客製化產品推薦模式之研究: 有限混合與層級貝氏邏輯斯模式之比較

(1)

行政院國家科學委員會專題研究計畫成果報告

客製化產品推薦模式之研究: 有限混合與層級貝氏邏輯斯

模式之比較

研究成果報告(精簡版)

計畫類別：個別型計畫編號： NSC 95-2416-H-002-049- 執行期間： 95 年 08 月 01 日至 96 年 07 月 31 日執行單位：國立臺灣大學國際企業學系暨研究所計畫主持人：任立中計畫參與人員：博士班研究生-兼任助理：邵功新報告附件：出席國際會議研究心得報告及發表論文處理方式：本計畫可公開查詢

中華民國 96 年 12 月 21 日

(2)

A Study of Customized Product Recommendation Models: Comparison of

Finite Mixture and Hierarchical Bayes Logit Models

ABSCRACT

The purpose of this article is to provide a set of solutions for customized new product

recommendation to improve the performance of CRM (Customer Relationship Management)

project. We proposed two customized new product recommendation models: finite

mixture Probit and hierarchical Bayes Probit model. The proposed methods are tested by

random selected customers in a home electronic retailer’s CRM database, and results show

that the presented customized new product recommendation models perform well. The

approach of this paper has its strength to be able to do new product recommendation and

cross-selling in database marketing on a one to one basis.

Keyword: Customer Relationship Management; New Product Recommendation; Finite

(3)

When you have a now product to sell, do you have trouble in identifying the customers

who might have higher propensity to buy this product in your CRM (customer relationship

management) database? A good new product recommendation system provides high-value

of service to customers, enhances cross-selling in database marketing, and generates higher

profits for this company. With the CRM database, company now can better understand

customer needs, can recommend right product for their customers, and can integrate

knowledge into their product design and marketing plans. However, the goal is often

difficult to achieve. Nearly half of U.S. implementations and more than 80 percent of

European implementations of CRM project are considered failures (Patron 2002). To

getting value back out of CRM, a good solution for customized new product

recommendation can not be ignored.

Current Solutions for Product Recommendation

Current solutions for product recommendation can be briefly categorized as two types

of filtering. The first is collaborative filtering which is based on the similarity between

customers’ rating. Despite it has been intensively used by internet retailers such as

Amazon.com, it still has limitations in practice. In order for collaborative

recommendation to be accurate, a large number of transaction data of a given product must

be prepared. Accordingly, it is completely limited when new products being encountered.

(4)

The other type is content-based filtering which match customer interest profiles with the

product attributes. In order for the approach to be effective, sufficiently rich product

information as well as personal preference profiles should be available. Accordingly, this

approach is limited when new customers are encountered.

Some scholars suggest to using hybrid models to overcome the limitations (Ariely,

Lynch and Aparicio 2002; Balabanovic and Shoham 1997; Pazzani 1999). However, the

filtering approach still be criticized (Ansari, Essegaier, and Kohli 2000; Iacobucci, Arabie,

and Bodapati 2000). First, the proposed filtering techniques are typically based upon

customers’ ratings data, instead of actual purchases record. This might limits their

application in practice. Second, filtering techniques does not base upon statistical method

so that they are unable to reflect uncertainty in predictions. To overcome this problem,

Ansari, Essegaier, and Kohli (2000) proposed a hierarchical Bayes regression model for

internet movie recommendations. The hierarchical Bayes approach can be used to provide

recommendations when either new products or new customers are being encountered.

Although their study provides great insights, two reasons might challenge the managers’

intension to put their idea into CRM practice. First, their study was based upon consumer’s

preference self-rating data instead of actual transaction records. Can we apply their

framework to improve the performance of new product recommendation in CRM project?

(5)

retailers, they sell multi-category products. Accordingly, they are more concern about how

to do new product recommendations for mulit-catogory products? The purpose of this

article is to provide a solution designed for doing customized new product recommendations.

The analytical solution is based upon statistical choice model and derived from our practical

experience during the implementation of CRM project of a home electronic appliance

retailer.

The outline of this paper is as follows. In the next section, we introduce a CRM case

from a home appliance retailer. The steps of proposed solution are illustrated. In the final

section, we present the results of estimation and compare the success of hit rates of different

methods, and a conclusion was provided in the last section.

A CRM Case form a Home Appliances Retailer

We use the data from the CRM database of a home appliance retailer in Taiwan.

