Integrating AHP and data mining for product recommendation
based on customer lifetime value
Duen-Ren Liu
a,*, Ya-Yueh Shih
a,ba
Institute of Information Management, National Chiao Tung University, Hsinchu 300, Taiwan
b
Department of Information Management, MingHsin University of Science and Technology, Hsinchu, Taiwan Received 5 March 2003; received in revised form 29 August 2003; accepted 6 January 2004
Available online 9 April 2004
Abstract
Product recommendation is a business activity that is critical in attracting customers. Accordingly, improving the quality of a recommendation to fulfill customers’ needs is important in fiercely competitive environments. Although various recommender systems have been proposed, few have addressed the lifetime value of a customer to a firm. Generally, customer lifetime value (CLV) is evaluated in terms of recency, frequency, monetary (RFM) variables. However, the relative importance among them varies with the characteristics of the product and industry. We developed a novel product recommendation methodology that combined group decision-making and data mining techniques. The analytic hierarchy process (AHP) was applied to determine the relative weights of RFM variables in evaluating customer lifetime value or loyalty. Clustering techniques were then employed to group customers according to the weighted RFM value. Finally, an association rule mining approach was implemented to provide product recommendations to each customer group. The experimental results demonstrated that the approach outperformed one with equally weighted RFM and a typical collaborative filtering (CF) method.
# 2004 Elsevier B.V. All rights reserved.
Keywords: Recommendation; Marketing; Analytic hierarchy process (AHP); Customer lifetime value; Collaborative filtering; Clustering; Association rule mining
1. Introduction
Intense competition is forcing companies to develop innovative marketing activities to capture customer needs and improve customer satisfaction and retention. The use of the Internet and the explosive growth of e-commerce have expanded marketing activities and made large volumes of customer data available for analysis. Businesses can benefit signifi-cantly from analyzing customer data to determine
their preferences and thus improve marketing decision support. Providing adequate support to meet customer needs can boost the success of on-line e-stores [18]
and web site success depends on enhancing informa-tion and service quality to serve customers better[21]. Recently, IT has been utilized to help companies maintain competitive advantage [36]. Data mining techniques[9]are a widely used information technol-ogy for extracting marketing knowledge and further supporting marketing decisions[4,5,33]. The applica-tions include market basket analysis, retail sales ana-lysis, and market segmentation analysis. Lin et al.[19]
applied data mining techniques to extract inter-orga-nizational retailing knowledge from POS information
*Corresponding author. Tel.:þ886-3-5712-121;
fax:þ886-3-572-3792.
E-mail address: dliu@iim.nctu.edu.tw (D.-R. Liu).
0378-7206/$ – see front matter # 2004 Elsevier B.V. All rights reserved. doi:10.1016/j.im.2004.01.008
in retail store chains. Moreover, Hui and Jha [14]
employed it to provide customer service support. The knowledge can support marketing decisions and cus-tomer relationship management.
The buying patterns of individual customers and groups can be identified via analyzing customer data
[38], but also allows a company to develop one-to-one marketing strategies that provide individual marketing decisions for each customer[24]. Recommender sys-tems are technologies that assist businesses to implment such strategies. They have emerged in e-commerce applications to support product recommen-dation[31]. The systems use customer purchase his-tory to determine preferences and identify products that a customer may wish to purchase. Schafer et al. presented a detailed taxonomy of recommender sys-tems in e-commerce, and determined how they can provide personalization to establish customer loyalty. Generally, recommender systems increase the prob-ability of cross-selling; establish customer loyalty; and fulfill customer needs by discovering products in which they may be interested.
Collaborative filtering (CF) has been successfully used in various applications. The CF method utilizes preference ratings given by various customers to determine recommendations to a target customer based on the opinions of other customers. The Group-Lens system [26]applied the CF method to recom-mend Usenet news and movies. Video recomrecom-mender
[12]also used this approach to generate recommenda-tions on movies. Examples of music recommender systems are Ringo [32] and MRS [8]. Siteseer [27]
provided recommendations based on the bookmarks of the user’s virtual neighbors. Content-based filtering provides recommendations by matching customer profiles (e.g. interests) with content’s features (e.g. product attributes). NewsWeeder[17]is an example of content-based recommender systems. Changchien and Lu[7]developed a procedure for mining association rules to support on-line product recommendations. Amazon.com [20]employed item-to-item collabora-tive filtering to provide recommendations of those products that are similar to the customer’s purchased and rated products. However, few have considered customer lifetime value (CLV).
From the perspective of niche marketing, all cus-tomers are not equal (they have different lifetime value or purchase behaviors), even if they purchase identical
products or services; market segmentation is therefore necessary. Firms are increasingly recognizing the importance of the lifetime value of customers [3]. Several studies have considered the use of CLV. Generally, recency, frequency, and monetary (RFM) methods have been used to measure it [16,23]. The concept has been applied to cluster customers for niche marketing[11].
Our work proposes a novel product recommenda-tion methodology that combines group decision-mak-ing and data mindecision-mak-ing. The analytic hierarchy process (AHP) [28] was applied to evaluate the importance (weight) of each RFM variable, according to a group of decision-makers. Clustering was then employed to group customers based on their weighted RFM value. Finally, association rule mining was used to provide product recommendations for each group of custo-mers.
