Information processing system, information processing method, learning model generation method, and computer program

Abstract

To provide an information processing system, an information processing method, a learning model generation method, and a computer program.SOLUTION: An information processing system 11 comprises a memory 42 storing a program code 45, and a processor 41 executing processing on the basis of a content of the program code 45. The processor 41 is configured to execute: classification processing for classifying a plurality of users into a plurality of segments; and estimation processing for estimating, from a plurality of attributes of the users, one or more transition attributes that are characteristic to a transitioning user who will transition to another segment as time elapses. The estimation processing includes input processing for inputting user data, including the plurality of attributes of the respective users 14, into a learning model 37U. The learning model 37U outputs a calculation result when the user data is input. The calculation result includes a result of prediction of whether each user 14 will become the transitioning user and the importance of each attribute in the prediction.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present disclosure relates to an information processing system, an information processing method, a learning model generation method, and a computer program. [Background technology]

[0002] In the field of marketing, potential users in a market are often classified into a number of groups (segments) based on their needs or characteristics. By implementing marketing measures according to the characteristics of each segment, it becomes possible to efficiently carry out sales promotion activities such as advertising. For example, Patent Document 1 discloses a method for sending an electronic message including segmented user data. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Special Publication No. 2018-536947 Summary of the Invention [Problem to be solved by the invention]

[0004] Each of the multiple users has various attributes, such as age, gender, educational background, or income level. If one or more attributes that are highly correlated with a target characteristic can be found from among these attributes, more appropriate segmentation becomes possible. For example, if multiple existing users are segmented based on the frequency and amount of use of a certain service, if a certain user's amount of use can be increased, the user will transition to a segment with a higher amount of use. Therefore, for example, if a common attribute can be found for users who have the characteristic of being "likely to transition to another segment," more effective sales promotion activities can be carried out.

[0005] The present disclosure aims to provide an information processing system, an information processing method, a method for generating a learning model, and a computer program capable of estimating one or more attributes that are highly correlated with a target characteristic from among multiple attributes possessed by a user. [Means for solving the problem]

[0006] An information processing system according to one embodiment of the present disclosure comprises at least one memory configured to store computer program code and at least one processor configured to execute processing based on the contents of the computer program code, wherein the processor is configured to execute a classification process for classifying a plurality of users into a plurality of segments, and an estimation process for estimating, from a plurality of attributes of the users, one or more transition attributes characteristic of a transition user who transitions to another segment over time, wherein the estimation process includes an input process for inputting user data including the plurality of attributes of each of the users into a learning model, and the learning model is configured to output a calculation result when the user data is input, and the calculation result includes a prediction result of whether the user will become the transition user and the importance of each of the attributes in the prediction.

[0007] An information processing method according to one embodiment of the present disclosure includes having one or more computers classify a plurality of users into a plurality of segments, and estimate, from a plurality of attributes of the users, one or more transition attributes characteristic of a transition user who transitions to another segment over time, wherein the estimation includes predicting whether the user will become the transition user, and calculating the importance of each of the attributes in the prediction.

[0008] In a method for generating a learning model according to one aspect of the present disclosure, when a plurality of users are classified into a plurality of segments, a user who transitions to another segment over time is a transition user, and the user data includes a plurality of corresponding attribute data for each of the plurality of users, and the learning model is configured to output, when the user data is input, a prediction result of whether the corresponding user will become the transition user and the importance of each attribute in the prediction, and the generation method includes obtaining training data and training a pre-training learning model using the training data, and the training data includes the user data and a correct answer label indicating whether each of the plurality of users included in the user data corresponds to the transition user.

[0009] A computer program according to one embodiment of the present disclosure includes causing one or more computers to execute the following steps: classify a plurality of users into a plurality of segments; and estimate, from a plurality of attributes of the users, one or more transition attributes that are characteristic of transition users who transition to another segment over time. The estimation includes predicting whether the user will become the transition user, and calculating the importance of each of the attributes in the prediction. [Brief description of the drawings]

[0010] [Figure 1] FIG. 1 is a schematic diagram for explaining an information processing system according to an embodiment. [Diagram 2] FIG. 2 is an example graph showing multiple segments and multiple transition paths. [Diagram 3] FIG. 3 is an example of a graph showing the number of months a user belonged to each of multiple segments. [Figure 4] FIG. 4 is an example of a graph showing the volume of each of a number of segments. [Diagram 5] FIG. 5 is another example of a graph showing multiple segments and multiple transition paths. [Figure 6]FIG. 6 shows an example of the result of outputting the importance of feature quantities as a SHAP value. [Figure 7] FIG. 7 is a table showing transition effect coefficients when a transition occurs between each of a plurality of segments. [Figure 8] FIG. 8 is a flow diagram illustrating an information processing method according to the embodiment. [Figure 9] FIG. 9 is a flow diagram illustrating a method for outputting a target group. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0011] Examples of the information processing system, information processing method, learning model generation method, and program of the present disclosure will be described below with reference to the drawings. The present invention is not limited to these examples, but is defined by the claims, and is intended to include all modifications within the meaning and scope of the claims.

[0012] [Information Processing System] FIG. 1 shows an example of an information processing system 11. The information processing system 11 includes one or more analysis devices 40. The information processing system 11 may include at least one of one or more service servers 20 and one or more information processing devices 30. The service server 20, the information processing device 30, and the analysis device 40 communicate with a plurality of user terminals 13 through a network 12. The service server 20, the information processing device 30, and the analysis device 40 may be realized by one or more servers or devices (e.g., computers) that realize their functions. When the service server 20, the information processing device 30, and the analysis device 40 are all realized by one device, for example, a computer, the information processing system 11 is realized by one device.

[0013] The network 12 includes, for example, the Internet, a Wide Area Network (WAN), a Local Area Network (LAN), a provider terminal, a wireless communication network, a wireless base station, a dedicated line, etc. The user terminal 13 is, for example, but is not limited to, a smartphone, a personal computer, a tablet, etc. Each user terminal 13 is operated by a user 14. The user 14 is a user of one or more services provided by a business operator.

[0014] Each service server 20 includes one or more processors 21, one or more memories 22, a communicator 23, and a communication bus 24 for interconnecting these. The communicator 23 is, for example, an interface for connecting to the network 12. The communicator 23 enables communication with other devices, for example, the user terminal 13 and the analysis device 40, via the network 12. The memory 22 stores an application 25 and a database 26 for providing the service.

[0015] When the information processing system 11 includes a plurality of service servers 20, the plurality of service servers 20 may provide different services. A database 26 including user information of the service provided by the corresponding service server 20 is stored in the memory 22 of each of the plurality of service servers 20. Personal information and service usage history of the user 14 who uses each service are included in the corresponding database 26. When a plurality of services are provided by the same or affiliated businesses, the plurality of databases 26 relating to the different services may be shared with each other.

[0016] The analysis device 40 is configured to perform various analyses mainly related to marketing measures. The analysis device 40 includes one or more processors 41, one or more memories 42, a communicator 43, and a communication bus 44 for connecting these to each other. The communicator 43 is, for example, an interface for connecting to the network 12. The communicator 43 enables communication with other devices, for example, the service server 20 and the information processing device 30, via the network 12.

[0017] The memory 42 may store a computer program code 45 for performing analysis (hereinafter simply referred to as "program code 45") and output results 46. The processor 41 reads the computer program code 45 stored in the memory 42 at a predetermined timing and executes processing based on the computer program code 45. That is, the processor 41 is configured to output various analysis results by operating as instructed by the program code 45. The processor 41 may also store the output results 46 in the memory 42 by processing based on the computer program code 45. The processor 41 may also be configured to store the output results 46 in the memory 42 by operating as instructed by the program code 45.

