

system. Because it was an experimental system, we

did not sell products or collect the data on requests

for purchase.

To test the validity of the proposed methods, we

compared our experimental results with those of the

conjoint analysis. To do this, we conducted a

questionnaire using the same set of subjects. We

used orthogonal design, and asked each participant

to rank eight configurations in their order of

preference.

Table 4: Products attributes in experiment.

# Attribute Level No. Level Price

1 900 MHz +0 yen A CPU

2 1.1 GHz +20,000 yen

1 256 MB +0 yen B Memory

2 512 MB +23,000 yen

1 20 GB +0 yen

2 40 GB +10,000 yen

C HDD

3 60 GB +20,000 yen

1 CD-RW &

DVD-ROM

+0 yen D DVD drive

2 DVD-RAM &

DVD±R/RW

+20,000 yen

1 1-year warranty +0 yen E Warranty

2 3-year warranty +18,000 yen

Base price 190,000 yen

(Note: Profile nos. of all 48 (2*2*3*2*2) kinds of product

configurations were calculated by 24*(attribute A’s level No.-1)

+12*(attribute B’s level No.-1) +4*(attribute C’s level No.-1) +2*

(attribute D’s level No.-1) + (attribute E’s level No.-1)

4.2 Results and Discussions

4.2.1 Results for preference estimation

The quotation data collected from 66 subjects with

the experimental system described in Section 4.1 are

listed in Table 5. We converted the collected data

into ordinal preference data by using the six methods

described in Section 3 and used the data to estimate

part-worth λ

and product preference U

for any m.

Figure 3 is a graph indicating the part-worth λ

estimated with Method 2 for five subjects. The

estimated part-worth value was calculated from the

corresponding preference number described in

Section 3. By data conversion, the higher the part-

worth is, the more the subjects preferred that level

for product attribute. The average part-worth in the

same attributes for each consumer is set to zero.

Figure 3 shows that subject No. 1 preferred a 1.1-

GHz CPU and 512-MB memory and that subject No.

5 preferred a 900-MHz CPU and 512-MB memory.

4.2.2 Discussions on results

We evaluated the accuracy of estimated product

preferences U

with the questionnaires, which

subjects had indicated the configurations they

preferred. The data collected from these

questionnaires is listed in Table 6.

Table 5: Collected requests for quotation data.

Subject No. Product profile of quote sequence

1 46, 22, 34, 46, 42, 45, 46, 48, 46

2 44, 48, 36, 48, 47

3 10, 48, 24, 22, 2 , 24 ,23, 19

… …

66 29

(Note: Starting from left, product profile is quoted earlier)

-.60

-.40

-.20

.00

.20

.40

.60

.80

Part-worth

Subject No.1 Subject No.2 Subject No.3

Subject No.4 Subject No.5

Figure 3: Part-worth λ

of five subjects estimated with

proposed Method 2.

Table 6: Preference data collected from questionnaire.

Subject No. Preference order for 8 product profiles

(4, 7, 10, 16, 25, 28, 42, 47)

1 4, 7, 3, 8, 6, 5, 1, 2

2 7, 5, 4, 6, 3, 8, 2, 1

3 1, 8, 6, 4, 7, 5, 3, 2

… …

66 3, 8, 6, 4, 7, 5, 2, 1

EVALUATION OF METHODS FOR CONVERTING REQUEST FOR QUOTATION DATA INTO ORDINAL

PREFERENCE DATA: Estimating Product Preference in Online Shopping System

We calculated Peason’s correlation coefficient

between product preference estimated with conjoint

analysis and product preferences estimated with the

six proposed methods. The estimated product

preferences and correlation coefficients for subject

No. 1 are listed in Table 7. Figure 4 shows the

percentages of subjects for whom the correlation

coefficients between preference estimated with

conjoint analysis and with the six proposed methods

are statistically significant at the 5% and 1% levels.

Note that 33 of the 66 subjects only requested one

configuration for quotes. Therefore, these six

methods generated the same ordinal preference data,

and the same correlation coefficients results.

The number of subjects for whom there was a

significant correlation between the preferences

estimated with conjoint analysis and with the six

proposed methods was greatest for Methods 3 and 4.

For Method 3, the correlation coefficients of 42 of

the 66 subjects were significant at the 5% level and

the correlation coefficients of 39 subjects were

significant at the 1% level.

50.0% 52.0% 54.0% 56.0% 58.0% 60.0% 62.0% 64.0% 66.0%

Method1

Method2

Method3

Method4

Method5

Method6

5% significance level 1% significance level

Figure 4: Percentage of subjects for whom correlation

between preferences estimated with conjoint analysis and

proposed methods was significant at 1% and 5% levels.

In terms of ratio, the correlation coefficients for

63.6% of the subjects were significant at the 5%

level and the correlation coefficients for 59.1% of

subjects were significant at the 1% level. Even with

the method that had the lowest correlation, 60.6% of

the subjects had correlation that was significant at

the 5% level, and 54.5% of the subjects had

significant correlation at the 1% level. These results

led us to conclude that any of the proposed methods

could approximate preference obtained by conjoint

analysis for about 60% of the subjects.

Outcome of the hypothesis test for difference of

correlation coefficients among the proposed methods

are listed in Table 8. Numeric data “4” in the table

where the row is Method 2 and the column is

Method 4 means that H

is rejected in favour of H

the 5% significance level by “4” subjects from the

outcome of testing the hypotheses.

H0 (null hypothesis): population correlation

coefficients between Methods 2 and 4 are

equal

H1 (alternative hypothesis): population

correlation coefficient of Method 2 is

greater than that of Method 4.

All six methods are almost the same, and none

can be said to be better than the others.

