Information processing apparatus, information processing method, and program
The information processing device improves customer loyalty analysis in electronic commerce by identifying core customers, clustering others based on distance, and using collaborative filtering to recommend products, thereby enhancing the accuracy and relevance of product recommendations.
Patent Information
- Application Number
- JP2023081048
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-05-16
- Publication Date
- 2024-11-28
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing customer loyalty analysis methods in electronic commerce struggle to accurately distinguish between customers based on their purchasing behavior, leading to inaccurate product recommendations.
An information processing device that identifies core customers by analyzing multiple characteristic variables, calculates their center of gravity in a feature space, and classifies other customers into clusters based on distance from this core, using collaborative filtering to recommend products tailored to each cluster, thereby improving analysis accuracy and product relevance.
Enhances the accuracy of customer purchase analysis and product recommendations, leading to improved customer loyalty by aligning product suggestions with individual purchasing behaviors.
Smart Images

Figure 2024165146000001_ABST
Abstract
Description
[Technical field]
[0001] The present invention relates to an information processing device, an information processing method, and a program for analyzing customers in electronic commerce. [Background technology]
[0002] In EC (Electronic Commerce) businesses, it is common to classify customers based on their loyalty and provide sales promotions to them that are suited to their classification. In addition, various techniques have been proposed for analyzing customer purchases in relation to customer loyalty (Non-Patent Documents 1 to 5).
[0003] Non-Patent Document 1 uses a method of defining three levels of customer rank, with customers who spend a lot being considered to have high customer loyalty. While this method is relatively simple, it is difficult to determine which customer has higher loyalty: customer A who buys a high-priced item worth 10,000 yen once a month, or customer B who buys a 2,500 yen item every week (4 times a month) for a total monthly purchase amount of 10,000 yen.
[0004] In addition, Non-Patent Document 2 proposes a method of classifying customers by RFM analysis using purchase frequency (Frequency) and time elapsed since the last purchase (Recency) in addition to the purchase amount (Monetary). The variables R, F, and M are characteristic variables commonly used in customer purchasing behavior analysis, and are considered to be effective in classifying customers according to the characteristics of their purchasing behavior. However, for example, in addition to the above-mentioned customer B, there is a customer C who purchases different products every week (purchases four times a month) and the total purchase amount for the month is 10,000 yen, and for both customers B and C, if the most recent purchase date was 7 days ago, as shown in Table 1, in the RFM analysis, customers B and C are both (R, F, M) = (7, 4, 10,000), and cannot be distinguished. In Table 1, the Recency column indicates how many days ago the most recent purchase date was, the Frequency column indicates the number of purchases per month, and the Monetary column indicates the total purchase amount (yen) for the month. [Table 1]
[0005] Therefore, Non-Patent Documents 3 to 5 propose techniques for extending RFM analysis when the above RFM analysis is not sufficient for analysis. For example, Non-Patent Document 3 proposes an extension of RFM analysis by adding customer duration to RFM as a variable in bank customer segmentation, Non-Patent Document 4 proposes adding purchased product items to RFM as a variable, and Non-Patent Document 5 proposes an extension of RFM analysis by adding membership period length and profits to RFM as variables. However, in these techniques, the correlation with each of the R, F, and M variables differs depending on the variables added to the RFM analysis, so adding variables may not change the accuracy of the analysis, or the analysis results may be affected more than necessary by the added variables, resulting in a decrease in the accuracy of the analysis. [Prior art documents] [Non-patent literature]
[0006] [Non-Patent Document 1] Hisatetsu Iizuka, Daisuke Yonemura, and Hideki Toyoda. Behavioral analysis by customer rank (in japanese). Operations Research: The Science of Management, 48:94-99, 2003. [Non-Patent Document 2] Ching-Hsue Cheng and You-Shyang Chen. Classifying the segmentation of customer value via rfm model and rs theory. Expert systems with applications, 36(3):4176-4184, 2009. [Non-Patent Document 3] Mehdi Bizhani and Mohammad Jafar Tarokh. Behavioral segmentation of bank's point-of-sales using rf* m* approach. In Proceedings of the 2010 IEEE 6th International Conference on Intelligent Computer Communication and Processing, pages 81-86. IEEE, 2010. [Non-Patent Document 4] Takamasa Tanaka, Tomohiro Hamaguchi, Takumi Saigo, and Kazuhiko Tsuda. Classifying and understanding prospective customers via heterogeneity of supermarket stores. Procedia computer science, 112:956-964, 2017. [Non-Patent Document 5] Kuanchin Chen, Ya-Han Hu, and Yi-Cheng Hsieh. Predicting customer churn from valuable b2b customers in the logistics industry: a case study. Information Systems and e-Business Management, 13:475-494, 2015. Summary of the Invention [Problem to be solved by the invention]
[0007] As described above, definitions of customer loyalty are diverse and complex, and it has been difficult to accurately analyze customer purchases in relation to customer loyalty.
