The Timely Product Recommendation Based on RFM Method

Size: px
Start display at page:

Download "The Timely Product Recommendation Based on RFM Method"

Transcription

1 The Timey Product Recommendation Based on RFM Method Li-Hua Li Fu-Ming Lee Wan-Jing Liu Department of Information Management, Chaoyang University of Technoogy 168, Jifong E. Rd., Wufong Township, Taichung County, Taiwan 41349, R.O.C. cyut.edu.tw cyut.edu.tw ABSTRACT The recent study of recommendation systems and RFM method has been appied to anayze customers consumption property and the re-purchasing abiity. The RFM method empoys Recency(R), Frequency(F), and Monetary(M) to measure customers consumption oyaty. And the recommendation systems mainy to promote products for increasing profit. However, there are some probems because they ignore the reationship between product property and purchase periodicity. That is to say that the combination of recommendation system and RFM method did not take the customers product-purchasing tig into consideration. If the periodicity of product-demand can be estimated based on each customer s buying behavior, then the product recommendation at the right tig sha match the buying requirement. This is the reason why the past product recommendation studies have difficuty of increasing the accuracy. To dea with the product periodicity, this research proposes a Timey RFM (TRFM) modue which takes product property and purchase periodicity into consideration. This research is intended to (1) anayze different products to each customer s demands in different times, (2) provide a recommendation mechanism to satisfy customers needs, (3) to improve the deficiency of existing combination with recommendation and RFM. To exae the practicabiity and to vaidate the method, the experimentation uses the Foodmart2000 database of Microsoft SQL2000 to verify the accuracy of TRFM. The resuts prove that our proposed method can provide a timey recommendation and creates better resuts. Keywords: Recommendation Systems, RFM, Artificia Neura Network (ANN), Adaptive Resonance Theory (ART)

2 1. INTRODUCTION 1.1 Background The marketing strategy on Internet has become an important issue since the prevaiing of e-commerce. Many e-commerce companies deveop their promotion activities for increasing the revenue. However, customers may sti ook for repacements when they are not satisfied with the service or products. This is why the personaized recommendation service is graduay emphasized and is used to match the customer s desire. To provide personaized service, information fitering technique (Godberg, 1992) is usuay utiized in the recommendation system (Sarwar, 2000). Information fitering (Godberg, 1992) is a technique that can resove the information overoading and the recommendation system is a system that promotes products to the customer. Through the personaized recommendation system, product recommendation is generated to match user s desire and to increase the revenue. Reated researches about recommendation systems are intensivey studied, such as user profie anayzing (Kim et a., 2004) which finds user s preference, Top-N items (Cho et a., 2004) which finds the simiar interests among customers and recommends suitabe items, etc. Recenty, combination of recommendation and RFM is studied to anayze customers consumption property (Liu et a., 2005). This is to empoy Recency (R), Frequency (F), and Monetary (M) measures to understand customers oyaty and, then, to segment customers so that the recommendation is generated according to customer s consumption property. However, there is sti a probem exist because these methods ignore the reationship between purchase periodicity and product property. The timey probem of the product consumption cyce sha be considered, because after the customers consumed, the periodicity of consug the products can cause customers stronger requirement. The past recommendation system usuay recommends the items without considering the product periodicity which coud resut in futie or inefficiency. 1.2 Motivation According to the probem above, we propose a Timey RFM (TRFM) method that uses Adaptive Resonance Theory(ART) of Artificia Neura Network(ANN) to custer customers based on their purchasing behavior. This method combines product property and purchase periodicity to measure customers oyaty and to anayze customer s buying behavior. The purposes of this research are as foowed. A. To anayze different products for each customer s demands in different times. B. To provide a recommendation mechanism to satisfy customer s need. C. To improve the deficiency of existing method which combine the recommendation and the RFM method. Experimenta anaysis uses the Foodmart2000 database of Microsoft SQL2000 to verify the recommend accuracy of TRFM. Finay, the experiment resut proves this