2. Background
2.1. Customer lifetime value analysis and RFM evaluation
Customer lifetime value (CLV) is typically used to identify profitable customers and to develop strategies to target customers[15]. Measuring RFM is an impor-tant method for assessing customer lifetime value. Bult and Wansbeek [6] defined the terms as: (1) R (Recency): period since the last purchase; a lower value corresponds to a higher probability of the cus-tomer’s making a repeat purchase; (2) F (Frequency): number of purchases made within a certain period; higher frequency indicates greater loyalty; (3) M (Monetary): the money spent during a certain period; a higher value indicates that the company should focus more on that customer.
Numerous studies have discussed the evaluation of CLV. Goodman[10]suggested that the RFM method avoided focusing on less profitable customers, allow-ing resources to be diverted to more profitable custo-mers. Hughes [13] proposed a method for RFM scoring that involved using RFM data concerning to sort individuals into five customer groups. Different marketing strategies could then be adopted for differ-ent customers. Stone [35] suggested that different weights should be assigned to RFM variables
depend-ing on the characteristics of the industry. In analyzdepend-ing the value of customers who used credit cards, he suggested placing the highest weighting on the Fre-quency, followed by the Recency, with the lowest weighting on the Monetary measure. However, he determined the RFM weightings subjectively, without employing a systematic approach.
2.2. Market segmentation
Clustering[25]seeks to maximize variance among groups while minimizing variance within groups. Many clustering algorithms have been developed, including K-means, hierarchical, fuzzy c-means, etc. We used the K-means method to group customers with similar lifetime value according to weighted RFM. K-means clustering [22] is a method com-monly used to partition a set of data into groups. This scheme proceeds by selecting m initial cluster centers and then iteratively refining them. (1) Each instance diis assigned to its closest cluster center; (2) each cluster center Cjis updated to the mean of its constituent instances. The algorithm has converged when the assignment of instances to clusters no longer changes.
2.3. Association rule mining
Association rule mining, which identifies associa-tions among a set of product items frequently pur-chased together, is a widespread approach for market basket analysis[1,34]. It attempts to find association rules that satisfied minimum support and minimum confidence requirements. Appendix A provides the formalization of association rule mining. The support of an association rule indicates how frequently that rule applies to the data. Higher support corresponds to a stronger correlation between the product items. The confidence is a measure of the reliability of an asso-ciation rule. It corresponds to a more significant correlation between product items. The apriori algo-rithm[2]is typically used to find association rules by discovering frequent itemsets (sets of product items). An itemset is considered to be frequent if the support of that itemset exceeds a user-specified minimum support. Association rules that meet a user-specified minimum confidence, can be generated from the frequent itemsets.
2.4. Association rule based recommendation Sarwar et al.[30]described the method of associa-tion rule-based recommendaassocia-tion as: for each custo-mer, a customer transaction is created to record all the products previously purchased by a customer. The association rule mining algorithm is then applied to find all the recommendation rules that satisfy the given minimum support and minimum confidence con-straints. The top-N products to be recommended to a customer u, are then determined according to the recommendation rules. A detailed illustration is pro-vided inAppendix B.
2.5. Collaborative filtering
A typical collaborative filtering (CF) method employs nearest-neighbor algorithms to recommend products to a target customer u based on the prefer-ences of neighbors, that is, those customers having similar preferences to customer u. Preferences gen-erally are defined in terms of customer purchasing behavior or taste (preference rating on products).
Appendix C lists one common approach to compute the similarity of preferences among customers based on the Pearson correlation coefficient.
Customers are ranked by their similarity measures in relation to the target customer u. The k most similar (highest ranked) customers are selected as the k nearest neighbors of customer u. The frequency count of products is calculated by scanning the purchase data of the k-nearest neighbors. The products then are sorted based on frequency count. The N most frequent products that have not yet been purchased by target customer u are selected as the top-N recommenda-tions.
3. Methodology: integrating AHP, clustering and association rule mining
The proposed recommendation methodology pri-marily utilizes AHP, clustering, and association rule mining techniques, as shown inFig. 1. The rationale of the proposed approach is that if customers have had similar purachasing behavior or purchases, then they are very likely also to have similar RFM values. However, RFM values could be similar given very
different product purchases. Thus, the approach devel-oped here employed two steps to identify similar purchase patterns. First, RFM values were used to cluster customers into groups with similar RFM values: The weighting (relative importance) of each RFM variable was evaluated using AHP. K-means clustering then was employed to group customers with similar lifetime value or loyalty, according to weighted RFM. Second, an association rule mining approach was applied to extract recommendation rules, namely, frequent purchase patterns from each group of custo-mers. The extracted frequent purchase patterns repre-sent the common purchasing behavior of customers with similar product purchases. Therefore, the approach presented in this work recommends products to customers based on frequent purchase patterns of customers with similar product purchases.