[0008] Therefore, the technology disclosed herein has been made in consideration of the above circumstances, and its purpose is to provide technology that improves the accuracy of analyzing customer purchases and recommends more appropriate products to customers. [Means for solving the problem]
[0009] The information processing device disclosed herein includes an acquisition unit that acquires multiple feature variables that are factors that characterize the purchasing behavior of customers in electronic commerce, an identification unit that identifies core customers among the customers for whom the multiple feature variables satisfy a predetermined condition, a calculation unit that calculates the center of gravity of the core customers in a feature space, a classification unit that classifies the other customers into clusters according to the distance between the center of gravity of the core customer and the coordinates of each of the other customers other than the core customer in the feature space, and a determination unit that determines, for each cluster, recommended products for each customer in the cluster based on the purchased products of the customers in the cluster.As a result, by dividing customers into clusters according to the distance from the core customer in the feature space and determining recommended products for each cluster, it is possible to recommend products that are more suitable for customers while reducing the processing load associated with the determination of recommended products in the information processing device.
[0010] In the above information processing device, the predetermined condition is that at least one of the following is a predetermined range of values: the number of purchases of products handled in the electronic commerce by the customer, the number of product categories of products purchased by the customer, the purchase frequency of products handled in the electronic commerce by the customer, and the purchase amount of products handled in the electronic commerce by the customer. This is expected to enable the determination of recommended products that more strongly reflects the purchasing behavior of the customer.
[0011] In the information processing device, the classification unit may classify the other customers into clusters such that the number of customers included in each cluster is equal. Also, the determination unit may apply collaborative filtering to products purchased by each customer in each cluster for each cluster. and determine a recommended product for each customer in the cluster. The determination unit may determine, for each customer, a predetermined number of products with top scores among products whose scores obtained by collaborative filtering are equal to or higher than a predetermined threshold and which the customer has not purchased, as the recommended products. This makes it possible to recommend products that are more appropriate for the customer based on the customer's purchasing behavior. The identification unit may identify the core customers for each type of e-commerce business, and the classification unit may classify the other customers into clusters for each type of business. Furthermore, the determination unit may determine specialized products in the business as the recommended products. This makes it possible to perform clustering of appropriate customers for each type of e-commerce business, and to determine recommended products targeting specialized products in the business.
[0012] The information processing method disclosed herein also includes an information processing method characterized in that a computer executes the steps of acquiring a plurality of feature variables which are factors characterizing the purchasing behavior of e-commerce customers, identifying core customers among the customers for whom the plurality of feature variables satisfy predetermined conditions, calculating the center of gravity of the core customers in a feature space, classifying the other customers into clusters in accordance with the distance between the center of gravity of the core customer and the coordinates of each of the other customers other than the core customer in the feature space, and determining, for each cluster, recommended products for each customer in the cluster based on the products purchased by the customers in the cluster.
[0013] In addition, the program disclosed herein includes a program for causing a computer to function as an acquisition unit that acquires multiple feature variables that are factors that characterize the purchasing behavior of e-commerce customers, an identification unit that identifies core customers among the customers for whom the multiple feature variables satisfy specified conditions, a calculation unit that calculates the center of gravity of the core customers in a feature space, a classification unit that classifies the other customers into clusters in accordance with the distance between the center of gravity of the core customer and the coordinates of each of the other customers other than the core customer in the feature space, and a determination unit that determines, for each cluster, recommended products for each customer in the cluster based on the products purchased by the customers in the cluster. Effect of the Invention
[0014] According to the technology disclosed herein, it is possible to improve customer loyalty by improving the accuracy of analysis of customer purchases and recommending more appropriate products to customers. [Brief description of the drawings]
[0015] [Figure 1] FIG. 1 is a diagram illustrating an example of an information processing device according to the first embodiment. [Diagram 2] FIG. 2 is a diagram showing a schematic diagram of correlations between customers in the feature amount space of feature variables. [Diagram 3] FIG. 3 is a flowchart illustrating an example of processing executed by the information processing device according to the first embodiment. [Figure 4] FIG. 4 is a diagram illustrating an example of a customer clustering algorithm executed by the information processing device according to the first embodiment. [Diagram 5] Figure 5 is a graph showing an assumed example of the correlation between core customer distance and LTV. [Figure 6] Figure 6 is a graph showing an example of the correlation between core customer distance and LTV in the medical sector. [Figure 7] Figure 7 is a graph showing an example of the correlation between core customer distance and LTV in the MRO sector. [Figure 8] Figure 8 is a graph showing an example of the correlation between core customer distance and LTV in the office sector. [Figure 9] FIG. 9 is a diagram showing an example of a collaborative filtering algorithm for determining recommended products executed by the information processing device according to the first embodiment. [Figure 10] FIG. 10 is a graph showing an example of a clustering result of the medical sector in the first embodiment. [Figure 11] FIG. 11 is a graph showing an example of a clustering result of the MRO sector in the first embodiment. [Figure 12]FIG. 12 is a graph showing an example of a clustering result of the office sector in the first embodiment. [Figure 13] FIG. 13 is a diagram showing an example of a simulation result when a specialized product is recommended in the medical sector. [Figure 14] FIG. 14 is a diagram showing an example of a simulation result when a general product is recommended in the medical sector. [Figure 15] FIG. 15 shows an example of a simulation result when a specialized product is recommended in the MRO sector. [Figure 16] FIG. 16 is a diagram showing an example of a simulation result when a general commodity is recommended in the MRO sector. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0016] Preferred embodiments of the technology disclosed herein will be described below with reference to the drawings. However, the configurations of the components described below should be appropriately changed depending on the configuration of the device to which the technology disclosed herein is applied and various conditions. Therefore, it is not intended to limit the technical scope of the technology disclosed herein to the following description.