3 method can provide a timey recommendation. In this paper, the recommendation system, Adaptive Resonance Theory(ART), RFM method, and the measure of recommendation performance are introduced in Section 2. Our propose TRFM method is expained in Section 3. The experimentation and the resuts are described in Section 4. Finay, the concusion and the contribution of this research are summarized in Section Recommendation System 2. LITERATURE REVIEW Recommendation system (Sarwar, 2000) is a system that provides information, services, or products to users. Usuay it utiizes Information Fitering technique to (Godberg, 1992) through anayze users behavior and users preference in order to provide information to fit in users needs. There are three incentives to use recommendation system for product recommendation as isted beow. A. Converting browsers into buyers: Recommendation system can hep customers finding products to match their needs. When a user is browsing the web site, there is aso possibiities that the user is wish to purchase something. Recommendation system can improve the possibiity that converting browsers into buyers. B. Increasing cross-seing: Recommendation system can recommend additiona products and, therefore, it can increase cross-seing opportunity. C. Buiding oyaty: Recommendation system can improve reationship between merchant and customer. Current researches about recommendation mechanism can be cassified into two categories, i.e., Content-Based (CB) recommendation and Coaborative Fitering (CF) recommendation. A. Content-based (CB) recommendation: CB recommendation stems from the concept of Information Retrieva (IR). This method finds reationships between user preference and item content as requested by user. It is aso caed Feature-based recommendations (Wang, 2004). B. Coaborative Fitering recommendation (CF) (Godberg, 1992): The CF technique is highy used in recommendation systems. This method first finds out a group of customers who have simiar interests with the target user. According to the group of simiar interest, CF method recommends information or products the group preferred to the target user. In our proposed method, we use recommendation system and CF to do the recommendation. To achieve the timey recommendation, RFM anaysis is aso utiized. The RFM method is introduced next.

4 2.2 RFM Method Definition RFM method is an approach of Response Mode of Market Segmentation (Suh et a., 1999). It empoys Recency(R), Frequency(F), and Monetary(M) to measure customer s oyaty and, then, segment customers into various groups for future personaization services. The content of these three parameters are expained as beow. A. Recency(R): The period since the ast purchase. A ower R vaue corresponds to a higher probabiity of repurchasing. B. Frequency(F): The number of purchasing made within a certain period. Higher F vaue means higher oyaty. C. Monetary(M): The money spent within a certain period. Higher M vaue means higher contribution Overview RFM method has been used to anayze customers consumption property. In eary studies, Sung et a.(1998) proposed a method which used Sef-organizing Map(SOM) of neura network to custer customers with their RFM vaue. However, this method needs to know the number of custer in advance which is hard to obtain. Therefore, the PBS (Purchase-based Segmentation) agorithm was proposed by Tsai et a.(2004) to avoid this probem. This method combines RFM and genetic agorithm to segment and anayze customers with purchase behavior. The combination of recommendation system and RFM were proposed by Liu et a. (2005), caed weight RFM-based (WRFM-based) method. This method mainy compared the reative importance of R, F, M vaue, respectivey. Athough these studies provide usefu methods for product recommendation, however, they ignore the reationship between product property and purchase periodicity. Hence, we propose a Timey RFM method to dea with probems of product buying periodicity Customer Quinties The customer quinties method (Migautsch, 2000) divides each customers R, F, M vaue into 5-eves. There are five equa groups for each R, F, M vaue and 125 equa size segments are generated. The initia anaysis woud be to obtain a customers R, F, M vaue, ooking at the performance of each individua ce (ces woud have definitions ike: R=4, F=3, M=5 or R=2, F=3, M=3) and understanding how different segments of the customers performed. 2.3 Adaptive Resonance Theory Adaptive Resonance Theory(ART) is a kind of Artificia Neura Network(ANN) that is cassified as an unsupervised earning network. The maor appication of ART is to do the custering. The ART network has two properties: stabiity and pasticity which is adusted by using the vigiance vaue. In ART network, the unit number of output ayer