A case study was used to illustrate the methodology. It concerns a hardware retailing company that man-ufactures wheels, casters, platforms, and hand trucks for industrial, medical, hospital and institutional use. This company produces over 3000 products. Its
deci-sion-makers must target customer groups and develop market strategies to satisfy customer needs and thereby increase the market share of the company. Two years of data on consumer transactions, approxi-mately 70,000 rows, have been collected. The data set was preprocessed to extract customer transactions. Unreasonable records, such as those of customers who have a non-zero purchase but have never made any transactions, were also removed. RFM values of the 895 customers were extracted from the database to measure the customers’ CLV.
3.1. AHP approach
The AHP was used to determine the relative impor-tance (weights) of the RFM variables, wR, wF, and wM. The three main steps of the AHP are as follows. 3.1.1. Step1: perform pairwise comparisons
This asks evaluators (decision makers) to make pairwise comparisons of the relative importance of RFM variables using the scale shown inTable 1.
3.1.2. Step 2: Assess the consistency of pairwise judgments
Evaluators may make inconsistent judgments when making pairwise comparisons. Before the weights are computed, the degree of inconsistency is measured by an inconsistency index. Perfect consistency implies a zero inconsistency index. However, perfect consis-tency is seldom achieved, since humans are often biased and inconsistent, when making subjective judg-ments. Therefore, an inconsistency index of less than 0.1 is acceptable. If the inconsistency index exceeds this, then the pairwise judgments may be revised before the weights of RFM are computed.
3.1.3. Step 3: Computing the relative weights This determines the weight of each decision ele-ment. This work employs Eigenvalue computations to derive the weights of the RFM.
In our study, the three groups of evaluators judge the RFM weightings: three administrative managers, two business managers in sales, and one marketing con-sultant, and five customers who had previously made at least one purchase. These groups were invited to evaluate the relative importance of the RFM variables. Data were gathered by interviewing the evaluators. Interviews were conducted using a questionnaire (Table 2), and the answers were expressed in the form of a pairwise comparison matrix (Table 3).
According to the assessments, the relative weights of the RFM variables are 0.731, 0.188 and 0.081, respectively. The implication of the RFM weightings is that recency is the most important variable; thus evaluators must mainly concentrate on whether cus-tomers purchase regularly. If some perform no trans-action for a long period, they may have been lost or transferred to a new vendor.
Table 1
Relative degree of importance for pairwise comparisons Comparative
importance
Description Explanation
1 Equally importance Two activities contribute equally to the objective
2 Intermediate between equal and weak Experience and judgment slightly favor one activity over another 3 Weak importance of one over another Experience and judgment slightly favor one activity over another 4 Intermediate between weak and strong Experience and judgment strongly favor one activity over another 5 Essential or strong importance Experience and judgment strongly favor one activity over another 6 Intermediate between strong and demonstrated An activity is strongly favored and its dominance is demonstrated in
practice
7 Demonstrated importance An activity is strongly favored and its dominance is demonstrated in practice
8 Intermediate between demonstrated and absolute The evidence favoring one activity over another is of the highest possible order of affirmation
9 Absolute or extreme importance The evidence favoring one activity over another is of the highest possible order of affirmation
Table 2
AHP questionnaire for RFM
Criteria Comparative importance Criteria 9:1 7:1 5:1 3:1 1:1 3:1 5:1 7:1 9:1
Recency 9 7 5 3 1 3 5 7 9 Frequency
Recency 9 7 5 3 1 3 5 7 9 Monetary
3.2. Clustering customers with similar lifetime value
Customers with similar lifetime values, in terms of weighted RFM, were next clustered using the K-means method. This must specify the number of clusters, m, in advance. The parameter was set to 8, since eight (2 2 2) possible combinations of inputs (RFM) can be obtained by assigning # or ", according to the average R (F, M) value of a cluster being less than or greater than the overall average R (F, M). The RFM values of customers were normalized as follows. The profit form, x0¼ ðx xSÞ/(xL xS), was used to normalize the F (frequency) and M (monetary) values, since F and M positively influenced CLV or loyalty. The cost form, x0¼ ðxL xÞ/(xL xS), was used to normalize the R value, since it negatively impacted CLV. x0 and x represented the normalized and original R (F, M) values, while xL and xS repre-sented the largest and smallest R (F, M) value of all customers. The normalized RFM values of each cus-tomer were then multiplied by the relative importance of RFM variable, wR, wFand wM, which were deter-mined by the AHP. The K-means method was then
applied to cluster the customers into eight groups, according to the weighted RFM values.
Table 4 presents the result, listing eight clusters, each with the corresponding number of customers and their average R, F and M values. The last row also shows the overall average for all customers. These, for each cluster, were compared with the overall averages. If the average R (F, M) value of a cluster exceeded the overall average R (F, M), then an upward arrow" was included. The last column ofTable 4shows the RFM pattern for each cluster.
Each cluster represents a market-segmentation. Customers in clusters with the pattern R# F " M " are considered to be loyal, purchased recently, pur-chase frequently, and spend regularly with the firm. They are gold customers. Clusters with the pattern R# F # M # may include new customers who have only recently visited the company. Customers in such clusters may be trying to develop closer rela-tionships with the company. These customers may become gold customers. Finally, clusters with the pattern R" F # M # include those who very rarely visited the site and made very few transactions. They are valueless customers, and may only make pur-chases during sales. Enterprises reduce prices to attract such customers.