[0017] (First embodiment) Fig. 1 is a diagram showing a schematic configuration of an information processing device according to the first embodiment. As shown in Fig. 1, the information processing device 1 has a control unit 11, a storage unit 12, an operation unit 13, a display unit 14, and a communication unit 15. In addition, the information processing device 1 is connected to a customer information DB (Database) 2, which is an external device.
[0018] The control unit 11 is realized by, for example, a CPU (Central Processing Unit). The control unit 11 acquires information used to execute the processing described below from the customer information DB2. The storage unit 12 stores programs and information used in the processing executed by the information processing device 1. In this embodiment, the storage unit 12 is realized by, for example, a flash memory. The operation unit 13 accepts input related to the processing executed by the information processing device 1 from a user of the information processing device 1. The operation unit 13 is realized by, for example, a touch panel, a camera, a microphone, etc., and accepts image input, manual input by the user, voice input, etc. as input information. The display unit 14 displays information related to the processing executed by the information processing device 1. The display unit 14 is realized by, for example, a liquid crystal display. Note that the information input to the operation unit 13 may be displayed on the display unit 14 by the control unit 11. The communication unit 15 communicates with an external device of the information processing device 1 and acquires information used in the processing executed by the information processing device 1. The communication unit 15 is realized by, for example, an antenna.
[0019] In the information processing device 1 of this embodiment, based on the customer information of customers in e-commerce acquired from the customer information DB2, in addition to the variables R (Recency; time elapsed since the last purchase), F (Frequency; purchase frequency), and M (Monetary; purchase amount) in the RFM analysis, multiple types of variables (e.g., tens to hundreds of types) that are thought to represent the characteristics of a customer's purchasing behavior, such as product type, membership period, and order price, are used to perform a process of mapping customers into a feature space using purchasing behavior as a feature variable.
[0020] In this embodiment, it is assumed that customers whose specific feature variables exceed a threshold value set for each feature variable are likely to have high customer loyalty, and this group of customers is called "core customers." In addition, the center of gravity coordinates of the core customers are calculated in the feature amount space of the feature variables of purchasing behavior, and the distance from the center of gravity coordinates of the core customers to the coordinates of each customer (also called "core customer distance") is calculated. In addition, customers other than the core customers in the customer group are called "general customers."
[0021] An example of the positional relationship between a core customer and other customers in the feature space of feature variables of purchasing behavior is shown in Fig. 2. Fig. 2 illustrates a three-dimensional feature space using feature variables x1, x2, and x3, but the number of feature variables constituting the feature space in this embodiment is not limited to this, and an n-dimensional feature space (n is the number of feature variables) can be used.
[0022] In the figure, points in the feature space represent customers, and a group of customers whose feature variables x1, x2, and x3 exceed their respective thresholds can be identified as core customer C. In addition, the core customer distances from the center of gravity G of core customer C in the feature space to non-core customers A and B are defined as LA and LB, respectively.
[0023] In this embodiment, a decile analysis is performed to divide general customers into 10 equal parts according to the core customer distance. Each collection of general customers obtained by this analysis is called a "cluster." Note that the number of clusters divided by this decile analysis is not limited to 10, and may be any number of clusters, and the number of general customers does not necessarily have to be equal in each cluster, and may differ for each cluster.
[0024] In addition, in this embodiment, using clusters classified based on the distance from the center of gravity coordinates of core customers to the coordinates of general customers, a product suitable for recommendation to each general customer is determined for each cluster using collaborative filtering. Since general customers belonging to the same cluster are likely to have similar purchasing behavior, it can be expected that each general customer in the cluster is highly likely to purchase the determined product.
[0025] 3 shows a flowchart of the process executed in the information processing device 1 according to this embodiment. The process described below is started by the control unit 11 of the information processing device 1 based on an input from the operation unit 13 by the user, for example.
[0026] In step S101, the control unit 11 acquires information on the above-mentioned several dozen kinds of characteristic variables, which are factors that characterize the purchasing behavior of each customer of the electronic commerce, from the customer information DB2. Next, in step S102, the control unit 11 judges whether or not some of the characteristic variables satisfy a predetermined condition that is determined in advance for each customer. Then, the control unit 11 specifies a group of customers in which all of the some of the characteristic variables exceed a threshold as a core customer. The number of core customers may be one or more. Here, the some of the characteristic variables are characteristic variables that are specified in advance. Here, the predetermined condition includes a condition that at least one of the number of purchases of products handled in the electronic commerce by the customer, the number of product categories of products purchased by the customer, the purchase frequency of products handled in the electronic commerce by the customer, and the purchase amount of products handled in the electronic commerce by the customer is within a predetermined range. Examples of the predetermined range include being equal to or greater than a predetermined threshold and / or being equal to or less than a predetermined threshold.
[0027] In step S102, the control unit 11 may specify as a core customer a customer having a predetermined number or more of feature variables that exceed a threshold among some or all of the feature variables. Alternatively, the control unit 11 may specify as a core customer a customer for which each feature variable included in a combination of a predetermined number of feature variables among the some of the feature variables satisfies a threshold condition. Here, examples of the threshold condition include conditions for determining that a plurality of feature variables have a specific relationship, such as "feature variable A exceeds threshold value ThA" or "feature variable B exceeds threshold value ThB and feature variable C exceeds threshold value ThC" when some of the feature variables include feature variables A, B, and C.