They have implemented their CRM system since Jan 01, 1999. Transaction data were

available from Jan 01, 1999 to Jan 01 2001. There are 23 products categories selected for

analysis including electronic television, VCD, DVD, digital studio, air conditioning, et al.

(see figure 1). A random sample of 400 customers was selected for analysis (up to 1153

transaction records). The data of each customer includes their transaction data and

demographic information (see table 1).

Figure 1 here

(6)

Table 1 (a) here

Table 1 (b) here

Marketing scholars have encouraged the employ of choice model to improve the

analytical CRM project (Kamakura et al. 2005). However, there are several challenges for

researchers to apply choice model in CRM database. First, lacking of product attributes

coding in database. Second, the alternative choices were unavailable because only the real

purchased products were recorded in database. Third, methods to integrate data from

several kinds of sources were often depended on researchers’ expertise as well as experience.

In the following section, we develop a system way of analytical solution designed to

overcome these challenges and to improve the performance of new product

recommendation in CRM practice.

The Procedures and Logic for Proposed Models

The underlying assumption of proposed models is the features of these products can

be identified into common attributes. For instance, the common attributes for consumer

electronic products are country of origin, product design, function, price level, et al. In

that case, the procedures and logic for the proposed model are as follows: (1) we have

access to a retailer’ customer database that consists customers’ transaction records of

multi-category products and demographic information; (2) in each transaction, customer

(7)

products can be identified and coded; (4) individual’s preference toward each particular

attribute are estimated; (5) the utilities of each alternative products are predicted in order to

recommend new products to customers.

In order to test the ability to recommend new products, we separate customers’

transaction records into two groups: in sample and holdout sample. The hold out sample is

the last one purchase record of each customer. The others transaction records of each

customer were used as in sample records to estimate parameters. Our method follows the

steps: coding of product attribute, developing pseudo choice set, statistical modeling,

estimation, and utilities prediction.

Coding of Product Attributes

From an information cue theoretic perspective, products may be conceived as

consisting of an array of information cues, such as design, brand name, price, and country

of origin (Bilkey and Nes 1982). Each cue provides customers with a basis for evaluating

the product, which might influence customers’ purchase decision. In order to decompose

customers’ preference toward these attributes, we collect and integrate data from three kinds

of sources: product managers’ opinions, CRM database and magazine. Product attributes

coding includes five categories. They are specified as follows: (1) Country of origin

(country of manufacturing & country of brand): US, Japan, Taiwan, China & East and South

Asia countries, and the other countries. (2) Functioning: superiority, middle level and

(8)

weak in function. (3) Product design: good or not good in design. (4) Price ratio: a

continuous variable, which is calculated form the database by comparing the price of

purchased product and the average prices in same product category for a given time period

closed to the transaction date. (5) word-of-mouth: good or not in word-of-mouth. Data

was collected form a magazine in which there is an annual survey regarding consumers’

favorite brands in each product categories in year 2000.

Pseudo Choice Set

In order to simulate the available choices faced by consumers when they made their

purchase decision in retailer store, we create pseudo choice set. In this step, we includes

and integrated the other customers’ transaction records form 90 days earlier than and 30

days later than the actual purchase record of the target product made by the customer. The

mean of pseudo choice set is 10.83, and the standard deviation is 5.28.

Model

To compare the performance of hit rates of different methods, we provide two basic

solutions for comparison: random recommendation and Probit model. The others are two

types of customized new product recommendation models: Finite Mixture Probit Model and

hierarchical Bayesian Probit model. In a finite mixture Probit model, the individual

preferences of a given customer can be obtained by the weighted combination of probability

(9)

attribute weight for each individual can be estimated by pooling of information form both

individual and across populations. Details of model specifications are stated in appendix.

Results

Table 2 is the results of parameters estimated from Probit model, finite mixture Probit

model, and hierarchical Bayes Probit model. Some variable of country origin (i.e., country

of brand and country of manufacturing) were deleted due to either multicollinearity or too

few cases that might ruin the results of estimations. In the Probit model, we can see that

consumers have positive preference toward made in US. However, consumers have

negative attitude toward either country of brand in Japan or US. Comparing with middle

level in function, consumers have higher preference in either superiority or weak in

functioning. Besides, they have positive preference toward good design and

word-of-mouth and negative preference toward price ratio. The table 2 also shows the

estimated preference parameters of three finite mixture latent classes. In the first segment,

consumers have positive preference toward good design, price ratio and word-of-mouth.

People in this segment prefer good design and word-on-mouth products. They are more

willing to pay for relative higher price. People in the second segment are price

consciousness; they have negative preference toward price ratio and word-of-mouth.