[0018] The information processing system 11 may include one or more analytical databases 15. The analytical database 15 may be provided in an information processing device such as a server device. When the information processing system 11 includes multiple service servers 20, the analytical database 15 may include some or all of the user data in the database 26 of each of the multiple service servers 20. The user data includes the usage history of one or more services of each user 14 and attribute data including multiple attributes of each user 14.

[0019] Analytical database 15 may include data generated by processing data contained in one or more databases 26. One or more analytical devices 40 and one or more information processing devices 30 may exchange data with analytical database 15 over network 12. Some or all of analytical database 15 may be temporarily or permanently stored in memory 42 of analytical device 40 or memory 32 of information processing device 30 for various uses, including analysis.

[0020] The attribute data includes static data items and dynamic data items related to the attributes of the user 14. The term "attribute" in this disclosure may refer to information (e.g., "age") including a combination of an "attribute" as a data item (e.g., "age") and an "attribute value" that is a value for the attribute (e.g., "age: 40 years old").

[0021] Static data items include, for example, but are not limited to, date of birth, place of origin, and date of first purchase (date of registration for the service). Dynamic data items may include name, age, address, family composition, annual income, occupation, hobbies, and information related to products or services of interest. The analysis database 15 may further include data items related to the behavioral history of the user 14. Data items related to the behavioral history include, for example, but are not limited to, service usage history, product purchase history, website browsing history, or search history.

[0022] The analysis device 40 is configured to perform various analyses based on the data included in the analytical database 15, mainly to optimize marketing measures. The analysis performed by the analysis device 40 may include four phases. The four phases are, for example, classification of users 14 (first phase), analysis of user attributes (second phase), expansion of target groups (third phase), and analysis of measure data (fourth phase).

[0023] Some of the first to fourth phases may be omitted or changed. For example, the fourth phase may be performed before the first phase, or the fourth phase and other phases may be performed alternately or simultaneously while changing the analysis conditions. Furthermore, the first to fourth phases may be performed repeatedly to optimize the measures. When the first to fourth phases are performed repeatedly, only some of the phases may be repeated, or the results of any one of the phases may be fed back to the other phases. The analysis results may be provided to the service provider managing the service server 20.

[0024] When the information processing system 11 includes a plurality of analysis devices 40, each analysis device 40 may perform an analysis of a different phase. For example, the information processing system 11 may include first to fourth analysis devices 40 corresponding to the first to fourth phases, respectively. The results of the analysis by the analysis device 40, data processed by the analysis, or data generated in association with the analysis may be stored as part of the analysis database 15. Furthermore, the analysis device 40 may use the data processed or generated in association with the analysis in another analysis or the next analysis.

[0025] The information processing device 30 is configured to generate machine learning models 37 (37U, 37T, 37E) for assisting the analysis. A machine learning model is also simply called a learning model. The information processing device 30 includes one or more processors 31, one or more memories 32, a communicator 33, and a communication bus 34 for connecting these to each other. The communicator 33 is, for example, an interface for connecting to the network 12. The communicator 33 enables communication with other devices, for example, an analysis device 40, via the network 12.

[0026] The memory 32 may store one or more learning programs 35 and training data 36 for machine learning. The memory 32 may further store one or more learning models 37 generated using the learning programs 35 and the training data 36. The trained learning models 37 may be stored in the memory 42 of the analysis device 40 or in the memory of another computer included in the information processing system 11 when used.

[0027] The training data 36 is used to update various weights when training the learning model 37. The training data 36 may include validation data and test data. The validation data is used, for example, for tuning hyperparameters. The test data is used, for example, to verify the accuracy of the trained learning model 37.

[0028] The processors 21, 31, and 41 each include an arithmetic unit such as a CPU, a GPU, and a TPU. The processors 21, 31, and 41 are processing circuits configured to execute various software processes. The processing circuits may include a dedicated hardware circuit (e.g., ASIC, etc.) that processes at least a part of the software processes. That is, the software processes may be executed by a processing circuitry that includes at least one of one or more software processing circuits and one or more dedicated hardware circuits.

[0029] The memories 22, 32, and 42 are computer-readable media. The memories 22 and 32 include non-transitory storage media such as, for example, a random access memory (RAM), a hard disk drive (HDD), a flash memory, and a read only memory (ROM). The processors 21, 31, and 41 execute a series of instructions included in a program stored in the memories 22, 32, and 42 in response to a given signal or in response to a predetermined condition being satisfied. The service server 20, the information processing device 30, and the analysis device 40 may use data stored in a memory of a device separate from each other, for example, a cloud server (not shown).

[0030] [Summary of analysis by phase] The main objective of the first phase is to classify users 14. More specifically, the analysis device 40 is configured to classify a plurality of users 14 in one or more services into a plurality of groups based on a plurality of indicators. The groups thus classified are also called segments, and classifying is also called segmenting. The plurality of segments may be set for a business domain related to one or more services.

[0031] The main purpose of the second phase is to analyze the attributes of users 14 belonging to each of the multiple segments. For example, each segment classified based on the usage of the service includes users 14 with various attributes. For more effective marketing activities, it is necessary to select users 14 to be targeted for marketing. As an example of analysis, the attributes of users 14 belonging to each segment may be analyzed and a group of users with common characteristics may be extracted. A group of users with common characteristics is also called a cluster, and classifying users 14 into one of multiple clusters is called clustering.

[0032] In the second phase, as part of the analysis, transition attributes for a transition user may be estimated. A transition user is a user 14 who transitions from one segment to another segment over time. A transition attribute is an attribute that is characteristic of a transition user among multiple attributes possessed by a user 14.

[0033] When a transition occurs from one segment to another among multiple segments, the transition route is also called a transition path. As data for distinguishing the transition users on each transition path from other users 14, for example, a data item may be created in which a label "0" is input to the transition users and a label "1" is input to the users 14 who have not made the transition.

[0034] The analysis device 40 may be configured to estimate transition attributes characteristic of transition users. Here, the transition path to be analyzed is called a target transition path, and the segment to be analyzed or marketed is called a target segment. The target segment is a source segment that is the starting point of the transition. The segment that is the end point of the transition is called a destination segment. A plurality of target transition paths or target segments may be set. Alternatively, an analysis may be performed on all segments and all transition paths, and the results may be compared.

[0035] The analysis device 40 may be configured to extract some users 14 having transition attributes as transition users from a group of users belonging to a target segment. A transition user has attributes that are the same as or similar to the transition attributes (including interests, concerns, purchasing tendencies, or price preferences) of a transition user. A transition user is a prospective user who is likely to transition to a better segment by encouraging the use of a service through some marketing measure. A transition user may be identified in the analysis or planning of a marketing measure to be implemented. In contrast, a transition user is a user 14 who has an actual history of transitions, and is identified in the analysis of past data.

[0036] The main purpose of the third phase is to expand the target group. The target group is a group of users 14 who are targets of marketing measures such as sales promotion activities. A plurality of target groups may be set. The target group may be set corresponding to one or more transition paths of one or more transition source segments. In addition, a plurality of different measures may be prepared for a plurality of users 14 included in one target group.