5 represents the custering groups which wi be affected by the vigiance vaue. The higher vigiance vaue wi produce more custers, which means higher pasticity. On other hand, the ower vigiance vaue wi ead to fewer custers. The framework of ART network is shown as Figure 1. Y 1 2 Output ayer Y Y m Weight connections X 1 X 2 Input ayer X n Figure 1. Framework of Adaptive Resonance Theory A. Input ayer: The input data is represented by X1, X2,..., Xn. The number of input vectors is decided by the question domain. B. Output ayer: The number of output units represents the custer. The training process is ended whie the output units are staby fixed in a steady number. C. Weight connections: It connected the input ayer and output ayer. Every connection has two direction inks. The input and output units are fuy connections. The ART network gets the training sampe from research domain. The network wi earn from the input data and automaticay cassified the data into various custers. If the input data is not incuded at existing custers, the network wi generate a new custer. The operation of ART is simiar to human neura system. Not ony the network earns the new exampe, but aso reserves the od memory. The ART network can provide speedy, pasticity, and stabiity. Hence, we use ART network to custer customers based on their purchasing behavior. 2.4 Recommendation Performance Estimation In genera, there are three measurements for recommendation performance, namey reca measure, precision measure, and F1 measure (Cho et a., 2005; Liu et a., 2005). A. The reca vaue is to measure the ratio of recommended products with the user interested. RC Re = (1) IP RC : Number of correcty recommended products

6 IP : Number of interesting products B. The precision vaue is to measure the ratio of product correcty recommended. RC Pr = (2) RP RP : Number of recommended products C. Because the reca measure(re) and precision measure(pr) are usuay conficted. That is to say, by increasing the size of the recommendation ist eads to increase the reca measure but decrease the precision measure (Sarwar, 2000). Therefore, to refect the true performance of the recommendation, many researchers used F1 measure instead. F1 measure can baance the tradeoff between precision measure and reca measure. The F1 formua is as beow. 2 Re Pr F1= (3) Re + Pr Since these measurements can be used for testing the recommendation, therefore, we aso use Re, Pr and F1 measure for verifying the performance.

7 3. THE TRFM MODULE In this section, we propose a TRFM process to take cares the periodicity probem and to recommend products. Before starting process of TRFM, we need to, first, coect user s transaction data and product information in order to generate a set of input for further anaysis. Once the data is prepared there are four processes are designed in this method, i.e., custering process, RFM anaysis, timey anaysis, and recommendation process. The process fow of TRFM is shown in Figure 2. Each process of TRFM is introduced in the foowing steps. ART Network Y... 1 Y2 Y m X... 1 X 2 X n Custering Process Y 1 : Custer 1 Y 2 : Custer 2 Y m : Custer Quinties Method RFM Anaysis High- RFM Low- RFM Recommendation Process Timey Anaysis High- RFM Low- RFM Top-N List T u = 1 i= np ( T T ) i n T = PT LPT p i 1 u 1, if T u T R = 0, otherwise Timeiness Products Figure 2. The architecture of TRFM modue Step 1. Data coection and preprocessing This step is to obtain products, product categories, and the transaction records from the TBp, TBC, TBT, and TB D tabe(see the definition of Tabe 1), correspondingy. The outier, the inconsistent data are deeted and the R(Recency), F(Frequency), M(Monetary) vaues are computed by extracting data from date, t_id, and store_saes as defined in Tabe 1.