[0028] Next, in step S103, the control unit 11 calculates the barycentric coordinates of the core customers in the feature space using the feature variables of each customer constituting the core customers identified in step S103. Then, in step S104, the control unit 11 calculates the distances (core customer distances) between the barycentric coordinates of the core customers calculated in step S103 and each of the general customers. The core customer distance is defined, for example, as the L2 norm (Euclidean distance) in the feature space, but may also be defined by the L1 norm (Manhattan distance), Lp norm (where p is an integer equal to or greater than 3), or the like.
[0029] Next, in step S105, the control unit 11 classifies the general customers into clusters according to the core customer distance calculated in step S104 by decile analysis. Specifically, the control unit 11 divides the core customer distance into 10 ranges so that the number of general customers included in each range is equal, and the general customers included in each range belong to each cluster. In addition, as an example, the control unit 11 assigns a label of cluster k (k=1 to 10) to each of the clusters divided by the decile analysis. Here, cluster 1 is the cluster of general customers closest to the centroid coordinates of the core customers, and the larger the value of k, the farther away it is from the centroid coordinates of the core customers, and cluster 10 is the cluster of general customers farthest from the centroid coordinates of the core customers.
[0030] The control unit 11 that executes the process of step S101 above corresponds to an acquisition unit of the present invention. The control unit 11 that executes the process of step S102 above corresponds to an identification unit of the present invention. The control unit 11 that executes the process of step S103 above corresponds to a calculation unit of the present invention. The control unit 11 that executes the process of step S105 above corresponds to a classification unit of the present invention.
[0031] FIG. 4 shows an example of an outline of an algorithm for performing customer clustering based on characteristic variables of purchasing behavior, which realizes the processing of steps S101 to S105.
[0032] The control unit 11 first acquires feature variables thought to characterize the purchasing behavior of a customer. Here, only quantitative variables are targeted, not categorical variables. The control unit 11 then creates a data frame df of the values of the feature variables of purchasing behavior, using the customer code as a unique key, based on the customer information acquired from the customer information DB2. The control unit 11 then clusters general customers using the data frame df as input data.
[0033] The algorithm in FIG. 4 will be described in detail. In the first line (similar to "1:" in the figure), the input data frame df is standardized by z-score normalization, which scales the input data frame df to a mean of 0 and a variance of 1, and the matrix of the general customers and the feature variables obtained is designated as matrix X. [CoreCondition] shown in the second line is a condition for extracting a group of core customers. More specifically, [CoreCondition] selects predetermined variables from among the feature variables of purchasing behavior, assigns threshold conditions to each variable, and connects them with an AND condition. As a guideline, the group of core customers should be about 1% of the total number of customers to be clustered. The variables used here can be selected in advance by the user of the information processing device 1 before executing this algorithm.
[0034] In lines 3 and 4, the k-means method is used to calculate the centroid coordinates of the core customer group, with the number of clusters set to 1. Then, in lines 5 to 10, the L2 norm (Euclidean distance) of each general customer is calculated from the centroid coordinates of the core customers, and the calculated result is used as the core customer distance. Finally, in line 11, the core customer distance is sorted in ascending order, and the data is divided into 10 groups so that the number of customers is equal. A cluster ID from c1 to c10 is then assigned to each divided customer, and a cluster column is added to the data frame df.
[0035] Returning to FIG. 3, next in step S106, the control unit 11 determines a recommended product for the customer for each cluster classified as described above.
[0036] Here, we will explain the correlation between core customer distance and customer loyalty. Lifetime Value (LTV) is a commonly used indicator of customer loyalty. LTV is an indicator that shows the profits that an e-commerce customer brings from the start to the end of a transaction in a specified period of time, and is sometimes called Customer Lifetime Value (CLV). For example, LTV can be used to perform customer segmentation based on customer loyalty in a specified business field, such as the health and beauty field.
[0037] Here, we will explain the correlation between core customer distance and LTV when using LTV as an index of customer loyalty. LTV can be calculated by the following formula (1).
number
[0038] where S is the expected annual total contribution margin per customer, which is equal to revenue minus cost of goods sold, and c is the churn rate.
[0039] As shown in Figure 5, there is a negative correlation between core customer distance and LTV. It is believed that the shorter the core customer distance, that is, the closer the purchasing behavior of the core customer, the larger the LTV will be, and conversely, the longer the core customer distance, the smaller the LTV will be.
[0040] However, as a result of the applicant's investigation into the correlation between core customer distance and LTV, it was found that the actual correlation between core customer distance and LTV is not simple, as shown in Figure 5, and that a negative correlation is observed up to a certain core customer distance, as described above, but when the core customer distance is even larger, the correlation between core customer distance and LTV reverses to a positive correlation.
[0041] As an example, we will explain the correlation between order history and LTV of general customers when a company that operates an e-commerce logistics business aggregates data on products recommended to general customers and products purchased by general customers over a period of several months. Here, we will list three customer industry sectors: the medical sector, the MRO (Maintenance Repair and Operations) sector, and the general office sector.
[0042] Figures 6, 7, and 8 show the relationship between average core customer distance and average LTV in the medical, MRO, and office sectors, which are industries of e-commerce, respectively. In Figures 6, 7, and 8, the X-axis (horizontal axis) shows the result of min-max normalization of core customer distance, with a normalized scale of 0.0 to 1.0. Similarly, the Y-axis (vertical axis) shows the scale of LTV, which is normalized to 0.0 to 1.0.