[0037] The target group may include at least some or all of the transitional users. "Expanding the target group" includes extracting additional users 14 to be included in the target group when the number of transitional users is insufficient. More specifically, one or more similar users are extracted from the remaining users 14 of the transition source segment from which one or more transitional users have been extracted, and are added to the target group. A similar user is a user 14 that has a combination of one or more transition attributes and similar attributes.

[0038] The number of users 14 (target users) included in the target group is set according to the scale or budget of the marketing initiative. The target group basically includes transitional users and similar users. The similar users are extracted to make up the number of the target group. Therefore, the number of similar users is the number of users in the target group minus the number of transitional users.

[0039] The main purpose of the fourth phase is to analyze the campaign data related to the campaign. In the fourth phase, the results of one or more marketing campaigns are assumed or estimated, and the campaign data showing the results is analyzed. The campaign data may include prediction data regarding which segment users 14 belonging to the target segment will transition to after the campaign is implemented.

[0040] The analysis device 40 may calculate the cost-effectiveness of a policy based on the policy data. When analyzing multiple target groups, the cost-effectiveness may be compared between them. Furthermore, the analysis result of the policy data may be fed back to other phases to plan the next policy.

[0041] [Data used for analysis] The analysis database 15 may include data items generated based on a plurality of behavioral indicators as an example of a service usage history. An example of a behavioral indicator is an indicator related to the usage history of one or more services provided by the service server 20. The behavioral indicator may be set for each service, or may include a behavioral indicator related to usage across a plurality of services (cross-use). In the present disclosure, an indicator used for segmentation is called a segment indicator. Various behavioral indicators can be used as a segment indicator.

[0042] The one or more services may include, but are not limited to, one or more of a point program service, a credit card service, an electronic payment service, a commercial transaction (including online shopping) service, a travel business (including accommodation reservations or ticket sales) service, a communication service, a banking service, a securities trading service, and an insurance service. The point program service is a service that provides various added value, including monetary value, in units of points depending on the usage status of one or more services. For example, the behavioral indicators may be the presence or absence of monthly cross-use, the number of services used, the amount of points provided, the amount of points used, the amount of points expired, or the customer lifetime value (LTV) in a group of services to which a common user ID and point program service are applied.

[0043] Examples of behavioral indicators related to credit card services or electronic payment services include, but are not limited to, monthly usage amount or monthly usage frequency. Further examples of behavioral indicators for credit card services include monthly installment payment amount, revolving payment amount, repayment amount by other payment methods, or renewal rate of paid services. Further examples of behavioral indicators for electronic payment services include monthly payment amount, payment amount by prepaid payment means, number of coupon usages, or discount amount by coupon.

[0044] Examples of behavioral indicators related to retail commerce services include, but are not limited to, transaction amount, maximum amount of purchased items, number of purchased items, or number of listed items. Examples of behavioral indicators related to travel business services include, but are not limited to, monthly accommodation fee, distance traveled / traveled from home, domestic travel expenditure, or overseas travel expenditure. Examples of behavioral indicators related to communications (e.g., mobile communications or optical communications used on mobile phones, smartphones, or tablets) services include, but are not limited to, monthly data traffic, call time, payment amount, option fee, or subscription contract continuity rate.

[0045] Examples of behavioral indicators related to banking services include, but are not limited to, the monthly amount of remittance to other banks or accounts, the number of remittances, or the amount of withdrawals from a credit card. Examples of behavioral indicators related to securities trading services include, but are not limited to, the monthly purchase amount or savings amount of financial products (e.g., investment trusts, stocks, or bonds), or the amount of savings from prepaid payment instruments. Examples of behavioral indicators related to insurance services include, but are not limited to, the monthly withdrawal amount for insurance products under contract, or the amount of withdrawal by insurance type (e.g., non-life insurance or small amount short-term insurance).

[0046] The multiple segment indicators may include indicators related to multiple services that are different from each other. For example, the multiple segment indicators may include an indicator related to at least one of a credit card service, an electronic payment service, a commerce service, a travel business service, a communication service, a bank service, a securities exchange service, and an insurance service, and an indicator related to a point program service. Alternatively, the multiple segment indicators may include indicators related to two or more services of a credit card service, an electronic payment service, a commerce service, a travel business service, a communication service, a bank service, a securities exchange service, and an insurance service.

[0047] [Phase 1: User Classification] The analysis device 40 is configured to classify the multiple users 14 into multiple groups based on two or more segment indicators, and output the classification results. For example, the program code 45 of the analysis device 40 includes a classification code, which is configured to cause the processor 41 to read data related to the multiple users 14, classify the users 14 into multiple segments, and output the classification results. The users 14 belonging to each segment may be further classified into multiple groups. The analysis device 40 can gradually narrow down the characteristics of the users 14 belonging to the group by performing classification multiple times using different indicators.

[0048] The classification result may include, for example, at least one of a number and a character indicating the group to which the user belongs, or may include data in the form of a graph or table indicating the volume of the multiple groups. The format of the output may vary depending on the classification method. When the classification result includes data in the form of a graph, the program code 45 includes graph output code configured to cause the processor 41 to output one or more graphs indicating the distribution of the users 14 included in each of the multiple segments.

[0049] The program code 45 may include a storage code that causes the processor 41 to store the classification data indicating which of a plurality of segments each user 14 belongs to for each time period as an output result 46. The storage code may be further configured to cause the processor 41 to add the classification data to the analysis database 15 as part of the user data.

[0050] The classification code may be configured to cause the processor 41 to classify the multiple users 14 based on multiple segment indicators. The multiple segment indicators may include a frequency indicator related to the frequency of use of the service and a quantitative indicator related to the sales amount of the service. The frequency indicator is an example of a behavior indicator of the user 14. A boundary value may be set for each of the multiple segment indicators so that no gaps are created between adjacent segments. In other words, the boundary value may be set so that no data is missing or overlapping between segments.

[0051] For example, when classifying users 14 using two or three segment indicators, the classification results can be displayed using a graph with two axes (for example, a first axis being the horizontal axis and a second axis being the vertical axis), making it easy to visually understand. The most important segment indicator (KPI: Key Performance Indicator) may be set on the first axis. The third segment indicator may be visualized on a two-dimensional graph, for example, by showing the volume with the size of a circle in a bubble chart, or by changing the color or color intensity of a figure (for example, a circle) showing the volume. When the classification results are not visualized using a graph, users 14 may be classified using four or more segment indicators.

[0052] In the classification, it is preferable to set the classification conditions so that the difference or variance in the volumes between the segments is appropriate. For example, it is preferable to adjust the volume of each segment using an index that can be expressed by a numerical value for classification. The program code 45 may include a combination of indexes and a boundary value used for classification that are preset in order to perform appropriate classification. Furthermore, the program code 45 may include several patterns of combinations of indexes, and the patterns may be selectable.

[0053] FIG. 2 shows a graph of the classification results obtained by segmenting users 14 based on the amount of service usage and frequency of usage. Boundary values ​​are set for the data relating to the amount of service usage and frequency of usage so that no data is missing or overlapping between segments. Numerical data can be converted into input values ​​for data items to be classified by dividing the data into numerical values ​​or numerical ranges divided by the boundary values. The input value may be any alphanumeric character or symbol.

[0054] In the example of FIG. 2, the vertical axis indicates the amount of service usage, which is a quantitative indicator. The amount of usage is, for example, the amount of usage within a specified period (e.g., one month), and specific amounts (e.g., 0 yen and 10,000 yen) are set as boundary values. One of three pieces of data (e.g., 0, 1, 2) is entered into the classification data item related to the amount of usage. For example, "0" indicates no usage (0 yen), "1" indicates an amount of usage less than 10,000 yen, and "2" indicates an amount of usage 10,000 yen or more.