8 Tabe 1. Variabes definition at preprocessing step Tabes or variabes Definition DB Database. DB = { TB, TB, TB, TB } p C T D TB p Product tabe. TB P = {p_id, pc_id, p_name}. p_id: product ID. pc_id: category ID. p_name: product name. TB C Cass tabe. TB C = {pc_id, p_cate}. pc_id: category ID. p_cate: product category name. TB Transaction tabe. T TB T = {c_id, t_id, p_id, store_saes, unit_saes}. c_id: customer ID. t_id: time ID. store_saes: the money of product sod. unit_saes: number of product sod. TB Time tabe. TB = {t_id, date}. D D k Number of product category. C C C X i X i = {0,1}. X i represents the i-th transformed binary code that indicates the purchase count of product which beongs to category c. i = 1, 2,3,..., k and C = 1, 2,3,..., n. C To encode the X i data for representing the buying pattern, we sum up the purchased quantity for each product and accumuate product quantities under each category. Therefore, for each customer we have a set of product purchased amounts to represent each category s buying pattern. Since the ART network takes ony binary vaue for input, hence, each customer s categories amount is encoded into four eves which is defined and divided into four equa ranges according to each category s - range. The binary code representing each category purchasing amount is shown in Tabe 2. Therefore, we wi generate a set of input with 4 k bits represented by I = { X1, X2, X3, X4,..., X C 1, X C 2, X C 3, X C 4}. For exampe, if we have 45 product categories, then we wi generate the input pattern with 45 4 = 180 bits. Tabe 2. The 4-eve binary code representing purchasing amount Leve 1 Leve 2 Leve 3 Leve Step 2. Using ART network for custering process C After the preprocessing step, we wi obtain a set of X i which represents the product purchasing behavior. For recommendation purpose, we need to custer customers into various groups based on their product purchasing behavior. We use ART network as introduced in Section 2.3 to do the custering. The purpose is to gather the simiar buying behavior customers into same group, such that we can utiize the group behavior for those customers who have ow RFM vaue for recommendation. The custering process is shown in Figure 3.

9 Purchasing Database Processing and Data Set Generating Y... 1 Y2 Ym X... 1 X2 X n Y 2 : Custer 1 Y 2 : Custer 2 Y m : Custer m Figure 3. Custering process using ART network In Figure 3, the input ayer (I) uses the product-buying pattern to generate custers. The output ayer (O) generates m custers. The variabes I is used in this step is defined as in Tabe 3. Tabe 3. Definition of variabes for custering process Variabe Definition I Input ayer. { 1, 2, 3, 4,..., C 1, C 2, C 3, C c I = X X X X X X X X4}, X i is input vaue, i {1, 2, 3, 4}, c [1, n], n is number of category. O Output ayer, O= { Y1, Y2,..., Y m }, Ym represents output custer, m is number of quantity of output units (number of custer). Note that the c X i and the output Y m are both binary vaues. Step 3. RFM anaysis To understand each customer s contribution and oyaty, we empoy the RFM method to perform the measurement. To do this, the Customer Quinties Method (Migautsch, 2000) proposed by Migautsch (2000) is appied. This method evauates the ast purchase, number of purchases in certain period, and money spent during a certain period of customers transaction data and transfers these data into five eves by using score 1 to 5. The process of this step is shown in Figure 4.

10 Customer Quinties Method Database Transaction data 5 R, F, M 4 R, F, M 3 R, F, M 2 R, F, M 1 R, F, M High-RFM Low-RFM G1 G2 G3 G4 Gn-2 Gn-1 Gn-3 Gn Figure 4. The processes of RFM anaysis Usuay, the R-measure (Recency) means measure the time between ast purchase and present time. However, in our method, R-vaue has different consideration. Because each product has its own periodicity, which means, short-cyce products wi be bought with higher frequency, such as food or daiy necessities, and the ong-cyce product wi not be bought frequenty, such as househod appiances. If we use the traditiona R-measure to understand the customer s buying frequency and to recommend products, then we are ikey to recommend products that customer don t need due to the wrong tig. Therefore, our method takes the product periodicity into consideration to udge the purchase frequency. And, we use the foowing formua to recognize the periodicity. T f ( Ti Ti 1) ( T 2 f T( f 1) ) ( ( f 1) ( f 2) )... ( 2 1 ) i + T T + + T T Tf T = 1 = = = f 1 f 1 f 1 v T = 1 PC RT = v RA if R = B PC θ otherwise (4) (5) (6) T i : The time of customer buying product of category at the i th purchase. T : The average time of customer buying product of category. f : The count frequency of customer buying product of category. PC : The periodicity of buying product of category for a customers. v : The number of customers who had purchased product of category. θ : Threshod vaue. This is to be used for cassifying the short-cyce or ong-cyce product. RT : RT { RA, RB}. If RT = RA means the product is a short-cyce product and if RT = R B means the product is a ong-cyce product