Analysis of variance is used to determine whether RFM variables could be used to distinguish the eight clusters (whether statistically significant). The analysis rejected the null hypothesis H0because the P-values were significant (P < 0:05). The result con-firmed that these eight clusters can be significantly distinguished by recency, frequency, and monetary.
Table 3
Example of RFM pairwise comparison matrix
Recency Frequency Monetary
Recency 1 5 7
Frequency 1/5 1 3 Monetary 1/7 1/3 1
Table 4
Eight clusters generated by K-means clustering
Cluster Number of customers Recency (days) Frequency Monetary (NT dollars) Type
1 212 79 36 199010 R# F # M # 2 150 69 54 306065 R# F " M " 3 190 66 95 593861 R# F " M " 4 123 92 41 152007 R" F # M # 5 47 147 18 100483 R" F # M # 6 100 108 23 130096 R" F # M # 7 28 162 10 71536 R" F # M # 8 45 135 25 67403 R" F # M # Overall average 89 48 270837
3.3. CLV ranking
The CLV ranking was derived to help develop more effective strategies for retaining customers and thus identify and compare market segments. The ranking of clusters proceeds as follows. The RFM values of each customer were normalized.Table 5shows the average normalized RFM values of each cluster, denoted as CjR, CjF, and CMj , respectively, for j¼ 1 to m (the number of clusters). CjR, CFj, and CjMwere computed by averaging the normalized RFM values of customers in cluster j. Let CIjbe the integrated rating of cluster j. CjIwas computed as the weighted sum of CRj, CjF, and CjM, that is, CIj¼ wRC j Rþ wFC j Fþ wMC j M, where wR, wF and wM are the relative importance of the RFM variables from AHP. Finally, the CLV ranking of the clusters was derived according to their integrated rating. The ranking indicated that cluster three had the highest rank, followed by cluster two. Customers in a cluster with a higher rank are more loyal. 3.4. Recommendation based on association rules
For each customer, a customer-transaction was created to record all the products previously pur-chased by him or her. The transactions were grouped according to the clusters of customers. Association rule mining was then used to extract the recommen-dation rule set RSjfrom transactions associated with each cluster, rather than from all customer transac-tions. The cluster Cjto which a customer, u, belonged was first identified. Then, RSj, the recommendation rule set extracted from Cjwas used to select the top-N candidate products to be recommended to customer u. Let Xu represent the set of products previously
purchased by customer u. For each recommendation rule X) Y in RSj, if X Xuthen all products in Y–Xu are the candidate products for recommendation to customer u. All candidate products were sorted and ranked according to the associated confidence of the recommendation rules. The N highest ranked candi-date products were selected as the top-N recom-mended products.
4. Experimental evaluation 4.1. Experimental setup
The proposed method was experimentally com-pared with three other methods—the non-weighted RFM method, the non-clustering method, and the typical CF method. The non-weighted RFM method does not consider the relative importance of RFM variables. The method initially sets wR¼ wF¼ wM, and then uses K-means clustering to cluster customers according to the RFM values of customers. Associa-tion rule-based recommendaAssocia-tion was applied to each cluster to recommend the top-N products. The non-clustering method did not perform non-clustering before making an association rule-based recommendation. The recommendation rules were extracted by mining association rules from the entire set of customer transactions. The typical CF method uses the prefer-ences on product purchases to compute the similarity between customers, and then employs the k-nearest neighbor (k-NN) approach to derive top-N recommen-dations.
Various experiments were performed to compare the quality of recommendations made by the proposed
Table 5
CLV ranking by weighted sum of normalized RFM values
Cluster Recency CjR Frequency CjF Monetary CjM Integrated rating CjI CLV ranking
1 0.777 0.0151 0.0228 0.573 3 2 0.856 0.0232 0.0352 0.633 2 3 0.883 0.0413 0.0684 0.658 1 4 0.667 0.0174 0.0174 0.492 4 5 0.204 0.0073 0.0115 0.151 7 6 0.527 0.0093 0.0149 0.388 5 7 0.077 0.0033 0.0081 0.058 8 8 0.301 0.0103 0.0075 0.222 6 CjI¼ wRC j Rþ wFC j Fþ wMC j MðwR¼ 0:731; wF¼ 0:188; wM¼ 0:081Þ:
method with those of the other three methods. In comparing the weighted with the non-weighted RFM method, clusters with the same order of CLV ranking were compared.
The hardware retailing data set was divided into a 75% training set and a 25% testing set. The training set included product items purchased by customers in a specified period and was used to extract recommenda-tion rules by associarecommenda-tion rule mining. The minimum confidence level was set to 0.8 and the minimum support to 0.1. Identifying all frequent itemsets was difficult, since the average number of product items purchased by customers exceeded 60. Hence, associa-tion rule mining explored only frequent itemsets with sizes less than or equal to three. Testing data were used to verify the quality of the recommendations of the various methods.