[0055] In the example of FIG. 2, the horizontal axis indicates the frequency of use of a service, which is a frequency index. The frequency of use is, for example, the number of times used within a specified period (e.g., one month), with boundary values ​​set at 1 time and 10 times. One of four data (e.g., X, S, M, L) is input into the classification data item related to the frequency of use. For example, X indicates less than 1 time of use (0 times), S indicates 1 time of use, M indicates 2 to 9 times of use (more than 1 time and less than 10 times), and L indicates 10 times or more.

[0056] In the example of FIG. 2, users 14 are classified into seven segments based on two segment indicators. In the graph of FIG. 2, the volume of each segment is visualized by illustrating circles whose size is proportional to the number of users. In FIG. 2, users 14 in the segment on the lower left (e.g., first segment X0) can be said to be light users of the service, while users 14 in the segment on the upper right (e.g., second segment L2) can be said to be heavy users of the service. The first segment X0 is a user group with no usage in the relevant month.

[0057] A user 14 who has not used the service for a certain period of time, such as the first segment X0, is also called a dormant user. In this way, the multiple segments may include a segment of a dormant user who has not used the service for a certain period of time. Alternatively or in addition, the multiple segments may include a segment of a new user who has no history of using the service.

[0058] The second segment S1 is a user group who used the service once in the relevant month for less than 10,000 yen. The third segment M1 is a user group who used the service 2 to 9 times in the relevant month for a total of less than 10,000 yen. The fourth segment L1 is a user group who used the service 10 times in the relevant month for a total of less than 10,000 yen. The fifth segment S2 is a user group who used the service once in the relevant month for more than 10,000 yen. The sixth segment M2 is a user group who used the service 2 to 9 times in the relevant month for a total of more than 10,000 yen. The seventh segment L2 is a user group who used the service 10 times in the relevant month for a total of more than 10,000 yen.

[0059] [Phase 2: Analysis of user attributes] The analysis device 40 may be configured to analyze attributes of users 14 belonging to a target segment and output an analysis result. In one example, the analysis device 40 may extract a group of users having common characteristics (attributes) as a cluster from among a plurality of users 14 included in the target segment. Clustering may be performed by a learning model employing an algorithm such as k-means.

[0060] In clustering, a number of attributes that may be common characteristics may be selected from among the many attributes possessed by each user 14. As an example, the users 14 in each segment may be classified into a number of clusters based on attributes that do not change or are unlikely to change, such as gender, age, place of residence, income, or family structure. Then, the occupancy rate of the number of users contained in each of the multiple clusters relative to the entire corresponding segment may be calculated. This makes it possible to know what attributes the users 14 have in each segment.

[0061] Furthermore, for example, the occupancy rate of a group with a high occupancy rate in a target segment may be compared with the occupancy rate of groups with the same characteristics in other segments. Alternatively, the occupancy rate of a specific group (for example, a group of "unmarried" "men") may be compared between the target segment and the entire user population. For example, when considering marketing measures aimed at a specific user group, it is advisable to perform a more detailed analysis of the target user group.

[0062] The analysis device 40 may perform the same classification using the same index every certain period (for example, one month). This makes it possible to analyze what attributes users 14 have that have remained in one segment or have transitioned to another segment (become transition users). Such transition situations can also be analyzed for users 14 that belong to one or more clusters.

[0063] For example, in Fig. 2, for segments M1 and L1 enclosed by a two-dot chain line, an arrow pointing from segment M1 to segment L1 indicates the corresponding transition path. In this manner, one or more graphs output by the graph output code may include a display of the transition history of user 14. In Fig. 2, the arrows show examples of transitions between adjacent segments that proceed in a direction in which the frequency of use or amount of money increases. Fig. 2 further shows examples of the proportion of users 14 who made a transition from the previous month to the current month for some transition paths, shown as a percentage.

[0064] Conversely, transitions in the direction of decreasing usage frequency or amount can also be analyzed by showing them with arrows, etc. If such a downward transition trend is observed, the cause can be further analyzed and marketing measures can be developed to suppress the decline. In addition, transition paths connecting non-adjacent segments can also be analyzed in the same way.

[0065] Such an analysis may be performed on a specific number of users 14 sampled from each segment. Alternatively, for dormant users in the current month, the actual usage status throughout the year may be understood, or transition trends in each segment may be understood. In addition, transition analysis may be performed by dividing users into new users who have just started using the service and existing users who have been using the service for some time.

[0066] 2 can be used to analyze the volume of a plurality of segments and its change over time, but it is also possible to clarify the transition situation over a longer period of time by analyzing the transition tendency of each user 14 included in each segment. For example, the analysis device 40 may output data on the segment membership frequency by calculating the number of months during the year that each user 14 belonged to each of a plurality of segments.

[0067] Figure 3 is an example of a bar graph showing the number of months that a user 14 belonged to each of the segments X0, S1, S2, M1, M2, L1, and L2 in a certain year. By displaying the data in a graph like this, the analysis results can be visualized. From this bar graph, the transition trends of each user 14 throughout the year can be read.

[0068] The increase or decrease in the number of users or the user ratio for each segment may be analyzed for a longer-term trend using a bar graph or a line graph. In this case, the vertical axis may show the number of users or the user ratio for each of a plurality of segments, and the horizontal axis may show the time course (e.g., month by month). Such a graph makes it possible to visualize the transition trend over a longer period of time.

[0069] The multiple indices used for classification may include indices related to multiple different services. For example, when a provider of a certain service starts providing a new service, the provider will prepare measures to encourage existing users to use the new service. In such a case, the users of the existing service are the subjects of analysis, and at the beginning of the launch of the new service, non-users, i.e., potential users 14, will be the majority.

[0070] FIG. 4 is an example of a graph showing the volume of each of a plurality of segments, and is particularly an example of a graph that can be used to show the transition state when potential users are in the majority. The vertical axis of FIG. 4 is the number of times of use, and the horizontal axis is the amount of use. In FIG. 4, the left side of the vertical axis shows the segment of non-users, and the right side of the vertical axis shows the segment of users using the new service. In this graph, by separately displaying non-users, it is possible to clearly show the comparison with other users 14.

[0071] As shown in FIG. 5, segmentation may be performed using the transition state and the usage status of the service as indicators. FIG. 5 shows an analysis plan for a case where a service provider starts offering a subscription service as a new service. In FIG. 5, two values, existing users U1 who have used the new service and non-users U0, are set on the vertical axis (second axis). More specifically, the first axis (most important KPI) for existing users U1 is set to the duration of the service (T0: service stopped, T1: within 1 month, T2: more than 1 month, within 3 months, T3: more than 3 months). The first axis for non-user users U0 is set to the awareness status of the new service (K1: aware, or K0: not aware).

[0072] For clustering, one or more personas, which are fictitious user images, may be set. For example, three clusters C1, C2, and C3 may be set as different marketing targets. The profile of the persona set in cluster C1 is, for example, "in their 20s to 30s," "employed," "living in rented housing," "unmarried," "drinks alcohol at restaurants," and "often uses electronic payment at convenience stores." The profile of the persona set in cluster C2 is, for example, "in their 40s to 50s," "employed," "owns a home," "married with children," "drinks beer," and "owns a car." The profile of the persona set in cluster C3 is, for example, "female, aged 35 to 45," "owns a home," "married with children," "owns a car," and "purchases skin care products and cosmetics."