11 The threshod vaue can be adusted according to the probem domain. After we recognize the periodicity of product, we appy RFM method by integrating the short-cyce ( R ) and ong-cyce ( ) into to cacuation. We need to obtain the ast A R B purchase time ( R ), the purchase frequency ( F ), and the amount of money spent in a period of time ( M ) for each customer. For a R, F, M vaues, we aso find out its imum vaue, i.e., R, F, M and its imum vaue, i.e., R, F, M. Its average vaue, R, F, M, wi be used to separate the measurement into five-eves. The reated formua and variabes are described next. R R F F M M R =, F =, M = (7) R = Min{ R1, R2,..., Ru }, F = Min{ F1, F2,..., Fu }, M = Min{ M1, M 2,..., M u} (8) = { }, F = Max{ F F F }, M Max{ M M M } R Max R1, R2,..., Ru 1 2,,..., u =,,..., u (9) 1 2 R : The average of a recency vaue, R F : The average of a frequency vaue, F M : The average of a monetary vaue, M R : Minimum R-vaue F : Minimum F-vaue M : Minimum M-vaue R : Maximum R-vaue. F : Maximum F-vaue. M : Maximum M-vaue. In order to figure out the oyaty of a customer, we empoy Customer Quinties Method to obtain the RFM score at shown in Tabe 4. Tabe 4. The 5-eve of R, F, M vaue R-vaue( R ) R-vaue( R ) F-vaue M-vaue A Range R ~ R R ~ R F ~ F M ~ M B 1 R ~ R 4 R ~ R 1 F ~ F 1 M ~ M 1 2 R 4~ R 3 R 1~ R 2 F 1~ F 2 M 1~ M 2 3 R 3~ R 2 R 2~ R 3 F 2~ F 3 M 2~ M 3 4 R 2~ R 1 R 3~ R 4 F 3~ F 4 M 3~ M 4 5 R 1~ R R 4~ R F 4~ F M 4~ M

12 Based on the Tabe 4, we can transform each R, F, M vaue into their corresponding eve, namey, RL, FL, ML, respectivey. According to the RFM method, we sum up three vaues as G average RFM vaue from a the customers RFM vaue as for customer (see the formua 10). Then, we obtain the (see the formua 11). We use the rue as formua 12 to categories customer into two groups, i.e., Low-RFM group and High-RFM group. The Low-RFM group means customers with ower oyaty and the High-RFM group means customers with higher oyaty. G = RL + FL + ML, [1, u], RL, FL, ML {1,2,3,4,5} (10) u GT GAvg =, GT = G u (11) = 1 GT : The tota vaue of a the G u : Tota number of customers G Avg If G < G g then customer Low-RFM Av otherwise customer High-RFM (12) Step 4. The Timey Anaysis We consider each product for each customer has its specific demand-tig. For instance, after a customer bought a product, the product has consumed, which means the periodicity of the product reached, hence, the recommendation of this product sha be introduced. Therefore, in this step, product property and purchase periodicity are combined to measure the proper recommendation tig for each customer. In timey anaysis phase, we cacuate the product periodicity by using the formua 13 and 14. TD = PT LPT (13) 1, if TD T P = 0, otherwise (14) PT : Present time. LPT : The ast time of customer purchased product in category TD : Distance between PT and LPT. P : If P = 1 which means the product is a timey product, if P = 0 which means the product is not a timey product.