4.2. Evaluation metrics
Two metrics, precision and recall, are commonly used to measure the quality of a recommendation. These are also used measures in information retrie-val [29]. Product items can be classified into pro-ducts that customers are interested in purchasing, and those that they are not interested in purchasing. A recommendation method may recommend inter-esting or uninterinter-esting products. The recall-metric indicated the effectiveness of a method for locating interesting products. The precision-metric repre-sented the extent to which the product items recom-mended by a method really are interesting to customers.
Recall is the fraction of interesting product items that can be located.
Recall¼number of correctly recommended items number of interesting items
Precision is the fraction of recommended products (predicted to be interesting) that are really found to be interesting.
Precision¼number of correctly recommended items number of recommended items Items interesting to customer u were those products purchased by u in the test set. Correctly recommended items were those that match interesting items. How-ever, increasing the number of recommended items tended to reduce the precision and increase the recall. An F1-metric[37]could be used to balance the trade-off between precision and recall. F1 metric assigned equal weight to precision and recall and was given by, F1¼2 recall precision
recallþ precision
Each metric was computed for each customer, and the average value computed for each cluster, as well as the overall average (over all customers) as measures of the quality of the recommendation.
4.3. Experimental results
4.3.1. Comparing weighted RFM with non-clustering method
The quality of the top-all recommendation gener-ated by the weighted RFM method was analyzed for
Table 6
Quality of recommendation by weighted RFM and non-clustering (top-all)
CLV ranking Weighted-RFM Non-clustering
Precision Recall F1-metric Precision Recall F1-metric
1 0.433 0.893 0.580 0.431 0.783 0.550 2 0.385 0.878 0.532 0.420 0.710 0.515 3 0.368 0.828 0.491 0.330 0.674 0.437 4 0.321 0.804 0.446 0.272 0.751 0.382 5 0.282 0.847 0.413 0.247 0.623 0.351 6 0.219 0.758 0.324 0.180 0.453 0.248 7 0.192 0.741 0.286 0.145 0.721 0.232 8 0.184 0.674 0.285 0.143 0.625 0.227 Overall average 0.346 0.836 0.476 0.326 0.697 0.430
each cluster. The top-all recommendation recom-mended all candidate products to the customer.
Table 6 presented the CLV ranking of clusters and the average performance values—Precision, Recall and F1-metric for each cluster. The average perfor-mance value of a cluster was computed over the customers in the cluster. The last row in the table gave the overall average for all customers. For the non-clustering method, clusters generated by the weighted RFM method were used to compute the average performance values of each cluster. The weighted RFM method extracted recommendation rules from customer-transactions in a cluster, while the non-clus-tering method extracted them from the entire training set. As presented inTable 6, the performance values (precision, recall, and F1-metric) for weighted RFM generally exceeded those for the non-clustering method. The weighted RFM method yields better recommendations.
4.3.2. Comparing weighted RFM with non-weighted RFM method
The top-all recommendation quality by the pro-posed methodology, weighted RFM, was compared with that by the non-weighted RFM. The clusters generated by weighted and non-weighted RFM are different. The two methods were compared using clusters of the same CLV ranking order.Table 7shows the result. For all clusters, the F1-metrics of weighted RFM exceeded those of non-weighted RFM, except for cluster six. The overall average precision, recall and F1 metrics of weighted RFM exceeded those of non-weighted RFM. Thus the weighted RFM method
outperforms the non-weighted RFM method. For weighted and non-weighted RFM, the relationship between CLV rank and F1-metric was positive. The F1 metrics of more highly ranked clusters generally exceeded those of the lower-ranked clusters; the clus-ters with a higher CLV rank included more loyal customers. This result implies that the proposed meth-odology is more effective for more loyal customers. However, those with a lower CLV ranking may not receive improved recommendations.
4.3.3. Effect of CLV ranking and top-N recommendations
Earlier experimental results indicated that, the F1-metrics of clusters were generally positively as compared with the CLV rankings. The quality of recommendation for clusters with a high CLV ranking exceeded that for clusters with a lower CLV ranking. This experiment examined the effect of varying N, the number of recommended items. Fig. 2compares the F1 metrics of the weighted RFM (WRFM) with non-weighted RFM (non-WRFM) for top-4, top-10, top-30 and top-50 recommended product items. The analy-tical results indicated that the positive relationship between CLV ranking and recommendation quality may not have applied for small N (top-4 and top-10). This implies that appropriately selecting the number of recommended items is critical in product recom-mender systems.
Fig. 3presents the effect of top-N on the quality of recommendation, when the weighted RFM method was used. For clusters with a high CLV rank (1, 2 or 3), the F1 metrics stopped rising at a large N (18–30).
