Using the RFM (Recency, Frequency, Monetary value) model can provide valuable insights about customer clusters which is the core of customer relationship management. Due to accurate customer segment coming from dynamic weighted applications, in-depth targeted marketing may also use type of dynamic weight of R, F and M as factors. In this paper, we present our dynamic weighted RFM approach which is intended to improve the performance of customer segmentation by using the factors and variations to attain dynamic weights. Our dynamic weight approach is a kind of Custom method in essential which roots in the understanding of the data set. Firstly, Analytic Hierarchy Process is used to calculate the subjective weight, then the entropy method is applied to calculate the objective weight. Finally, we use comprehensive integration weighting method to combine the subjective and objective weight to obtain the final weight of the index to calculate the individual user value and quantify the user value difference. The experiment shows that the dynamic weight we used in RFM model is vital, affects the customer segmentation performance positively. Also, this study indicates that customer segments containing dynamic weighted RFM scores bring about stronger and more accurate association rules for the understanding of customer behavior. At last, we discuss the limitations of RFM analysis.
In recent years, the number of online shopping users in China has shown rapid growth. CNNIC (China Internet Network Information Center) released the 47th China Internet Development Statistics Report in February 2021. The penetration rate was 70.4%, an increase of 5.9 percentage from March 2020. In 2020, Chinese online retail sales reached 11.76 trillion yuan, an increase of 10.9% over 2019. Among them, the online retail sales of physical goods were 9.76 trillion yuan, accounting for 24.9% of the total retail sales of consumer goods. As of December 2020, the number of online shopping users in China has reached 782 million, an increase of 72.15 million compared to March 2020, accounting for 79.1% of the total Internet users, and the number of people using the Internet for shopping in China has reached 782 million.
A large number of online shoppers bring great opportunities to the e-commerce industry, but also make the industry competition intensified [
The analysis of user purchase behavior has very important practical significance in social and economic development [
Although there are many methods for customer value calculation such as Share of Wallet, Past Customer Value, and Life Time Value, RFM is generally used to obtain customer values, which has been showing good flexibility and adaption to each business situation [
RFM analysis is an effective model to study customer value and customer profitability, which can improve the complexity of the customer value analysis model with the help of relatively few consumer behavior characteristics [
RFM model combines three different indicators to calculate customer’s value, and then segment them with it. The RFM values can be calculated in several ways. The common ways are Pareto 80/20 cut, Percentiles e.g., Quantiles and Custom.
Pareto principle is commonly referred to as the 80/20 rule. About 20% of customers contribute to 80% of the total revenue [
According to RFM model, the weight of each index used to identify customer value is the same. Studies have shown that proximity and frequency are good indicators of loyalty [
① Analytic Hierarchy Process
The Analytic Hierarchy Process (AHP) is a practical systematic analysis and decision-making method proposed by professor Thomas Seti from the University of Pittsburgh in the early 1970 s [
Step 1. Building hierarchy model.
Step 2. Building judgment matrix (pairwise comparison). The properties of the judgment matrix are as follows:
Step 3. Hierarchical single sort and its consistency test.
Step 4. Hierarchical total order and its consistency test. The scale table of the judgment matrix element
Scale | Definition and description |
---|---|
1 | Two elements are equally important to attributes |
3 | Comparing two elements, one element is slightly more important than the other |
5 | Comparing two elements, one element is obviously more important than the other |
7 | Comparing two elements, one element is highly more important than the other |
9 | Comparing two elements, one element is extremely more important than the other |
2, 4, 6, 8 | Indicates the need to compromise between two elements |
1/aij | An inverse comparison between the two elements |
② Entropy Method
The entropy method is an objective weighting method, which determines the index weight according to the information provided by the observed values of each index [
③ Comprehensive Integration Weighting Method
The comprehensive integration weighting method integrates the weights calculated by the subjective weighting method and the objective weighting method through a certain equation. The integrated results reflect the subjective intention of the evaluator and the objectivity of the evaluation data at the same time, avoiding the defect that the subjective weight is too subjective and the objective weight lacks subjective information [
Our dynamic weight approach is a kind of Custom method in essential which roots in the understanding of the data set. Therefore the first step aims to identify the requirements from data, of course, the best is big data. For example, we collect selling data of CD from CDnow website, January 1, 1997 to June 30, 1998. The data set contains 69660 purchase records. We randomly chose the date December 8, 2019 for index calculation. The calculation results are shown in
Recency is the number of days from the latest purchase date of the user to the date of data collection, that is R indicator value; Frequency is the total number of user consumption, i.e., F indicator value; Monetary is the total amount of user consumption, that is M indicator value. Exploratory analysis is made on the calculated results, and the program operation results are shown in
From
To distinguish the value differences of different users more precisely, Dynamic weights are assigned to R, F and M. Firstly, AHP is used to calculate the subjective weight. Then the entropy method is used to calculate the objective weight. Finally, the comprehensive integration weighting method is used to combine the subjective and objective weight to obtain the final weight of the index to calculate the individual user value and quantify the user value difference.