13 Step 5. Recommendation Process Based on the resuts generated from the step 3, i.e., RFM anaysis process, Low-RFM group and High-RFM group are used in this recommendation process. The Low-RFM group contains customers with ower oyaty; on the contrary, the High-RFM has higher contribution for purchasing. In this step, we find the top-n products based on the neighbor of customers preference for Low-RFM group. The High-RFM group is customers with high oyaty. Therefore, we use the timeiness products as obtain in step 4 to recommend to the customer in this group. The process of this step is shown in Figure 5. C 1 C 2 C 3 : : : : : : C n Low-RFM High-RFM Find the products uses the preferences of neighbors. Find the products that the customers needs based on the timeiness products. Top-N Recommendation List Timeiness Products Recommendation List Figure 5. The of recommendation process 4.1 Data source and toos 4. EXPERIMENTATION The experimentation is conducted by using Microsoft Office Access 2003 for data preprocessing and Matab 6.5 with ART1 neura network earning too for custering. The data source uses Foodmart2000 which is obtained from Microsoft SQL2000. From this Foodmart2000 database, records and fieds of saes_fact_1998, product, product_cass, and time_by_day tabes are coected. We use data records from January 1998 to February, In tota, there are 935 customers, 7691 transaction records, and 45 product categories used in this experimentation. 4.2 Data Anaysis In custering process, we use ART to custer customers. The training data with categories of January 1998 is used as the input, I = { X1, X2, X3, X4,..., X1, X2, X3, X4 }, for ART network. To verify the network, a set of test data using transaction records of February 1998 is used. The input vaue represents the frequency of purchase. The encoding process is described in Step2 of Section 3.

14 The vigiance vaue of ART is decided based on the point of Steepest Knee (Pa et a., 1997). The reationship between the vigiance vaue and the number of custers are shown in Figure 6. When the vigiance vaue is set to 0.8, ART network wi generate 62 custers. But the number of custer increase to 212 groups when the vigiance vaue is set to 0.9. This indicates that at the point of vigiance vaue 0.8, this is a turning point for custer size. For this reason, we choose 0.8 as our vigiance vaue. Number of custer Vigiance vaue Figure 6. Reationship between the vigiance vaue and the number of custers We cacuate TRFM vaue and anayze distribution of three measures (R-vaue, F-vaue, and M-vaue) individuay. The resuts of R, F, M vaue is shown as Tabe 5. Let take M-vaue as an exampe, its distribution range ies between 1 and 250, hence, the 5-eve range uses 50 to separate each eve. Tabe 5. The resuts of 5-eve of R, F, M vaue Range Leve 1 Leve 2 Leve 3 Leve 4 Leve 5 R-vaue 1~30 25~30 19~24 13~18 7~12 1~6 (Frequency product) R-vaue 1~30 1~6 7~12 13~18 19~24 25~30 (Non-frequency product) F-vaue 1~39 1~8 9~16 17~24 25~32 33~39 M-vaue 1~250 1~50 51~ ~ ~ ~250 In RFM anaysis, the customers average RFM vaue is Therefore, a customer is in Low-RFM group if the customer s RFM vaue ies between 3 and 5 and a customer is in High-RFM group if the customer s RFM vaue ies between 6 and 15.

15 According to Low-RFM group and High-RFM group, different recommendation approaches are used, where the Low-RFM group utiizes CF method for recommendation and the High-RFM group gets the timey product recommendation. Customers recommended items resuted from this experimentation are shown as exampe in Tabe 6. Tabe 6. Recommendation ist Customer ID Group Recommendation items 1495 H 1, 4, 5, 13, 18, 23, 25, 31, 32, 36, 42, 3762 H 12, 18, 22, 25, 31, 32, 36, 37, 42, 45, 4198 L 18, 25, 32, 42, L 18, 25, 32, 42, L 18, 25, 32, 42, H 2, 13, 14, 18, 20, 26, 32, 37, 42, 45, 6527 H L L L 18, 25, 32, 42, 45 The recommendation performance estimation is as introduced in Section 2.4. There are three measures (reca, precision, and F1 measure) to be exaed. The anaysis resut of this experimentation is shown in Tabe 7. The reca vaue of customer 4410 and customer 9715 is 100%, which means a the products recommended from the system is aso interested by the customer. The precision from customer 4198 is 100% which means a the products recommended from the system are bought by the customer. The F1 measure of customer 3762 and 6632 is higher than 60%. In past studies of recommendation, the F1 measure usuay ower than 60%, such as WRFM-based method (Liu et a., 2005). Therefore, this proves that a TRFM method can offer a better recommendation resut. Tabe 7. The reca, precision, and F1 measure Customer ID Reca Precision F1 Measure