[0073] As an example of analysis, the analysis device 40 may be configured to estimate transition attributes. For example, the program code 45 of the analysis device 40 may include estimation code that may be configured to cause the processor 41 to estimate, from among a plurality of attributes of the user 14, one or more transition attributes characteristic of a transition user that transitions to another segment over time.

[0074] The transition attributes may be estimated for the entire target segment, or for one or more clusters in the target segment. When estimating transition attributes for a cluster, some of the user attributes are limited by clustering, making it easier to analyze the transition attributes, especially when there are a large number of user attributes. In addition, a transition attribute related to one transition path may be estimated from one segment or one cluster, or transition attributes related to some or all of the transition paths may be estimated.

[0075] The transition attribute may be estimated using a trained learning model. The learning model may be a learning model 37U generated by the information processing device 30. In this case, the estimation code may include an input code. The input code may be configured to cause the processor 41 to input user data including a plurality of attributes of each user 14 to the learning model 37U. When the user data is input, the learning model 37U may be configured to output, as a calculation result, a result of predicting whether or not the user 14 will become a transition user and the importance of each attribute in the prediction.

[0076] The learning model 37U may be a neural network model trained to predict whether the user 14 related to the input user data corresponds to a transition user. The importance of each attribute in the prediction may be the probability of corresponding to the transition user. As an algorithm for predicting whether the user 14 will be a transition user, for example, boosting such as xgboost may be used.

[0077] The method for generating the learning model 37U includes acquiring training data and training the learning model before training using the training data. The training data includes user data and a correct answer label indicating whether each of the multiple users 14 included in the user data corresponds to a transition user.

[0078] Xgboost is an example of an ensemble learning algorithm, and is a model that combines boosting and decision trees. When training this learning model 37U, historical data of user transitions belonging to a target segment is used as correct answer data. When the trained learning model 37U receives attribute data of a certain user 14, it outputs a prediction result of whether the user 14 will transition to another segment.

[0079] The importance of each attribute in the prediction may be calculated as the impact or importance of the feature, as a SHAP (SHapley Additive exPlanations) value that applies the Shapley Value to machine learning. The SHAP value quantitatively represents the contribution of each feature in the prediction. It can be said that outputting the quantitative contribution of such features is an additional function of the learning model 37U, and the estimation code is configured to estimate the transition attribute using this additional output result.

[0080] In this analysis, among the calculation results of the learning model 37U, the SHAP value, i.e., the importance of each attribute in the prediction, is required, rather than the prediction result of "whether or not to transition". Therefore, instead of the user data of the target segment for which future transitions are to be predicted, past data of users who have actually transitioned may be input to the learning model 37U. In this case, the user data including the transition status used as training data for the learning model 37U may be divided into verification data, test data, and prediction data prior to training the learning model 37. Then, after the generalization performance of the learning model 37U is confirmed using the verification data and test data, the prediction data may be input to the trained learning model 37U to obtain the calculation results.

[0081] FIG. 6 shows an example of a result of outputting the importance of a feature as a SHAP value. The greater the SHAP value is, the easier the transition is, and the smaller the SHAP value is, the harder the transition is. Thus, the larger the SHAP value is, the more important the feature is in the transition. In FIG. 6, the SHAP values ​​of a plurality of feature values ​​are shown in a graph with the SHAP value on the horizontal axis. The data label on the vertical axis (the character "attribute" + a four-digit number) shows a plurality of different attributes (attribute + attribute value) as feature values. In FIG. 6, generally, the feature values ​​with high importance are arranged at the top of the vertical axis, and the importance decreases toward the bottom. The program code 45 or the learning model 37U may be configured to output a graph such as that shown in FIG. 6 as a part of the calculation result.

[0082] The estimation code may be configured to output one or more (e.g., two to three) feature quantities having high SHAP values ​​from among the output feature quantities as estimation results of one or more transition attributes characteristic of a transition user. In this case, instead of simply outputting the transition attribute based on the SHAP value alone, for example, a target user of marketing or an attribute related to the target user (e.g., a usage history of a characteristic service) may be preferentially selected as the transition attribute. The estimation code may further be configured to extract multiple users having one or more transition attributes as transition users from a target segment (at least one transition source segment among multiple segments, or a cluster in the target segment).

[0083] [Phase 3: Expansion of target group] The analysis device 40 may execute a process for expanding the target group and output the process result. The analysis device 40 reads the program code 45 and executes a process for expanding the target group based on the contents of the program code 45. More specifically, the program code 45 may include an acquisition code and an extraction code. The acquisition code is configured to cause the processor 41 to acquire user data including data of a plurality of users 14. At this time, the user data may include attribute data indicating a plurality of attributes possessed by each user 14 and output results in the first and second phases. For example, the user data may include classification data indicating which of a plurality of segments each user 14 belongs to for each period, and data indicating whether each user 14 corresponds to a transition user or a transitional user.

[0084] The extraction code may be configured to cause the processor 41 to extract one or more similar users from one segment from which transition users have been extracted, among the multiple segments. In other words, the analysis device 40 extracts one or more similar users having a combination of attributes similar to one or more transition attributes from the target segment (or a cluster in the target segment). For example, in FIG. 2, when planning a marketing measure to transition user 14 of the third segment M1 to the fourth segment L1, the third segment M1 is the target segment.

[0085] A combination of one or more transition attributes and similar attributes is also called a plurality of similar attributes. When outputting the similarity of attributes between a transition user having a transition attribute and another user 14, the output method may be set on a rule basis. Alternatively, the similarity may be calculated using a trained learning model 37T.

[0086] The extraction code or learning model 37T may include a conversion code and a calculation code. The conversion code is configured to convert attributes of multiple transition users into a vector representation in a feature space, an embedding space, or a latent space. The calculation code is configured to calculate a similarity between the vectors (a distance between data in a vector space). The extraction code or learning model 37T may be configured to extract similar users based on the similarity. In this case, the similarity between the vectors can be calculated using, for example, Euclidean distance or cosine similarity.

[0087] The transform code is an encoder for obtaining a vector representation. As an example of such an encoder, a transformer encoder, which is an example of the learning model 37T, may be adopted, but is not limited to this. As another example of the encoder, a graph neural network (GNN) may be adopted. An example of a GNN is a knowledge graph or a GCNN (Graph Convolutional Neural Network).

[0088] In one example, multiple attributes of all users in the transition source segment may be converted into a vector representation. Then, the cosine similarity of other users 14 to the vector of the transition user (or transition user) is calculated. Then, a vector with a higher cosine similarity than a predefined threshold, or a predefined number of vectors in descending order of cosine similarity, are selected. In this case, the user 14 corresponding to the selected vector is extracted as a similar user having a combination of attributes similar to one or more transition attributes.

[0089] The extraction conditions, for example, the number of similar users and the threshold of similarity, can be optimized as necessary. If the number of similar users is limited to those with higher similarity, the number of similar users will be reduced, and if the number of similar users is increased, users 14 with relatively low similarity will also be extracted as similar users. For example, the extraction conditions can be changed depending on the volume of the target group, or the content or cost of the targeted marketing measures.

[0090] The extracted code or learning model 37T may include a target output code. The target output code is configured to cause the processor 41 to output a target group. In this case, the target group includes a plurality of transitional users and one or more similar users. As a result, a target group is output in which the transitional users extracted as having transitional attributes are supplemented with similar users having attributes similar to those of the transitional users.