Table 7
Quality of recommendations for weighted RFM and non-weighted RFM (top-all)
CLV ranking Weighted RFM Non-weighted RFM
Precision Recall F1-metric Precision Recall F1-metric
1 0.433 0.893 0.580 0.397 0.912 0.543 2 0.385 0.878 0.532 0.366 0.903 0.519 3 0.368 0.828 0.491 0.351 0.822 0.482 4 0.321 0.804 0.446 0.320 0.802 0.442 5 0.282 0.847 0.413 0.168 0.838 0.257 6 0.219 0.758 0.324 0.216 0.820 0.334 7 0.192 0.741 0.286 0.177 0.734 0.264 8 0.184 0.674 0.285 0.176 0.659 0.273 Overall average 0.346 0.836 0.476 0.317 0.844 0.445
Thus, recommending more items helped to increase the F1 metric and improved the quality of recommen-dation for clusters with a high CLV rank—for more loyal customers. For clusters with a low CLV rank, such as 6 and 7, the F1 metrics stopped rising at a small N (6–14). Thus recommending more product items may not improve the quality of the recommendation for less loyal customers.
4.3.4. Comparing weighted RFM with typical CF method
Experiments were conducted to compare the weighted RFM method with the typical CF method. The typical CF method has been widely used and is a
representative recommendation method. The method uses product purchase preferences to compute simi-larity among customers, and then employs the k-nearest neighbor (k-NN) approach to derive top-N recommendations. Table 8 lists the overall average F1 metrics of weighted RFM and the typical CF method, respectively, for different k and N. From
Table 8, the F1 metrics of weighted RFM exceeded those of the typical CF method. This result indicated that the proposed method provided better recommen-dations.
An RFM-based k-nearest-neighbor method was used to evaluate its effect on recommendation quality. The method resembles the typical CF method that
Fig. 2. Comparisons under various top-N.
selected k-nearest neighbors to obtain top-N recom-mendations. However, the RFM-based k-NN method used the weighted RFM values of customers to com-pute the similarity measures between customers rather than using product purchase preferences.Table 9lists the experimental result, and shows the F1 metrics of the RFM-based k-NN method and the typical CF method. The RFM-based k-NN method performed better than the typical CF method. The relative impor-tance of RFM variables contributed to improving product recommendation quality.
4.3.5. Experiments on three clusters of customers Experiments were also performed on placing cus-tomers into three clusters.Table 10andFig. 4show the experimental results which exhibited trends similar to those of the experiments using eight clusters. The weighted RFM method outperformed the non-cluster-ing, non-weighted RFM and typical CF methods. The F1 metrics of the more highly ranked clusters exceeded those of the lower-ranked clusters. Further-more, recommending more items helped to increase the F1 metrics and improve the quality of
recommen-Table 8
F1 metrics for weighted RFM and typical CF method Top-N Weighted RFM Typical CF method
90-NN 100-NN 110-NN 130-NN 150-NN Top-4 0.333 0.285 0.286 0.291 0.300 0.296 Top-6 0.413 0.376 0.381 0.380 0.386 0.392 Top-10 0.499 0.484 0.487 0.488 0.491 0.491 Top-20 0.524 0.514 0.515 0.517 0.516 0.517 Top-30 0.504 0.497 0.498 0.498 0.501 0.503 Top-40 0.484 0.467 0.467 0.467 0.470 0.470 Top-50 0.477 0.422 0.422 0.422 0.424 0.425 Table 9
F1 metrics for RFM-based k-NN and typical CF method
Top-N Neighbors-90 Neighbors-100 Neighbors-110 Neighbors-130 Neighbors-150 RFM-based k-NN Typical CF RFM-based k-NN Typical CF RFM-based k-NN Typical CF RFM-based k-NN Typical CF RFM-based k-NN Typical CF Top-4 0.303 0.285 0.307 0.286 0.311 0.291 0.305 0.300 0.313 0.296 Top-6 0.393 0.376 0.404 0.381 0.409 0.380 0.410 0.386 0.410 0.392 Top-10 0.491 0.484 0.492 0.487 0.500 0.488 0.495 0.491 0.498 0.491 Top-20 0.520 0.514 0.520 0.515 0.516 0.517 0.520 0.516 0.519 0.517 Top-30 0.500 0.497 0.500 0.498 0.499 0.498 0.503 0.501 0.503 0.503 Top-40 0.470 0.467 0.470 0.467 0.470 0.467 0.470 0.470 0.472 0.470 Top-50 0.423 0.422 0.422 0.422 0.424 0.422 0.425 0.424 0.426 0.425 Table 10
F1 metrics of various methods for three clusters under top-30 and 110 nearest neighbors
CLV ranking Weighted RFM Non-clustering Non-weighted RFM Typical CF method
1 0.736 0.617 0.663 0.698
2 0.533 0.469 0.492 0.520
3 0.393 0.363 0.355 0.386
dation for clusters with a high CLV ranking. However, recommending more product items may not improve the quality of recommendation for customers of lower loyalty.
5. Conclusions
Our work involved the introduction of a novel recommendation methodology that combines AHP, clustering, and association rule-based methods. It clusters customers into segments according to their lifetime value expressed in terms of weighted RFM. Applying AHP to determine the relative importance of RFM variables proved important, since the RFM weights vary with the characteristics of product and industry. Moreover, clustering customers into different groups not only improves the quality of recommenda-tion but also helps decision-makers identify market segments more clearly and thus develop more effec-tive strategies. The experimental results show that the proposed methodology indeed can yield recommenda-tions of higher quality. However, the methodology is not effective for all customer groups. It is more effective for more loyal customers. Recommending more items helps to improve the quality of recom-mendation for more loyal customers, but may not do so for less loyal customers.