16 5. CONCLUSION In this research, we propose a TRFM method to sove probem of produce periodicity. The TRFM method incudes four processes: (1) custering process using ART network, (2) the RFM anaysis to anayze each customer s purchasing period for each product and then cacuating RFM score, (3) the timey anaysis process to find out products with timeiness, and (4) the recommendation to produce recommendation ist. The contributions of this timey product recommendation are as foows. A. TRFM takes the product purchasing cyce into consideration. After the customer purchase a product, the consumption of the products or the periodicity of consug the products is cacuated. B. TRFM combines the product property and the periodicity of purchasing to provide a better recommendation to satisfy customers needs. C. TRFM can reduce the deficiency of the recommendation system based on RFM technique. ACKNOWLEDGEMENTS The authors wish to thank the co-editors and the reviewers for their inspection. This research was supported by the Nationa Science Counci of the Repubic of China under the grant NSC H REFERENCES Cho, Ho Yoon and Kim, Jae Kyeong Appication of Web usage ing and product taxonomy to coaborative recommendations in e-commerce. Expert Systems with Appications. 26(2), Cho, Yeong Bin, Cho, Yoon Ho, and Kim, Soung Hie Mining changes in customer buying behavior for coaborative recommendations. Expert Systems with Appications. 28(2) Godberg, David, Nichos, David, Oki, Brain M., and Terry, Dougas Using coaborative Fitering to Weave an Information TAPESTRY. Communications of The ACM. 35 (12), Kim, Won, Ko, I-Ju, Yoon, Jin-Sung, and Kim, Gye-Young Inference of recommendation information on the internet using improved FAM. Future Generation Computer Systems. 20(2), Liu, Duen-Ren and Shih, Ya-Yueh Integrating AHP and data ing for product recommendation based on customer ifetime vaue. Information and Management. 42(3), Migautsch, John R Thoughts on RFM Scoring. Journa of Database Marketing. 8(1), Min, Sung-Hwan and Han, Ingoo Detection of the customer time-variant pattern for improving recommender systems. Expert Systems with Appications. 28(2).

17 Pa, N. R., and Biswas, J Custer Vaidation Using Graph Theoretic Concepts. Pattern Recognition. 30(6), Roh, Tae Hyup, Oh, Kyong Joo, and Han, Ingoo The coaborative fitering recommendation based on SOM custer-indexing CBR. Expert Systems with Appications. 25(3) Sarwar, Badru, Karypis, George, Konstan, Joseph, and Ried, John Anaysis of Recommendation Agorithms for E-commerce. Proceedings of The 2nd ACM Conference on Eectronic Commerce Suh, E. H., Noh, K. C., and Suh, C. K Customer ist segmentation using the combined response mode. Expert Systems with Appications. 17(2), Sung, Ho Ha and Sang, Chan Park Appication of data ing toos to hote data mart on the Internet for database marketing. Expert Systems with Appication Tsai, C.-Y. and Chiu, C.-C A purchase-based market segmentation methodoogy. Expert Systems with Appications. 27(2) Wang, Feng-Hsu and Shao, Hsiu-Mei Effective personaized recommendation based on time-framed navigation custering and association ing. Expert Systems with Appications. 27(3) Weng, Sung-Shun and Liu, Mei-Ju Feature-based recommendation for one-to-one marketing. Expert Systems with Appications. 26(4),