[0043] As shown in Figures 6, 7, and 8, in the medical sector, LTV bottoms out in cluster c4, and the correlation between core customer distance and LTV reverses to a positive correlation from cluster c5 onwards. In the MRO sector, LTV bottoms out in cluster c5, and the correlation between core customer distance and LTV reverses to a positive correlation from cluster c6 onwards. In the office sector, LTV bottoms out in cluster c7, and the correlation between core customer distance and LTV reverses to a positive correlation from cluster c8 onwards.
[0044] Furthermore, as shown in Figures 6, 7, and 8, the correlation between core customer distance and LTV has inverted to a positive correlation, suggesting the possibility that there may be general customers with higher customer loyalty apart from general customers with short core customer distance. However, in order to identify such general customers with higher customer loyalty, it would be necessary to verify each combination of variables in order to determine which variables should be selected from the vast number of purchasing behavior characteristic variables and to identify how to determine the threshold for each variable, but such verification would be impractical due to the enormous number of combinations.
[0045] Therefore, in this embodiment, with the aim of bringing general customers closer to clusters with higher customer loyalty, in step S106, the control unit 11 performs collaborative filtering on the products purchased by each general customer in the cluster, and determines recommended products that are candidates for recommendations for each general customer in the cluster.
[0046] FIG. 9 shows an example of an outline of an algorithm for determining a recommended product by collaborative filtering, which realizes the process of step S106. The control unit 11 of the information processing device 1 selects a recommended product by collaborative filtering using a Non-negative Matrix Factorization (NMF) algorithm. NMF is an algorithm widely used in multivariate analysis, and is a process for approximately factorizing an n×m matrix X given an n×m matrix W and an r×m matrix H. Usually, r is selected to be smaller than m, and the original data matrix is compressed by making W and H smaller than the original matrix X. In addition, the control unit 11 executes the algorithm of FIG. 9 for each cluster, and determines a recommended product for each general customer in the cluster. Note that the matrix X in FIG. 9 is a matrix different from the matrix X in FIG. 4. In the following description, NMF is used as the collaborative filtering algorithm, but a recommended product for each general customer may be determined using an algorithm different from NMF.
[0047] The algorithm in Figure 9 will be explained in detail. In this algorithm, data on the products and quantities purchased by general customers during a certain period is converted into a matrix with general customers as rows and products as columns, and this is used as input data. In the first line, an NMF model is created with "n components=16". This parameter corresponds to r above. In this example, the parameter value is set to 16, but if the value is too large, the error due to approximation will be small but it will be difficult to extract products to be recommended, and if the value is too small, the error due to approximation will be large and the features will become unclear. Therefore, this value is adjusted to an appropriate value depending on the size of the number of general customers (m) and the number of products (n) for each cluster.
[0048] In lines 2 and 3, the factorized W and H are obtained. In line 4, the dot product of W and H is calculated, and the result is the matrix Xp. The numerical value obtained here is the "score." In line 5, the column names of matrix X, i.e., the array of product IDs, are obtained. Next, in line 6, the row names of matrix X, i.e., the array of customer IDs, are obtained.
[0049] Next, in line 7, "threshold=0.05" is the cutoff threshold when extracting recommended products for general customers. This threshold is adjusted according to the number of products identified as recommended products. Then, in lines 8 to 15, a loop process is executed to extract recommended products for each general customer. Specifically, the process is executed repeatedly until i ranges from 0 to the number of general customers (m). In line 11, the product score array for the i-th general customer[i] is passed to the "sort_by_score" function, and the recommended products for the i-th general customer[i] are sorted in descending order of score, and the sorted result is returned as items_i.
[0050] Next, in line 12, the extraction conditions for the recommended products are defined. Specifically, the conditions are defined for products that have not been purchased in the past but may be purchased in the future to be extracted as recommended products. As shown in FIG. 9, the extraction condition "condition" is set as a condition that the product is one that a general customer has not purchased in the past and the product score exceeds the cutoff threshold "threshold."
[0051] Then, in line 13, the extraction condition defined in line 12 is applied to determine the recommended product[i] for general customer[i]. In line 14, the recommended product is added to the result array user_items, which is a two-dimensional array of general customers and recommended products. After the above loop process is completed, the result array user_items is generated as output data.
[0052] This allows the control unit 11 to determine recommended products for each general customer by collaborative filtering for each cluster classified in step S105.
[0053] Returning to FIG. 3, next, in step S107, the control unit 11 determines whether or not each of the determined recommended products for each general customer is a specialty product. Then, in step S108, the control unit 11 determines the recommended products determined to be specialty products as products to be recommended to the customer. For example, the control unit 11 determines, from among the recommended products identified in step S106, a predetermined number of products that are specialty products and have high product scores as products to be recommended to the customer. Information on the products to be recommended determined in step S108 may be stored in the memory unit 12, and may be acquired by the control unit 11 at any timing. Note that the control unit 11 that executes the process of the above step S106 corresponds to the determination unit of the present invention.
[0054] Next, an outline of a specific processing example of extracting recommended products for customers by collaborative filtering for each cluster obtained by the above customer clustering in this embodiment will be described.