As for the three indexes of RFM, the importance of the indexes should be determined by pairwise comparison. This paper combines the characteristics of CD itself to make an artificial qualitative judgment: ① CD belongs to durable goods, it is different from daily necessities, after a purchase, users can use a CD for a long time, do not need to buy twice. ② For users who love CDs, the purchase of CDs is affected by many factors, such as the purchase preference of users, the date of CDs release date and so on. The buyback cycle is irregular, and users’ demand for CD is discontinuous. Therefore, for such a non-high-frequency, non-consumer industry, the impact of the most recent purchase date R on customer value is relatively little in the contribution to corporate profits, the customer consumption is the most intuitive reflection. So the total consumption amount M has the highest importance. And F focuses on measuring customer loyalty, the higher the loyalty, the greater the customer value, F will also account for a certain proportion. The final judgment matrix is shown in
R | F | M | |
---|---|---|---|
R | 1 | 1/5 | 1/7 |
F | 5 | 1 | 1/2 |
M | 7 | 2 | 1 |
After normalization of the judgment matrix, the vector T = (0.23, 0.99, 1.78) is obtained. After normalization of the vectors, the eigenvectors of the judgment matrix of R, F and M were obtained. W = (0.08, 0.33, 0.59) T. At this point, whether the weight of each element of the feature vector can become an index is decided by the consistency test. Therefore, in this paper, we at first calculate the maximum characteristic root of the judgment matrix. Let’s say that original matrix:
B is equal to
where
n is the order of the matrix, and the test coefficient:
Therefore, the subjective weights of R, F and M were determined to be 0.08, 0.33 and 0.59 respectively through the consistency test.
Since the date of the data set is December 8, 2019 in this paper, the R index value has been increasing, while the other two indexes are constant. To eliminate the impact of the order of magnitude, the data is first standardized as follows.
Using standardized data to calculate the objective weight of each indicator, the steps are as follows:
Step1. Calculate the proportion of the index in item j of the
Step2. Calculate the information entropy of the
Step3. Calculate the difference coefficient of item
Step4. Calculate the weight of each index :
Going through the above 4 steps, finally we get the objective weights of R, F and M are 0.19, 0.48 and 0.33 respectively.
Due to the strong subjective will to give weight to individuals, and the lack of objective analysis and judgment of the real situation, in this paper we combine the subjective and objective weight using the additive integration of the comprehensive integration method, to establish the final weight. Let
Where, k1 and k2 are the synthesis coefficients (k1, k2 > 0, and k1 + k2 = 1), which are used to measure the relative importance of subjective and objective weights. In general, the values of k1 and k2 can be determined by a mathematical model or according to the subjective preference of decision-makers. The calculation equations of k1 and k2 are as follows:
The final weights of R, F and M are 0.15, 0.42 and 0.43 respectively. According to the weights of the indicators, the value of a single customer calculated as follows.
Value score is the customer’s value score for sure, WR, WF and WM respectively represent the weight of R, F and M, and XR, XF and XM respectively represent R, F and M after standardization. The first 20% of users are screened in descending order of customer value, as shown in
By using the sales data from CDnow website, we calculated the value of each user. Then we applied the model and weighted different transaction data with differentiated indexes. The key to the establishment of dynamic weight is the establishment of a judgment matrix. In this paper, a dynamic weighted RFM model was applied to the Brazilian e-commerce transaction data, combining the characteristics of the region, data and trading products judgment index importance: ① According to foreign media reports, over the past few years Brazil has been plagued by a political and economic crisis, in the face of economic downward pressure, Brazil’s electricity is strong, it gains made tremendous contributions to the economic development of the whole Brazil; ② From the characteristics of the data itself, many users only consume once and purchase only one piece of goods. Therefore, the traditional RFM model cannot be used for general weighting; ③ Through the observation of the trading products, it is found that most of the products are household items and electronic products, such as health beauty, auto, computers accessories, housewares, etc., which are durable goods and the buyback cycle is irregular. The trade product screenshot is shown in
Therefore, we believe that the last purchase date and total consumption times have little impact on user value. So, the judgment matrix is shown in
R | F | M | |
---|---|---|---|
R | 1 | 1/2 | 1/3 |
F | 2 | 1 | 1/3 |
M | 3 | 3 | 1 |
According to the experimental results, after the dynamic weight is assigned to the traditional RFM model, the most valuable customers are not necessarily the important ones. At this point, the customer value combines the data characteristics to judge the importance of indexes, instead of classifying users in a general way, so the user value evaluation results are more reasonable.
As an effective tool for identifying customer value and customer profitability, the RFM model can easily and quickly identify the most valuable customers (Important value users) and the least valuable customers (General retained users). For the middle six categories of customers, the RFM model cannot effectively judge the category ranking of user value, because different businesses have different emphases on R, F and M, so they need to differentiate the weight of each index. The dynamic weight RFM model is built based on the traditional RFM model. Through the customer value calculation, the value difference of each type of user and even each user can be clarified to meet the requirements of dynamic adjustment of index weight of different businesses, to help enterprises select the best users for precision marketing. Experiments on different data in this paper show that the dynamic weight RFM model applies to the transaction data of different businesses, and the user value evaluation results are more scientific, effectively making up for the defects of the traditional RFM model.
Meanwhile, our research discovery the limitations of RFM analysis is obvious. RFM only uses historical data about customers, which results in RFM being unable to identify future customer behavior that predictive methods may be able. Using RFM modeling can provide valuable insights about customers. But it does not take into account many other factors about the customer. Customer demographics such as age, sex and ethnicity are not covered in RFM analysis either. In-depth targeted marketing may also use the type of item purchased or customer campaign responses as factors.