Finally, people in the third segment are value consciousness; they have positive preference

toward function and word-of-mouth but negative preference toward price ratio. The last

(10)

two columns in table 2 are the results of hierarchical Bayes Probit model. The parameters

were the posterior mean and posterior standard deviation of individual beta parameters.

The posterior standard deviation of beta can represents the heterogeneity in consumers’

preference structure rather than represents the standard error in beta estimation. In

hierarchical Bayes Probit model, the individual’s demographic or behavior information can

be included to predict individualized parameters. Table 3 shows the results of these

coefficients estimated form hierarchical Bayes Probit model. For example, compared

with customers age below 20, customers age above 60 have negative preference toward

price ratio. Γ

Table 2 here

Table 3 here

After obtaining the estimated beta coefficients, we predict the utilities for each product

in pseudo choice set. Then, the set of utilities in each pseudo choice set were ranked form

high to low. The ranked number of actual purchased record can be used to represent the

numbers of recommended products that is required in order to hit the target product in each

choice set. As shows in table 4, the results were aggregated to test the predictive power of

different methods. In the in sample hit rates, the hierarchical Bayes Probit model out

perform finite mixture Probit model, and they both out perform Probit model and random

(11)

three recommendations of finite mixture Probit model is quite close to hierarchical Bayes

Probit model. The hierarchical Bayes Probit model still out performs the other models.

Table 4 here

Conclusion

The purpose of this article is to provide a solution designed for new product

recommendation. We propose two customized new product recommendation methods.

Both methods can help us to decompose consumers’ preference toward particular attribute,

and then help us to predict the purchase probabilities. The results of hit rates comparison

show that both of our proposed customized new product recommendation models perform

well in either in sample or hold out sample prediction. Thus, we suggest managers can

apply these customized new product recommendations when they want to improved their

performance in new product recommendation.

(12)

Reference

Allenby, M. Greg and James L. Ginter 1995. Using Extremes to Design Products and

Segment Markets, Journal of Marketing Research, 32 (November), 392–403.

Allenby, M. Greg, Neeraj Arora and James L. Ginter 1998. On the Heterogeneity of Demand,

Journal of Marketing Research, 35 (August), 384–89.

Allenby, M. Greg and Peter E. Rossi 1999. Marketing Models of Consumer Heterogeneity,

Journal of Econometrics, 89 (March/April), 57–78.

Ansari, Asim, Skander Essegaier and Rajeev Kohli 2000. Internet Recommendation Systems,

Journal of Marketing Research, 37 (August), 363-75.

Ariely, Dan, John G. Lynch Jr., and Manuel Aparicio IV 2004.., Learning by Collaborative

and Individual-Based Recommendation Agents, Journal of Consumer Psychology,

14(1&2), 81–94.

Balabanovic, Marko and Yoav Shoham 1997. Fab: Content-Based, Collaborative

Recommendation, Communications of the Association for Computing Machinery, 40

(3), 66-72.

Bilkey J. Warren and Erik Nes 1982. Country-of-Origin Effects on Product Evaluations,

Journal of International Business Studies, Vol. 13(1), 89-100.

Iacobucci, Dawn, Phipps Arabie, and Anand Bodapati 2000. Recommendation Agents on the

(13)

Kamakura, A. Wagner and Gary J. Russell 1989. A Probabilistic Choice Model for Market

Segmentation and Elasticity Structure, Journal of Marketing Research, 26 (November),

379-390.

Kamakura, A. Wagner, Carl F. Mela, Asim Ansari, Anand Bodapati, Pete Fader, Raghuram

Iyengar, Prasad Naik, Scott Neslin, Baohong Sun, Peter C. Verhoef, Michel Wedel and

Ron Wilcox 2005. Choice Models and Customer Relationship management, Working

Paper, Duck University.

Patron, Mark 2002. Editorial: If Database Marketing Was So Good, Why Is CRM So Bad?

Journal of Database Marketing, Vol. 10 (2), 102-103.

Pazzani, J. Michael 1999. A Framework for Collaborative, Content-Based and Demographic

Filtering, Artificial Intelligence Review, 13, 393-408.

Rossi, E. Peter and Greg M. Allenby 2003. Bayesian Statistics and Marketing, Marketing

Science, 22(3), 304–328.