Table 9 indicates the number of subjects who had

a high correlation between the outcome for conjoint

analysis and that for each of the six methods. Here,

the subjects are categorized by the number of

requests for quotes. We cannot see any relationships

between the number of requests for quotes and the

proposed methods. From the viewpoint of

computational effort, Method 6 (based on the

hypothesis that quote sequence and frequency are

not related to configuration preference) is the most

efficient because Steps 4 and 6 for the algorithm

described in Section 3.2 are not required to process

for each consumer data.

Table 7: Examples of estimated product preferences and correlation coefficients.

Product preference of subject No. 1

estimated with proposed methods

Product No.

Method 1 Method 2 Method 3 Method 4 Method 5 Method 6

Preference on

questionnaires

1 0.40 0.21 0.38 0.54 -0.33 -0.17 0.36

2 0.75 0.75 1.25 1.25 0.00 0.00 2.71

… … … … … … … …

48 3.33 2.92 3.00 2.58 0.90 0.48 8.75

Peason’s

correlation

0.761 0.77 0.722 0.721 0.741 0.764 -

Significance

probability of

correlation

0 0 0 0 0 0 -

ICEIS 2005 - SOFTWARE AGENTS AND INTERNET COMPUTING

Table 8: Outcome of the hypothesis test for difference of

correlation coefficients among proposed methods

Method

Method 1 - 0 4 4 1 2

Method 2 0 - 4 4 1 2

Method 3 3 3 - 0 0 2

Method 4 3 3 0 - 0 1

Method 5 0 0 3 3 - 1

Method 6 0 0 3 3 0 -

Table 9: Number of significantly correlated subjects.

Quote

frequency

Method

1 20 20 20 20 20 20

2 3 3 3 3 3 3

3 4 4 4 4 3 4

4 6 6 7 7 6 6

5 1 1 1 1 1 1

6 3 3 3 3 3 3

7 1 1 1 1 1 1

8 2 2 2 2 2 2

10 1 1 1 1 1 1

11 0 0 0 0 0 0

5 CONCLUSION

We proposed and examined hypotheses regarding

the relationship between preference for

configurations and the sequences and frequency of

requests for quotes. These hypotheses are critical in

estimating individual consumer product preference

from their browsing data. Ono and Matsuo (2000)

assumed Hypotheses 1a and 2b. However, the

validity of their assumptions over others that are

conceivable has not examined, and it was the focus

of this paper.

The experimental results indicated that, for about

60% of subjects, any of the six proposed methods

was able to approximate explicit preference obtained

by conjoint analysis. Therefore, any of these

methods could be used to track shifts in consumer

preference in a timely fashion.

The six methods had almost equal accuracy.

From the viewpoint of computational effort, the

method based on the hypothesis that there is no

relationship between configuration preference and

the quote sequences and frequency was the most

efficient.

Our evaluation only used experimental data on

requests for quotes, and thus did not include

information on purchase order. Evaluation based on

data that includes the order information from an

actual online shopping system is an issue for further

investigation.

The proposed methods can not only be applied to

PCs but also other products consisting of several

components. As long as we can collect data on the

browsing history of individual consumers, the

proposed methods may also be able to be applied to

shopping systems that use multimedia kiosks in

stores and on the street.

REFERENCES

Andrews, R. L. & Manrai, A. K., 1999. MDS Maps for

Product Attributes and Market Response: An

Application to Scanner Panel Data. Marketing Science,

18(4), 584-604.

Bucklin, R. E. & Gupta, S., 1999. Commercial Use of

UPC Scanner Data: Industry and Academic

Perspectives. Marketing Science, 18(3), 247-273.

Cooper, L. G., 1993. Market-Share Models. In Eliashberg,

J. & Lilien, G. L. (Eds.), Marketing Handbooks in

Operations Research and Management Science Vol. 5

(pp. 299-309). Elsevier Science Publishers B. V.

Green, P. E. & Krieger, A. M., 1993. Conjoint Analysis

with Product-Positioning Applications. In Eliashberg,

J. & Lilien, G. L. (Eds.) Marketing Handbooks in

Operations Research and Management Science Vol. 5

(pp. 469-477). Elsevier Science Publishers B. V.

Green, P. E., Krieger, A. M. & Wind, Y., 2001. Thirty

Years of Conjoint Analysis: Reflections and Prospects.

Interfaces, 31(3), 56-73.

Kalakota, R. & Whinston, A., 1997. Electronic Commerce

(pp. 217-243). Addison Wesley,

Kurawarwala, A. A. & Matsuo, H., 1996. Forecasting and

Inventory Management of Short Life Cycle Products.

Operations Research, 44(1), 131-150.

Lemay, L., Murphy, B. K. & Smith, E. T., 1996. Creating

Commercial Web Pages (p. 269). Sams.net Publishing.

Lilien, G. L., Kotler, P. & Moorthy, K., 1992. Marketing

Models (pp. 241-244). Prentice Hall.

Marsden, J. R., Tung, Y. A. & Keeney, R. L., 1999. The

Value of Internet Commerce to the Customer.

Management Scinece, 45(4), 533-542.

Miller, T. W. & Dickson, P. R., 2001. On-line Market

Research. International Journal of Electronic

Commerce, 5(3), 139-168.

Ono, T. & Matsuo, H., 2000. A Method for Analyzing

Customer’s Product Preference Based on Browsing

Data Collected in an Online Shopping System. The

Institute of Electrical Engineers of Japan, 120-C (8/9),

1230-1235 (in Japanese).

EVALUATION OF METHODS FOR CONVERTING REQUEST FOR QUOTATION DATA INTO ORDINAL

PREFERENCE DATA: Estimating Product Preference in Online Shopping System