[0055] Here, we use the matrix UserItemMatrix(X) in which the number of customers is 3 (m=3) and the number of product types is 5 (n=5), as shown in the following formula (2). For simplicity, we assume that the number of products purchased by each customer is either 1 or 0.
number
[0056] When we factor X using NMF with r = 3, we obtain a 3 × 2 matrix W and a 2 × 5 matrix H (equation (3)) (lines 1 to 3 in Figure 9).
number
[0057] Specifically, W and H are calculated as shown in the following formula (4).
number
[0058] The dot product of W and H is calculated to obtain the result shown in the following equation (5) (fourth line in FIG. 9).
number
[0059] In the original matrix X, create a matrix in which elements with a value of 0 are set to 1 and elements with a value of 1 are set to 0, and call this matrix NOT SIGN X (equation (6)).
number
[0060] The logical product of the matrices obtained by formula (5) and formula (6) is the matrix indicating the recommended products (formula (7)). Note that the logical product with NOT SIGN X is taken in order to determine unpurchased products as recommended products, excluding products previously purchased by the customer.
number
[0061] Each row of the matrix in formula (7) represents the score of each product for each customer. Here, elements in each row that have a value equal to or greater than the cutoff threshold value are the recommended products for that customer. The above process essentially corresponds to the process in lines 5 to 15 of FIG. 9.
[0062] The products recommended to the customer can be a predetermined number of the recommended products that are specialized products and have the highest scores. For example, if the predetermined number is 1, then the specialized products with the highest scores among the recommended products will be recommended to the customer. Assuming that all the products shown in formula (7) are specialized products, it is determined that the product recommended to customer 1 (the customer in the first row) is product 4 (the product in the fourth column), the product recommended to customer 2 is either product 1 or product 3 (the selection criteria can be arbitrary), and the product recommended to customer 3 is product 5.
[0063] In addition, in the above step S106, the control unit 11 may determine the recommended products to be recommended to general customers by targeting only specialty products. Specifically, in the above formula (2), the matrix X is formed by only specialty products. In this case, the recommended products determined by formula (7) are always specialty products, so the control unit 11 can omit the process of step S107.
[0064] Next, we will explain an example of the results of a simulation of customer clustering and product recommendations for each of the three sectors mentioned above (medical, MRO, and office) using data collected by the company that operates the e-commerce logistics business mentioned above.
[0065] In this example, in order to improve the accuracy of the simulation, the following three criteria are set to narrow down the feature variables to be adopted from among the listed feature variables. (Criterion 1) The feature variables to be used should be quantitative only, excluding categorical variables. (Criterion 2) Exclude variables with missing values above a certain value (Criterion 3) Exclude variables with the same values
[0066] Table 2 below shows an example of the number of characteristic variables initially listed for each sector in the above example ("Number of initial variables" in the table) and the number of characteristic variables finally adopted ("Number adopted" in the table). [Table 2]
[0067] FIG. 10 shows an example of the clustering results for the medical sector. In FIG. 10, the X-axis shows clusters divided according to the core customer distance, and are arranged from c1 to c10 in order of proximity to the center of gravity coordinates of the core customers. The Y-axis shows scaling normalized by the min-max method. In the graph, the solid line 1001 shows the results of normalizing the average LTV value of customers belonging to each cluster using the min-max method ("norm_LTV_mean" in the figure). In FIG. 10, , the LTV value is a downward convex curve that is smallest near clusters c4 and c5. One possible reason for this type of curve is the influence of the number of specialty products purchased in the medical sector (n medical). In FIG. 10, the dotted line 1002 shows the result of normalizing the average of n medical ("norm_n_medical_mean" in the figure). As shown in FIG. 10, the LTV The correlation between the average of LTV and the average of n medical is stronger in clusters c4, c5, and onwards. On the other hand, the correlation between the average of LTV and the average of n medical is weaker in clusters c1 to c3.
[0068] Here, as another hypothesis, we assume that the membership period (duration) of each customer affects the LTV, and in Fig. 10, the results of normalizing the average customer duration for each cluster are shown by the dashed-dotted line 1003 ("norm_duration_mean" in the figure). As shown in Fig. 10, duration is the lowest in cluster c6, which is different from cluster c4, which has the lowest LTV, but considering the transition of duration in clusters c1 to c4, it can be considered that LTV is being pushed up by duration in these clusters.
[0069] Similarly, FIG. 11 shows an example of the clustering results for the MRO sector. The MRO sector is a product category that includes, for example, work supplies, packaging materials, disaster prevention and crime prevention supplies, tools, measuring instruments, and parts. In FIG. 11, as in FIG. 10, the solid line 1101 shows the result of normalizing the average LTV value of customers belonging to each cluster using the min-max method ("norm_LTV_mean" in the figure), and the dotted line 1102 shows the number of purchases of specialized materials in the MRO sector, n, mr The result of normalizing the mean of o is shown (in the figure, "norm_n_mro_mean"). The dashed line 1103 indicates The results of normalizing the average customer duration for each cluster are shown ("norm_duration_mean" in the figure). As shown in Figure 11, compared to the medical sector example in Figure 10, there are some differences, such as the LTV being the smallest in cluster c5 and the LTV being large in cluster c1, but the results are generally similar to those of the medical sector.