(14)

200 131 120 125 95 87 84 203

Other Small Home Electronic Appliances Other Kitchen Appliances

Oven/Toaster Electronic Rice Cooker Washing Machine/Dryer Air Conditioning Appliances Digital Stereo Video/CD/DVD player

Frequency

(15)

14 ember ID Purchase Amount Quantity Purchase date Category Code Store

ID Brand Product Model 20009425 8500 1 19990704 12502 1 RCA 20 inches TV F21634TW 20009425 680 1 19990916 1303 1 ALIGN Oven OCE-8037 20009425 1980 1 20000729 1204 2 AIWA Digital Stereo XRAKH300 20012511 9500 1 20000816 22101 3 TECO Refrigerator RE-1601S 20015739 18000 1 20001027 22302 4 TECO Refrigerator RE6001N 20015739 13500 1 20001027 23103 11 SAMPO Washing machine ES-103SBF 20016493 19000 1 19991121 10401 8 AIWA Digital Stereo XRAKH100 20016493 2490 1 20000618 2101 7 HITACHI Vacuum Cleaner PV-C25 20016493 690 1 20000824 3301 7 PUMP Electric Iron TSK-750CS 20020521 1690 1 20000104 25101 10 SANYO DVD Player R-CF01T

…….. …….. …….. …….. …….. …… …….. …….. ……..

20940532 2988 1 20000131 10401 14 SANYO Digital Stereo DC-LU6

Table 1(a) is an example of data available in CRM transaction database.

Member ID

Card

Starting Date Gender Age

Post No. 20009425 10/14/2000 1 32 557 20012511 07/07/1997 1 38 820 20015739 11/18/2000 1 40 557 20016493 12/10/2000 1 27 241 20019338 09/16/1999 1 30 330 20002102 01/27/1990 2 31 830 20002383 07/28/2000 1 34 812 20013243 06/24/1994 2 64 704 20013421 01/24/1992 1 84 802 20015621 10/28/2000 1 32 356 …….. …….. …….. …….. …….. 20020532 08/09/1999 2 42 360

(16)

Probit Model Finite Mixture Probit Model Hierarchical Bayes Probit Model Segment 1 Segment 2 Segment 3

Beta Coefficient Standard Error Beta Coefficient Standard Error Beta Coefficient Standard Error Beta Coefficient Standard Error Posterior Mean of Beta Posterior Standard Deviation of Beta Constant - 1.47*** 0.07 - 2.56*** 0.48 - 0.54 1.27 - 1.12 0.13 - 1.31 1.54 Made in China - 0.06 0.05 - 0.07 0.25 - 0.01 0.61 - 0.08 0.09 - 0.11 0.44 Made in Japan - 0.10 0.08 - 0.15 0.47 - 1.17 1.33 - 0.15 0.14 - 0.19 0.55 Made in US 0.29*** 0.12 0.31 0.49 - 0.20 3.09 0.18 0.18 - 0.03 0.98 Brand in Japan - 0.16*** 0.06 - 0.10 0.34 0.57 0.80 - 0.15* 0.09 - 0.15 0.42 Brand in US - 0.32*** 0.09 - 0.44 0.34 - 0.18 1.07 - 0.25* 0.14 - 0.53 1.10 Superiority in Function 0.18*** 0.07 0.12 0.31 1.21 0.79 0.16* 0.09 0.20 0.47 Weak in Function 0.16*** 0.06 - 0.12 0.24 0.94 1.19 0.21*** 0.08 0.18 0.39 Good Design 0.17*** 0.05 0.45*** 0.18 0.33 0.63 0.03 0.07 0.13 0.40 Price Ratio - 0.15*** 0.05 0.80*** 0.25 - 1.62* 1.11 - 0.51*** 0.11 - 0.47 0.90 Word-of- Mouth _0.44*** _0.04 _0.36* _0.23 _{- 1.62* 0.93}_{0.60*** 0.07 0.48} _0.71

In sample is an unbalance panel that includes 400 individuals. The total number of observations is 8481 *p < .1, ** p < .05, *** p < .01

Made in Japan means product manufactured in Japan COO in Japan means the country of origin of brand is Japan

(17)

Constant Made in China Made in Japan Made in US COO in Japan COO in US Superiority in Function Weak in Function Superiority in Design Price Ratio Word-of- Mouth

Constant Posterior Mean 4.72 *** - 2.10 *** - 2.05 *** - 2.89 *** 0.26 - 2.65 *** 1.25 1.45 ** - 1.94 *** - 4.27 *** 0.28 Posterior STD (1.18) (0.87) (0.91) (1.08) (0.89) (1.03) (0.88) (0.83) (0.85) (0.96) (0.80) Gender Posterior Mean 0.02 - 0.26 0.03 0.59 - 0.08 - 0.17 - 0.29 - 0.02 0.10 0.12 0.07