Acknowledgements
The authors gratefully acknowledge the Editor and anonymous reviewers for their valuable comments and constructive suggestions. This research was supported in part by the National Science Council of the Repub-lic of China under the grant NSC 92-2416-H-009-010.
Appendix A. Formalization of association rule mining
Agrawal et al. formalized the problem of finding association rules. Let I be a set of product items and D be a set of transactions, each of which includes a set of products that are purchased together. An association rule is an implication of the form: X) Y, where X I, Y I, and X \ Y ¼ F. X is the antecedent (body) and Y the consequent (head) of the rule. Two measures, support and confidence, are used to indicate the quality of an association rule. The support of a rule is the percentage of transactions that contain both X and Y, whereas the confidence of a rule is the fraction of transactions that contain X, that also con-tain Y.
Appendix B. Association rule based recommendation
Let Xube the set of products previously purchased by customer u. First find all the recommendation rules X) Y, for which X Xu; i.e. customer u purchased all the products in X. Then, for each extracted recom-mendation rule, all the products in Y that have not yet been purchased by customer u are candidate products for recommendation. Each candidate product is asso-ciated with the confidence of the corresponding recommendation rule. If the candidate product is associated with multiple rules, then the highest con-fidence is used. Let Pu be the set of such candidate products. The candidate products in Puare sorted by associated confidence value. Candidate products with higher confidence are ranked higher, and the N highest ranked candidate products are selected as the recom-mendation set.
Appendix C. Computing pearson correlation coefficient
Customer purchase history is represented as a cus-tomer-item matrix R such that, rij is one if the ith customer purchased the jth product; and is zero other-wise. The similarity of preferences among customers can then be measured by computing the Pearson correlation coefficient defined as:
corrðci; cjÞ ¼ P s2Iðrci;s rciÞðrcj;s rcjÞ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi P s2Iðrci;s rciÞ 2P s2Iðrcj;s rcjÞ 2 q
The notations rciand rcjdenote the average number of
products purchased by customers ci and cj, respec-tively. Moreover, the variable I denotes the set of pro-ducts. Additionally, the rci,sand rcj,sindicate whether customers ciand cjpurchased product item s.
References
[1] R. Agrawal, T. Imielinski, A. Swami, Mining association between sets of items in large database, in: Proceedings of the ACM-SIGMOD International Conference on Management of Data, Washington, DC, USA, 1993, pp. 207–216.
[2] R. Agrawal, R. Srikant, Fast algorithms for mining associa-tion rules, in: Proceedings of the 20th Internaassocia-tional Conference on very large Data Bases, Santiago, Chile, 1994, pp. 407–419.
[3] P. Berger, N. Nasr, Customer lifetime value: marketing models and applications, Journal of Interactive Marketing 12 (1), 1998, pp. 17–30.
[4] I. Bose, R.K. Mahapatra, Business data mining-a machine learning perspective, Information and Management 39 (3), 2001, pp. 211–225.
[5] R.J. Brachman, T. Khabaza, W. Kloesgen, G. Piatetsky-Shapiro, E. Simoudis, Mining business databases, Commu-nications of the ACM 39 (11), 1996, pp. 42–48.
[6] J.R. Bult, T.J. Wansbeek, Optimal selection for direct mail, Marketing Science 14 (4), 1995, pp. 378–394.
[7] S.W. Changchien, Z.C. Lu, Mining association rules procedure to support on-line recommendation by customers and products fragmentation, Expert Systems with Applica-tions 20 (4), 2001, pp. 325–335.
[8] H.C. Chen, A.L.P. Chen, A music recommendation system based on music data grouping and user interests, in: Proceedings of the ACM International Conference on Information and Knowledge Management, Atlanta, GA, November 2001, pp. 231–238.
[9] M.S. Chen, J. Han, P.S. Yu, Data mining: an overview from a database perspective, IEEE Transactions on Knowledge and Data Engineering 8 (6), 1996, pp. 866–883.
[10] J. Goodman, Leveraging the customer database to your com-petitive advantage, Direct Marketing 55 (8), 1992, pp. 26–27. [11] S.H. Ha, S.C. Park, Application of data mining tools to hotel data mart on the Intranet for database marketing, Expert Systems with Applications 15 (1), 1998, pp. 1–31. [12] W. Hill, L. Stead, M. Rosenstein, G. Furnas, Recommending
and evaluating choices in a virtual community of use, in: Proceedings of ACM CHI’95 Conference on Human Factors in Computing Systems, Denver, CO, 1995, pp. 194–201. [13] A.M. Hughes, Strategic Database Marketing, Probus
Publish-ing, Chicago, 1994.
[14] S.C. Hui, G. Jha, Data mining for customer service support, Information and Management 38 (1), 2000, pp. 1–13. [15] S. Irvin, Using lifetime value analysis for selecting new
customers, Credit World 82 (3), 1994, pp. 37–40.