[0070] The medical and MRO sectors are more affected by the characteristics of specialized products specific to each industry than the office sector. In order to compare the above three sectors, an analysis similar to that in Figures 10 and 11 was also conducted on the office sector, which covers general office supplies purchased regardless of industry. Figure 12 shows an example of the results. In Figure 12, as in Figures 10 and 11, the solid line 1201 indicates the average LTV of customers belonging to each cluster. The results of normalizing the LTV score using the min-max method are shown in the figure ("norm_LTV_mean"). The dotted line 1202 shows the normalized average of the number of purchases of office supplies in the office sector ("norm_n_order_mean" in the figure), and the dashed line 1203 shows the average of customers in each cluster. The results show the normalized average of the duration of each cluster ("norm_duration_mean" in the figure). As shown in Figure 12, there are differences from the examples in the medical and MRO sectors, such as the fact that LTV is smallest in cluster c7 and that the correlation between membership duration and LTV is low from cluster c7 onwards. Also, in the case of the office sector, it can be seen that the closer a customer is to the center of gravity of core customers, the higher the LTV is for clusters c1 to c7.
[0071] Here, we focus on the clusters with the smallest LTV (called "bottoming clusters") in the medical sector and MRO sector. The bottoming cluster in the medical sector is cluster c4, and the bottoming cluster in the MRO sector is cluster c5. As LTV is an indicator of customer loyalty, customers belonging to the bottoming clusters are considered to be a customer group with low customer loyalty. Furthermore, since the bottoming cluster with the shortest membership duration is cluster c6 in both the medical sector and the MRO sector, it can be said that LTV, and therefore customer loyalty, does not necessarily depend solely on the shortness of the membership duration.
[0072] Therefore, taking into account the strength of the correlation between LTV and the number of specialized products purchased, we conducted a simulation to examine whether it is possible to move customers in a cluster with a low LTV to a cluster with a higher LTV by recommending specialized products to those customers. In the simulation, first, only specialized products are selected from the products extracted by NMF and recommended to the customer. Then, the above clustering is performed again while changing the probability that the customer will purchase the recommended product from 0.0 to 1.0 in increments of 0.1. In other words, this simulation makes it possible to investigate to what extent customers in the LTV-bottoming cluster will move to which cluster as the number of specialized product purchases increases.
[0073] FIG. 13 shows an example of the results of the above simulation for the medical sector. In FIG. 13, the X-axis is the purchase probability, the Y-axis on the left is the percentage of the number of customers who moved from cluster c4 to cluster c3 or cluster c5, and the Y-axis on the right is the percentage of the number of customers in cluster c4. Also, the dotted line 1301 is the percentage of the number of customers who moved from cluster c4 to cluster c3 ("moved to c3" in the figure), the dashed and dotted line 1302 is the percentage of the number of customers who moved from cluster c4 to cluster c5 ("moved to c5" in the figure), and the solid line 1303 shows the percentage of the number of customers who decreased in cluster c4 ("decrease in c4" in the figure). As shown in FIG. 13, it can be seen that the percentage of the number of customers who moved from cluster c4 to cluster c5 almost exceeds the percentage of the number of customers who moved from cluster c4 to cluster c3. Therefore, from the clustering simulation, it can be seen that the correlation between LTV and the number of specialized products n medical is stronger from cluster c5 onwards.
[0074] As a comparison with the example of the simulation result of FIG. 13, FIG. 14 shows an example of the simulation result when the same number of general products as the specialized products in the above simulation are recommended to customers in the medical sector. In the figure, the dotted line 1401 indicates the ratio of the number of customers who moved from cluster c4 to cluster c3 ("moved to c3" in the figure), the dashed line 1402 indicates the ratio of the number of customers who moved from cluster c4 to cluster c5 ("moved to c5" in the figure), and the solid line 1403 indicates the ratio of the number of customers who decreased in cluster c4 ("decreased number in c4" in the figure). As shown in FIG. 14, as in the case of FIG. 13, the movement of customers from cluster c4 to cluster c3 or c5 is seen, but it can be seen that the ratio of the number of customers who move between clusters is significantly smaller than the case of recommending specialized products in FIG. 13.
[0075] Similarly, FIG. 15 shows an example of the above simulation results for the MRO sector. Since the cluster c5 is the cluster where LTV bottoms out in the MRO sector, in the simulation, specialized products in the MRO sector are recommended to customers in cluster c5, and the number of customers who move from cluster c5 to cluster c4 ("moved to c4" in the figure) or cluster c6 ("moved to c6" in the figure) and the percentage of the number of customers who decrease in cluster c5 are tracked. In the figure, the dotted line 1501 indicates the percentage of the number of customers who move from cluster c5 to cluster c4 ("moved to c4" in the figure), the dashed and dotted line 1502 indicates the percentage of the number of customers who move from cluster c5 to cluster c6 ("moved to c6" in the figure), and the solid line 1503 indicates the percentage of the number of customers who decrease in cluster c5 ("decrease in c5" in the figure). As shown in FIG. 15, the simulation results were similar to the simulation results for the medical sector shown in FIG. 13.