[0091] [Fourth phase: Analysis of policy data] The analysis device 40 may use the policy data to execute a process for calculating the cost-effectiveness of the proposed policy, and output the process result. The analysis device 40 may also execute a process for selecting a transition path to be targeted by the policy, and output the process result. For example, a segmentation transition performance coefficient may be calculated for all transition patterns that may occur by the execution of the marketing policy. The segmentation transition performance coefficient is an example of an index value indicating the economic effect. Depending on the content of the analysis, the economic effect may be estimated using an index value other than the transition effect coefficient.

[0092] The program code 45 may include estimation code configured to cause the processor 41 to estimate an economic effect associated with a transition of the user 14. The estimation code may include calculating a transition effect coefficient, where the transition effect coefficient indicates an economic effect associated with a transition from one of the plurality of segments to another segment. The estimation code may further include outputting a transition path having a higher economic effect among the plurality of transition paths by comparing the transition effect coefficients for the plurality of different transition paths with each other.

[0093] The transition effect coefficient may be calculated on the assumption that all or some of the users 14 included in each segment transition to another segment. For example, when it is assumed that some of the users 14 included in each segment transition, a tentative transition probability (e.g., 5% of the total) may be set.

[0094] The transition probability of the user 14 may be estimated by a learning model. This learning model may be the first learning model 37E generated by the information processing device 30. In this case, the trial calculation code may include an input code. The input code may be configured to cause the processor 41 to input user data to the first learning model 37E. The user data includes information regarding a segment (segment i) that each of the multiple users belongs to in a certain month and a segment (segment j) that belongs to the next month.

[0095] The method of generating the first learning model 37E includes acquiring training data and training a pre-training learning model using the training data. The first learning model 37E is configured to convert data of a transition user who transitions from one segment to another segment among a plurality of segments into a probability distribution, for example, by normalization. The first learning model 37E is configured to output the probability that the user 14 transitions from one segment to another segment when the user data is input.

[0096] The transition effect coefficient may be calculated by a learning model. This learning model may be the second learning model 37E generated by the information processing device 30. In this case, the trial calculation code may be configured to cause the processor 41 to input the input data to the second learning model 37E.

[0097] The input data to the second learning model 37E may include first input data and second input data. The first input data may include, for example, the number of transition users through a certain transition path during a certain period. The second input data may include a value indicating the quality of transition users of a certain transition path. The second input data is, for example, an average usage cost of a certain segment. The second learning model 37E may be configured to output a transition effect coefficient of a certain transition path when the input data is input, as shown in Equation (4) described later.

[0098] The second learning model 37E may be configured to estimate the value that each of the multiple users 14 will bring in the next year, for example, the annual expenditure amount, as an example of an economic effect, based on the first input data and the second input data. The annual expenditure amount of the user may be an estimated value calculated by the formula (2) described later. This annual expenditure amount can be used to calculate the total distribution transaction amount, revenue, or return on investment.

[0099] In the following formulas (1) to (4), T is a segment transition matrix for calculating the transition probability of user 14, and tij denotes the segment transition matrix indicating the transition rate from segment i to segment j. s is the average monthly expenditure, and si denotes the average monthly expenditure of segment i. ni denotes the number of users of segment i. u denotes a user vector, and ui=1 if it is a user vector of target segment i, and ui=0 if not.

[0100] The expenditure amount of a certain user 14 belonging to segment i in a certain month (month “m”) is calculated by the following formula (1).

[0101]

number

[0102] The annual expenditure of a user 14 belonging to segment i is calculated using the following formula (2).

[0103]

number

[0104] The transition difficulty from segment i to segment j is calculated based on the first input data by the following formula (3).

[0105]

number

[0106] Then, the difference between the annual expenditure in segment i and the annual expenditure in segment j calculated by formula (2), the transition difficulty calculated by formula (3), and the scale of the segment transition are multiplied as shown in the following formula (4), thereby calculating the transition effect coefficient for the transition from segment i to segment j.

[0107]

number

[0108] In addition, as shown in Figure 7, the service usage fee (sales) for a specified period (e.g., one month) may be calculated for each of the seven segments X0, S1, M1, L1, S2, M2, and L2, and the increase in sales when a user 14 transitions from each segment to another segment may be calculated as the transition effect coefficient.

[0109] In the example of Fig. 7, the transition effect coefficient (229.3) when transitioning from the fourth segment L1 to the sixth segment M2 is higher than the transition effect coefficients of the other transition paths. Therefore, the analysis device 40 may output the transition path from the fourth segment L1 to the sixth segment M2 as a transition path with a higher economic effect. Analysis of all transition paths may be performed in the early stage of policy planning, and then more detailed analysis may be carried out in stages.

[0110] The analysis device 40 may be configured to evaluate how the transition effect coefficient changes depending on which transition path is targeted by the measure. A measure for a segment with a large volume is expected to increase sales, but the cost of the measure may also be high. For example, even if the increase in sales is the highest when transitioning from the fourth segment L1 to the sixth segment M2, if the measure to promote the transition is very expensive, the measure may not be cost-effective.

[0111] Therefore, for example, the relationship between the cost of measures that change according to the size of the transition source segment and the transition effect coefficient may be analyzed to identify the most cost-effective transition path. In this case, the analysis of the policy data may be advanced while changing the analysis conditions in other phases. For example, in the first phase, multiple patterns of segmentation with different segment sizes may be performed to evaluate the trade-off between the transition effect coefficient and the policy cost. Alternatively, in order to analyze the change in the transition effect coefficient depending on the size of the target group, in the third phase, multiple target groups with different numbers of users may be extracted from one target segment, and the transition effect coefficients calculated for each target group may be compared.

[0112] Here, an example is given of an analysis of data related to a measure for a target group set for the transition path from the third segment M1 to the fourth segment L1. In this example, the target group includes three user groups corresponding to the three clusters C1, C2, and C3, respectively.

[0113] Different marketing measures are prepared for each of the three clusters C1, C2, and C3. For example, measures are prepared for cluster C1 to encourage the use of electronic payments at specific restaurants, and measures are prepared for cluster C2 to increase the frequency of credit card use. Furthermore, measures are prepared for cluster C3 to encourage the use of electronic payments at specific family restaurants. For example, a measure could be the awarding of service points according to the amount spent on target services during the campaign period.

[0114] It is assumed that the three measures will cause 5% of the target group to become transition users and transition to the fourth segment L1. As a result, the number of transition users is calculated, and the total merchandise sales (GMS) or revenue that increases with the transition is calculated as the economic effect. Here, the return on investment (ROI) may be calculated as the economic effect, taking into account the costs related to the measures. The annual expenditure calculated by the above formula (2) can be used to calculate the total merchandise sales, revenue, or return on investment. The estimation code may be configured to compare the economic effects calculated for multiple measures with each other and output the more economical measure.

[0115] [Information processing method] FIG. 8 is a flow showing the procedure of the information processing method according to this embodiment. The analysis device 40 reads the program code 45, and executes the process flow shown in FIG. 8 based on the contents of the program code 45. As shown in FIG. 8, the processor 41 may continuously execute each process of segmentation (step S22), analysis of user attributes (second phase, step S25), expansion of the target group (third phase, step S28), and analysis of the policy data (fourth phase, step S32). Alternatively, the processor 41 may individually execute the processes corresponding to the first to fourth phases (steps S22, S25, S28, S32) without succession, or may execute only some of them continuously.