[16] R. Kahan, Using database marketing techniques to enhance your one-to-one marketing initiatives, Journal of Consumer Marketing 15 (5), 1998, pp. 491–493.
[17] K. Lang, NewsWeeder: Learning to Filter Netnews, in: Proceedings of the 12th International Conference on Machine Learning, Tahoe City, CA, 1995, pp. 331–339.
[18] T.P. Liang, H.J. Lai, Effect of store design on consumer purchases: an empirical study of on-line bookstores, Informa-tion and Management 39 (6), 2002, pp. 431–444.
[19] Q.Y. Lin, Y.L. Chen, J.S. Chen, Y.C. Chen, Mining inter-organizational retailing knowledge for an alliance formed by competitive firms, Information and Management 40 (5), 2003, pp. 431–442.
[20] G. Linden, B. Smith, J. York, Amazon.com recommenda-tions: item-to-item collaborative filtering, Internet Computing IEEE 7 (1), 2003, pp. 76–80.
[21] C. Liu, K.P. Arnett, Exploring the factors associated with Web site success in the context of electronic commerce, Information and Management 38 (1), 2000, pp. 23–33. [22] J.B. MacQueen, Some methods for classification and analysis
of multivariate observations, Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Prob-ability 1, 1967, pp. 281–296.
[23] J. Miglautsch, Thoughts on RFM scoring, Journal of Database Marketing 8 (1), 2000, pp. 67–72.
[24] D. Peppers, M. Rogers, The One to One Future: Building Relationships One Customer at a Time, Bantam Doubleday Dell Publishing, 1997.
[25] G.N. Punj, D.W. Stewart, Cluster analysis in marketing research: review and suggestions for application, Journal of Marketing Research 20, 1983, pp. 134–148.
[26] P. Resnick, N. Iacovou, M. Suchak, P. Bergstrom, J. Riedl, GroupLens: an open architecture for collaborative filtering of Netnews, in: Proceedings of the CSCW’94, Chapel Hill, NC, 1994, pp. 175–186.
[27] J. Rucker, M.J. Polanco, Siteseer: personalized navigation for the Web, Communications of the ACM 40 (3), 1997, pp. 73–75. [28] T.L. Saaty, Fundamentals of Decision Making and Priority Theory with the Analytic Hierarchy Process, RWS Publica-tions, Pittsburgh, PA, 1994.
[29] G. Salton, M.J. McGill, Introduction to Modern Information Retrieval, McGraw-Hill, New York, 1983.
[30] B. Sarwar, G. Karypis, J. Konstan, J. Riedl, Analysis of recommendation algorithms for e-commerce, in: Proceedings of the Second ACM Conference on Electronic Commerce, Minneapolis, October 2000, pp. 158–167.
[31] J.B. Schafer, J.A. Konstan, J. Riedl, E-commerce recommen-dation applications, Journal of Data Mining and Knowledge Discovery 5 (1–2), 2001, pp. 115–152.
[32] U. Shardanand, P. Maes, Social information filtering: algorithms for automating ‘world of mouth’, in: Proceedings of ACM CHI’95 Conference on Human Factors in Computing Systems, Denver, CO, 1995, pp. 210–217.
[33] M.J. Shaw, C. Subramaniam, G.W. Tan, M.E. Welge, Knowledge management and data mining for marketing, Decision Support Systems 31 (1), 2001, pp. 127–137. [34] R. Srikant, R. Agrawal, Mining generalized association
rules, in: Proceedings of the 21st International Conference on Very Large Data Bases, Zurich, Switzerland, 1995, pp. 407–419.
[35] B. Stone, Successful Direct Marketing Methods, Lincoln-wood, NTC Business Books, IL, 1995.
[36] R.W. Stone, D.J. Good, The assimilation of computer-aided marketing activities, Information and Management 38 (7), 2001, pp. 437–447.
[37] C.J. Van Rijsbergen, Information Retrieval, second ed., Butterworths, London, 1979.
[38] J.D. Wells, W.L. Fuerst, J. Choobineh, Managing information technology (IT) for one-to-one customer interaction, Informa-tion and Management 35 (1), 1999, pp. 53–62.
Duen-Ren Liu is a professor of the Institute of Information Management, National Chiao Tung University, Taiwan. He received the BS and MS degrees in Computer Science and Information En-gineering from the National Taiwan Uni-versity, Taiwan, in 1985 and 1987, respectively. He received the PhD degree in Computer Science from the University of Minnesota in 1995. His research interests include information systems, electronic commerce, workflow systems, and knowledge manage-ment. Dr. Liu is an associate member of the IEEE, and a member of the ACM.
Ya-Yueh Shihis a PhD student of the Institute of Information Management, National Chiao Tung University. She is currently an instructor of the Department of Information Management, Ming Hsin University of Science and Technology. She received her BS and MS degrees in Department of Information Management form the National Yunlin University of Science and Technology, Taiwan, in 1996 and 1998, respectively. Her research interests include data mining, consumer behavior, and electronic commerce.