Posterior STD (0.46) (0.32) (0.48) (0.63) (0.39) (0.43) (0.38) (0.34) (0.36) (0.44) (0.34) Age 21-30 Posterior Mean - 0.55 0.85 1.42 *** 3.24 *** - 0.95 0.04 - 0.59 - 0.18 0.93 - 0.12 - 0.71

Posterior STD (0.86) (0.69) (0.68) (0.73) (0.95) (1.14) (0.67) (0.68) (0.60) (0.73) (0.67) Age 31-40 Posterior Mean - 0.96 1.00 1.67 *** 2.76 *** - 1.00 1.75 ** - 0.59 - 0.20 0.74 0.03 - 0.08

Posterior STD (0.79) (0.70) (0.79) (0.80) (0.83) (1.01) (0.59) (0.61) (0.54) (0.74) (0.59) Age 41-50 Posterior Mean - 0.95 1.27 ** 1.67 *** 1.72 *** - 0.89 2.12 *** - 0.28 0.10 0.67 - 0.24 - 0.29

Posterior STD (0.78) (0.68) (0.66) (0.77) (0.89) (0.96) (0.60) (0.59) (0.58) (0.72) (0.58) Age 51-60 Posterior Mean - 1.00 1.00 1.81 *** 2.20 *** - 0.93 2.06 *** - 0.81 - 0.16 0.81 - 0.10 0.26

Posterior STD (0.84) (0.70) (0.75) (0.79) (0.85) (0.99) (0.64) (0.62) (0.56) (0.73) (0.58) Age above 60Posterior Mean - 0.59 1.40 *** 1.02 1.59 - 0.15 2.59 *** 0.01 0.12 0.26 - 1.16 ** 0.53

Posterior STD (0.93) (0.62) (0.92) (0.90) (0.96) (1.04) (0.72) (0.73) (0.66) (0.67) (0.69) Frequency Posterior Mean - 0.03 0.00 0.02 0.02 - 0.02 0.01 0.00 0.01 0.01 0.01 - 0.01

Posterior STD (0.10) (0.09) (0.12) (0.14) (0.10) (0.10) (0.10) (0.09) (0.08) (0.09) (0.08) Log_Amount Posterior Mean - 0.59 * 0.13 0.01 0.02 0.08 0.04 - 0.04 - 0.15 0.14 0.45 0.04

Posterior STD (0.36) (0.27) (0.30) (0.32) (0.25) (0.29) (0.28) (0.26) (0.26) (0.30) (0.25)

*p < .1, ** p < .05, *** p < .01

Log_amount means log average purchase amount

Table 3: This table shows the posterior mean and posterior standard deviation of Γcoefficients

(18)

In Sample Hit Rates (cumulative percentage) Out Sample Hit Rates (cumulative percentage) Number of Product Recommended (1) Random Recommendation (2) Probit Model (3) Finite mixture Probit Model (4) Hierarchical Bayes Probit (1) Random Recommendation (2) Probit Model (3) Finite mixture Probit Model (4) Hierarchical Bayes Probit 1 0.1128 % 0.2229 % 0.3325 % 0.4187 % 0.1207 % 0.2076 % 0.2398 % 0.2632 % 2 0.2242 % 0.3756 % 0.5406 % 0.6392 % 0.2414 % 0.3743 % 0.4561 % 0.4415 % 3 0.3371 % 0.5000 % 0.6761 % 0.7783 % 0.3621 % 0.5029 % 0.5819 % 0.5731 % 4 0.4472 % 0.6108 % 0.7685 % 0.8719 % 0.4828 % 0.6287 % 0.6754 % 0.7076 % 5 0.5553 % 0.7131 % 0.8349 % 0.9187 % 0.6013 % 0.7368 % 0.7661 % 0.8216 % 6 0.6527 % 0.7956 % 0.8805 % 0.9421 % 0.6993 % 0.8187 % 0.8509 % 0.8977 % 7 0.7304 % 0.8461 % 0.9089 % 0.9631 % 0.7710 % 0.8626 % 0.8918 % 0.9328 % 8 0.7902 % 0.8830 % 0.9335 % 0.9791 % 0.8194 % 0.8977 % 0.9240 % 0.9678 % 9 0.8369 % 0.9076 % 0.9483 % 0.9852 % 0.8772 % 0.9240 % 0.9591 % 0.9883 % 10 0.8695 % 0.9409 % 0.9557 % 0.9902 % 0.8793 % 0.9415 % 0.9678 % 0.9942 %

A random sample of 400 customers’ transactions was selected for analysis. The last one purchase records of sampled customers were selected as hold out sample. There are 58 customer samples with only one transaction record. Accordingly, the hold out sample contains 342 actual transaction records of 342 customers. The in sample contains 811 actual transaction records of 400 customers

Table 4 is the comparison of cumulative percentage of hit rates. The hierarchical Bayes Probit model out performs the other models in either in sample and hold out sample prediction.