[0116] Segmentation (step S22) is a classification process corresponding to the above-mentioned first phase processing. Analysis of user attributes (step S25) corresponds to the above-mentioned second phase processing and may include an estimation process for estimating one or more transition attributes. Expansion of the target group (step S28) may include an output process (output process) of the target group in the above-mentioned third phase. Analysis of policy data (step S32) corresponds to the above-mentioned fourth phase processing and may include an estimation process for estimating the economic effect associated with the transition.

[0117] After executing one of the four processes, the classification process, the estimation process, the output process, and the trial calculation process, the processor 41 may feed back the result of the executed process to one of the remaining three of the four processes. In this case, the processor 41 may re-execute the process that received the feedback.

[0118] The processor 41 may perform a process (steps S23, S26, S29, S33) of determining whether or not the corresponding process has been completed after steps S22, S25, S28, and S32. The processor 41 may determine whether or not the corresponding process has been completed in steps S23, S26, S29, and S33 based on a condition for determining whether or not the output results of steps S22, S25, S28, and S32 are appropriate. If the processes are not completed in steps S23, S26, S29, and S33, the processor 41 may change the conditions for calculation in the following steps S24, S27, S30, and S34. In the change of conditions in steps S24, S27, S30, and S34, the processor 41 may predefine the conditions to be changed and automatically perform the condition change process, or may change the conditions based on an input from an operator.

[0119] Regarding the segmentation (step S22), for example, a segmentation threshold related to the difference or variance in volume between segments may be set in advance, and the judgment in step S23 may be made depending on whether the condition of the segmentation threshold is met. If the completion condition is met in step S23, the process proceeds to the next step S25, and if the completion condition is not met in step S23, the segmentation condition may be changed in step S24, and the segmentation process (step S22) may be performed again.

[0120] The analysis of user attributes (step S25) includes at least one of outputting the analysis results including charts and graphs and estimating transition attributes (estimation step). For example, in estimating transition attributes, threshold conditions related to the SHAP value, the number of combinations of transition attributes, or the number of transition users may be set. Then, the judgment in step S26 may be made depending on whether or not the condition is met.

[0121] If the completion condition is satisfied in step S26, the processor 41 may proceed to the next step S28, and if the completion condition is not satisfied in step S26, the processor 41 may change the conditions related to the estimation in step S27. Fig. 8 shows an example in which the processor 41 returns to the segmentation in step S22 after step S27. Alternatively, the processor 41 may change the analysis conditions in step S27 and perform the user attribute analysis process (step S25) again.

[0122] Regarding the expansion of the target group (step S28), for example, the ratio between the number of similar users and the transition users may be set in advance as an extraction condition, and the judgment in step S29 may be made depending on whether or not the extraction condition is met. Then, if the completion condition is met in step S29, the process proceeds to the next step S31, and if the completion condition is not met in step S29, the condition related to the extraction of similar users may be changed in step S30, and the process in step S28 (or steps S47 to S49) may be performed. FIG. 8 shows an example in which the process returns to step S28 after step S30. Alternatively, after a negative determination in step S29, the condition may be changed, and segmentation (step S22) or analysis of user attributes (step S25) may be performed again.

[0123] Regarding the analysis of the policy data (step S32), for example, a significant difference in the evaluation value of the policy may be set in advance as a threshold condition, and the judgment in step S33 may be made depending on whether the threshold condition is satisfied. If the completion condition is satisfied in step S33, the process may be terminated. On the other hand, if the completion condition is not satisfied in step S33, the condition may be changed in step S34, and feedback may be given to the previous process. FIG. 8 shows an example of returning to step S22 after step S34. Alternatively, the condition may be changed in step S34, and the analysis process of the user attributes (step S25), the expansion of the target group (step S28), or the analysis of the policy data (step S32) may be performed again.

[0124] Even if the completion condition is satisfied in step S33, processor 41 may end the process after outputting data for feedback to any one or more of steps S22, S25, S28, and S32. For example, after the positive determination in step S33, processor 41 may end the process after outputting data for changing one or more analysis conditions included in the first to fourth phases.

[0125] The output process of the target group includes estimating transition attributes, which is a part of the second phase, and expanding the target group in the third phase (step S28). Therefore, instead of consecutively performing the first to fourth phases in FIG. 8, the processor 41 may separately perform the output of the target group as shown in the flow in FIG. 9 below.

[0126] 9 is a flow chart showing the procedure of the target group output method according to the present embodiment. The analysis device 40 reads the program code 45, and executes the process flow shown in FIG.

[0127] First, in step S41, the processor 41 inputs user data to the learning model 37U. In the following step S42, the learning model 37U outputs a calculation result. The calculation result includes the importance of multiple attributes in predicting whether a certain user 14 will transition to another segment.

[0128] In step S43, the processor 41 estimates transition attributes based on the importance of the multiple attributes output for the user predicted to transition. More specifically, the processor 41 selects one or more multiple attributes with high importance as transition attributes from the multiple attributes (attributes and attribute values).

[0129] In the next step S44, the processor 41 extracts transitional users from the target segment. More specifically, the processor 41 extracts a plurality of users 14 having a combination of a plurality of transition attributes from among the users 14 belonging to the target segment. In step S44, the processor 41 may input user data of the target segment to the learning model 37U and extract users 14 predicted to be transition users as transitional users.

[0130] In step S45, the processor 41 determines whether or not it is necessary to expand the target group based on the number of transitional users extracted in step S44 and the number of set users who are the targets of the marketing measures. If the number of transitional users alone is not enough to meet the set number of users, the processor 41 determines that expansion is necessary and proceeds to step S46. On the other hand, if the number of transitional users alone is sufficient to meet the set number of users, the processor 41 proceeds to step S49. In this case, in step S49, the processor 41 outputs a list of transitional users as the target group. Note that, in step S45, if the number of transitional users exceeds the set number of users, the number of transitional users may be reduced by, for example, changing the conditions in step S43 or step S44.

[0131] In step S46, the processor 41 vectorizes a plurality of attributes of the user 14 belonging to the target segment. In the following step S47, the processor 41 calculates the similarity of the vector of the other users 14 to the vector of the transition user. Then, in step S48, the processor 41 extracts one or more similar users from among the other users 14 based on the similarity calculated in step S47. Thereafter, in step S49, the processor 41 outputs a list of users including the transition user and the similar user as a target group, and ends the process.

[0132] [Effects of the present disclosure] By classifying a plurality of users 14, it is possible to narrow down the users 14 to be targeted for marketing. For example, by segmenting based on a segment indicator indicating the usage status of the service, it is possible to clarify the distribution of users 14 based on the usage status of the service. Furthermore, by clarifying the characteristic user groups included in each segment and the number of users, it is possible to find the most influential user groups or the user groups that are most important in terms of business. Influential or important user groups can be promising candidates for targets of effective marketing activities.

[0133] For example, by estimating the transition attribute of a transition user for a transition path that transitions from one segment to a better segment among multiple segments, it becomes possible to extract transition users having the transition attribute as more promising potential users. In addition, by extracting similar users having characteristics similar to the transition user, it is possible to adjust the number of users in the target group while appropriately maintaining the characteristics of users 14 belonging to the target group.

[0134] When a transition user has multiple transition attributes, a user 14 having any of the multiple transition attributes can be selected as a similar user, or a user 14 having an attribute similar or close to any of the transition attributes can be selected as a similar user. However, since it is not clear what influence each of the multiple transition attributes may have, these selection methods are not persuasive. In this respect, in this example, a user 14 having a combination of one or more transition attributes and similar attributes is extracted as a similar user by using, for example, a vector representation. Therefore, it is more persuasive and appropriate than analyzing the individual attributes of the user 14 separately.