[0076] In addition to the simulation shown in FIG. 15, a simulation is performed in the same manner as in FIG. 14, for the MRO sector, in which the same number of general products as the specialized products in the above simulation are recommended. FIG. 16 shows the simulation results. In the figure, the dotted line 1601 indicates the proportion of customers who moved from cluster c5 to cluster c4 ("moved to c4" in the figure), the dashed and dotted line 1602 indicates the proportion of customers who moved from cluster c5 to cluster c6 ("moved to c6" in the figure), and the solid line 1603 indicates the proportion of the decrease in the number of customers in cluster c5 ("decrease in c5" in the figure). As can be seen from FIG. 16, as in the case of FIG. 15, the cluster movement of customers is observed, but the proportion of the number of customers who move between clusters is significantly smaller than the case of recommending specialized products in FIG. 15.
[0077] Table 3 shows an example of the results of the above simulation in which the purchase rate of a recommended product when the recommended product is received by a customer is set to 100%. [Table 3]
[0078] As shown in Table 3, in simulations of recommending specialized products to customers in the medical and MRO sectors, a larger proportion of customers move to sectors with higher LTV than when general products are recommended to customers. Also, in the case of the medical sector, when moving from cluster c4, moving to cluster c5 ("c4 → c5") has a higher LTV than moving to cluster c3 ("c4 → c3"). Similarly, in the case of the MRO sector, when moving from cluster c5, moving to cluster c6 ("c5 → c6") has a higher LTV than moving to cluster c4 ("c5 → c4"). From this, it can be expected that recommending specialized products to customers in each sector will improve LTV.
[0079] From the above, as shown in Figures 10 to 12, the curves of core customer distance and LTV are downward convex, and the LTV of clusters farther away is greater than the LTV of clusters closer to the center of gravity of the core customer, which indicates that there may be core customers that are more suitable than the core customer. Furthermore, the above simulation suggests that in sectors where the specialization of the industries to be clustered is higher than in the general sector, such as the medical and MRO sectors, the number of purchases of specialized products is more likely to contribute to improving LTV, or customer loyalty.
[0080] Therefore, according to the information processing device 1 of this embodiment, by executing the process of the illustrated flowchart using the above algorithm, it is possible to identify core customers for each e-commerce industry and cluster customers based on their relationships with the core customers and their purchasing behavior characteristics. Furthermore, according to the information processing device 1, by performing collaborative filtering in each cluster, it is possible to recommend to customers specialized products that are expected to improve LTV and purchase probability as recommended products. As a result, the information processing device 1 can analyze customer purchasing behavior with higher accuracy than before and recommend products to customers that will improve customer loyalty. [Explanation of symbols]
[0081] 1. Information processing device 11 Control section 2 Customer information DB
Claims
1. an acquisition unit for acquiring a plurality of feature variables which are factors characterizing the purchasing behavior of customers in electronic commerce; an identification unit for identifying core customers among the customers, the core customers being characterized by the plurality of characteristic variables satisfying a predetermined condition; A calculation unit that calculates a center of gravity of the core customers in a feature space; a classification unit that classifies other customers into clusters according to a distance between the center of gravity of the core customer and coordinates of each of the other customers other than the core customer in the feature space; a determination unit that determines, for each of the clusters, a recommended product for each customer in the cluster based on products purchased by the customers in the cluster; 13. An information processing device comprising:
2. 2. The information processing device according to claim 1, wherein the specified condition is that at least one of the number of purchases of products handled in the electronic commerce by the customer, the number of product categories of products purchased by the customer, the purchase frequency of products handled in the electronic commerce by the customer, and the purchase amount of products handled in the electronic commerce by the customer is within a specified range of values.
3. The information processing apparatus according to claim 1 , wherein the classification unit classifies the other customers into clusters such that the number of customers included in each cluster is equal.
4. The information processing apparatus according to claim 1 , wherein the determination unit applies collaborative filtering to products purchased by each customer in the cluster for each of the clusters to determine recommended products for each customer in the cluster.
5. The information processing device according to claim 4, characterized in that the determination unit determines, for each customer, as the recommended products, a predetermined number of products with top scores obtained by collaborative filtering that are equal to or greater than a predetermined threshold and that have not been purchased by the customer.
6. The identification unit identifies the core customers for each type of electronic commerce industry, The classification unit classifies the other customers into clusters for each business type.
2. The information processing apparatus according to claim 1,
7. The determination unit determines a specialty product in the industry as the recommended product.
7. The information processing apparatus according to claim 6,
8. The computer A step of acquiring a plurality of characteristic variables which are factors characterizing the purchasing behavior of an electronic commerce customer; Identifying core customers among the customers, each of the plurality of characteristic variables satisfying a predetermined condition; Calculating a center of gravity of the core customers in a feature space; classifying the other customers into clusters according to a distance between the center of gravity of the core customer and coordinates of each of the other customers other than the core customer in the feature space; determining, for each cluster, recommended products for each customer in the cluster based on products purchased by the customers in the cluster; 2. An information processing method comprising:
9. Computer, an acquisition unit for acquiring a plurality of feature variables which are factors characterizing the purchasing behavior of customers in electronic commerce; an identification unit for identifying core customers among the customers, the core customers being characterized by the plurality of characteristic variables satisfying a predetermined condition; A calculation unit that calculates a center of gravity of the core customers in a feature space; a classification unit that classifies other customers into clusters according to a distance between the center of gravity of the core customer and coordinates of each of the other customers other than the core customer in the feature space; A decision unit for deciding, for each of the clusters, recommended products for each customer in the cluster based on products purchased by the customers in the cluster. A program to function as a