(19)

Appendix

To compare the performance of hit rates, we provide two basic solutions and two

customized new product recommendation model for comparison.

Random Recommendation

The first one is random recommendation. It is assumed that no information regarding

customer’ preferences is available. If there are 10 products for choice, the hit rate for

randomly recommending one product is 1/10.

Probit model

It is assumed no knowledge regarding individual’s preference. However, the

knowledge regarding the preference structure of aggregate market is available. Thus, their

new product recommendations are based upon the same preference structure of their

customers rather then customized new product recommendation. The Probit model is

specified as follows:

y_ij =x'_ijβ ε_i+ _ij y_ij =0,1, i=1, 2,..., , n j=1, 2,...,J_i

whereε_ijm follow normal distribution. Let to denote the choice made by

individual i in Ji choice occasions, and x is a set of common product attributes.

ij y

Finite Mixture Probit Model

A finite mixture model that employs a finite set of mass points to capture heterogeneity

has a history for the analysis of individual heterogeneity. It is assumed that individuals are

(20)

implicitly sorted into a set of S classes, s=1, 2,…..S. In marketing application, these classes

can be regarded as customer segments in the market. The following is a finite mixture

Probit model for choice made by individual i (i=1, 2,…., N) observed in Ji choice situations,

where x is a set of common product attributes. Let to denote the specific choice made by

individual i in choice situation Ji, so that the model provides

ij y ij ij exp(x ' ) Prob( 1| class ) = 1 exp (x ' ) s ij s y s β β = = +

The individual specific parameter vector isβˆ_i =

∑

S_s₌₁Hˆ_{s i}_|βˆ_s (Kamakura and Russell 1989). H is the individual i ‘s probabilities of being class s. This formula will be used ˆ_{s i}_|

to estimate individual preference toward product attributes to help us to predictive the

purchase probability of any selected new or existing products.

Hierarchical Bayes Probit Model

The hierarchical Bayes approaches to modeling consumer heterogeneity have been

conducted over a wide range of marketing problems (e.g., Allenby and Ginter 1995;

Allenby, Arora, and Ginter 1998; Rossi and Allenby 2003). The model we will employ is

the hierarchical Bayes Probit model. Let to denote the specific choice made by

individual i in choice situation Ji,, x is a set of common product attributes, ij

y

i zi i

β = Γ + ζ

(21)

occasions of subject i. β_i is a matrix of individualized preference coefficients, and Γ is a matrix of coefficients that relate β_i to the value of , and is a vector of covariates that account for observed heterogeneity. In this study, the covariate includes

demographic variables (i.e., age, gender) and observed behavior variables in database (log

average purchase amount and frequency).

i

z z_i

i

ζ is unobserved heterogeneity component, which is assumed to be multivariate normal distribution (Allenby and Ginter 1995). β_i will be used in this study to estimate individual preference toward product attributes to

help us to predictive the purchase probability of any new products.

(22)

Presented at 2007 INFORMS Marketing Science Conference, June 27-30, 2007, Lee Kong Chian School of Business, Singapore Management University, Singapore.

Jen, Lichung, Hsiu-Wen Liu, and Kung-Hsin Shao, 2007, Customized New Product

Recommendation in CRM Database, 2007 INFORMS Marketing Science Conference, June 27-30, 2007, Lee Kong Chian School of Business, Singapore Management University, Singapore.

Customized New Product Recommendation in CRM Database

ABSCRACT

The purpose of this article is to provide a set of solutions for customized new product recommendation to improve the performance of CRM (Customer Relationship Management) project. We proposed two customized new product recommendation models: finite

mixture Probit and hierarchical Bayes Probit model. The proposed methods are tested by random selected customers in a home electronic retailer’s CRM database, and results show that the presented customized new product recommendation models perform well. The approach of this paper has its strength to be able to do new product recommendation and cross-selling in database marketing on a one to one basis.

Keyword: Customer Relationship Management; New Product Recommendation; Finite

(23)