[0135] It is desirable for a service provider that a certain user 14 transitions from a segment with a low usage amount to a segment with a higher usage amount. In addition, cost-effectiveness is improved by conducting promotional activities targeting transitional users who are likely to transition to a more desirable segment and similar users. In particular, by outputting the analysis results of the campaign data in the fourth phase, the cost-effectiveness of multiple marketing campaigns can be compared and evaluated. Furthermore, optimization of campaigns can be achieved by repeating the analysis while mutually feeding back the output results of the first to fourth phases.

[0136] [Effects of this disclosure] According to the present disclosure, the following effects can be achieved. (1) The learning model 37U can estimate one or more transition attributes that are characteristic of a transition user.

[0137] (2) The prediction code can extract transition users who are likely to become prospective users from the target segment. (3) By segmenting users 14 based on their service usage history, transition trends can be appropriately analyzed.

[0138] (4) By using a combination of quantitative and frequency indices as multiple segment indices, it is possible to appropriately analyze changes in usage over time. (5) By using multiple segment indicators related to multiple services, it is possible to appropriately analyze the cross-use situation across multiple services.

[0139] (6) By adding segments of new users or inactive users, it is possible to consider marketing strategies to encourage them to start or resume using the service. (7) By using the SHAP value in prediction of transition attributes, the contribution of each attribute in the prediction can be quantified.

[0140] (8) The number of users included in the target group can be adjusted by expanding the target group by adding similar users. (9) By extracting transition users from the target segment, it is possible to include potential users who are likely to transition in the target group.

[0141] (10) Similar users have attributes similar to those of transitional users, and therefore can be potential users. By including transitional users and similar users in a target group, the target group can be set more broadly and appropriately.

[0142] (11) By using the similarity of vector representation to extract similar users, it is possible to quantitatively determine whether or not a wider variety of attributes are similar. In addition, since it is possible to prioritize combinations of multiple attributes using a single index called similarity, it is easy to adjust the volume of the target group.

[0143] (12) By estimating the economic effect associated with user transitions, the economic effects of multiple marketing measures can be compared with each other without implementing multiple marketing measures.

[0144] (13) The second learning model 37E can calculate a transition effect coefficient that takes into account the number of transition users and the quality of the transition users. (14) The distribution of users in each of a number of segments can be visualized by outputting one or more graphs.

[0145] (15) A graph including a display of the user's 14 transition history allows visualization of how transitions are occurring between multiple segments. (16) By repeating the analysis while mutually feeding back the output results of phases 1 to 4, marketing measures can be optimized.

[0146] [Example of change] This embodiment can be modified as follows: This embodiment and the following modifications can be combined with each other to the extent that there is no technical contradiction.

[0147] [Modification 1] In the second phase, in addition to classifying all users, a group of users having a specific attribute may be extracted and the extracted group of users may be classified. In this case, an attribute that does not change or is difficult to change, such as place of residence or gender, may be selected. For example, one or more attributes may be selected from a plurality of attributes such as age, gender, place of residence, income, and family structure, and users 14 having the selected attributes in common may be extracted. For example, a group of users extracted under the condition of being "unmarried" and "male," or a group of users extracted under the condition of being "home-owning," "in their 40s or 50s," and "having an annual income within a certain range" may be segmented.

[0148] [Modification 2] The analysis performed by the analysis device 40 may not include any one or more of the first to fourth phases, or may be divided into phases different from those in the above embodiment. For example, the second phase may be divided into an analysis of data including graph output and an estimation of transition attributes. Alternatively, the output of the target group shown in FIG. 9 may be one phase.

[0149] [Modification 3] Analysis device 40 may perform another analysis in addition to all or part of phases 1 to 4. For example, a phase for evaluating the results of at least one of phases 1 to 4, or a phase for feeding back the results to the other phases, may be added.

[0150] [Modification 4] Each of the first to fourth phases, or a computer program relating to some or all of the first to fourth phases, may be provided online or offline as an application for data analysis. Also, some of the first to fourth phases may be provided by the application, and the other parts may be executed by the analysis device 40. For example, the calculation process for analysis may be executed by the analysis device 40, and the output results including graphs may be provided by download.

[0151] [Modification 5] For example, in the case where the service server 20 handles products or services of multiple businesses, such as an online commerce site, the whole or part of the series of analyses including segmentation may be performed by focusing on products or services provided by a specific business. The results of such analyses may be provided to the specific business, or, as in modification 4, the analysis function itself may be provided to the specific business as an application.

[0152] [Modification 6] The format and type of the output graph, the analysis method, and the algorithm of the learning model can be changed as desired. [Modification 7] In the embodiment, the target group is set mainly on transition users who are likely to transition, but the target characteristic is not limited to transition possibility. For example, depending on the desired purpose such as promoting the use of a specific service, promoting use in a specific area, expanding a trade area, uncovering latent needs, activating dormant users, acquiring new users, and promoting cross-use, attributes highly correlated with the target characteristic can be estimated and used for analysis. In this case, it is preferable to extract users 14 having attributes similar to the attributes highly correlated with the target characteristic as similar users. In addition, in the fourth phase, it is preferable to use an effect coefficient according to the target characteristic instead of the transition effect coefficient.

[0153] [Example 8 of Modification] In expanding the target group, instead of or in addition to extracting similar users based on the vectorization and similarity of attributes, users 14 having highly important transition attributes may be extracted as similar users based on the importance of multiple attributes output by the learning model 37U.

[0154] Aspects grasped based on the above embodiment and modified examples are listed below. [1] at least one memory configured to store computer program code; at least one processor configured to perform operations based on the content of the computer program code, The processor, A classification process for classifying a plurality of users into a plurality of segments; An estimation process of estimating one or more transition attributes characteristic of a transition user who transitions to another segment over time from among a plurality of attributes of the user; configured to run The estimation process includes an input process of inputting user data including the plurality of attributes of each of the users into a learning model; The learning model is configured to input the user data and output a calculation result; The calculation result includes a prediction result of whether the user will become the transition user, and an importance of each of the attributes in the prediction. Information processing system.

[0155] [2] The estimation process further includes extracting a transition user having the one or more transition attributes from at least one of the plurality of segments. The information processing system according to [1] above.

[0156] [3] The user is a subscriber of one or more services provided by the business operator; The one or more services include one or more of a points program service, a credit card service, an electronic payment service, a commerce service, a travel business service, a communication service, a banking service, a securities exchange service, and an insurance service; The user data includes a usage history of the one or more services. The information processing system according to the above [1] or [2].

[0157] [4] The classification process includes classifying the plurality of users based on a plurality of segment indicators; the plurality of segment indicators include a frequency indicator related to a frequency of use of the service and a quantitative indicator related to sales of the service; A boundary value is set for each of the plurality of segment indicators so that no gap occurs between adjacent segments. An information processing system according to any one of the above [1] to [3].

[0158] [5] The plurality of segment indicators include indicators relating to a plurality of services that are different from each other, The information processing system according to [4] above. [6] The plurality of segments includes at least one of a segment of new users who have no history of using the service or a segment of dormant users who have not used the service for a certain period of time. An information processing system according to any one of [1] to [5] above.

[0159] [7] The importance of each attribute in the prediction is calculated as a SHAP (SHapley Additive exPlanations) value. An information processing system according to any one of the above items [1] to [6].

[0160] [8] the processor: An acquisition process for acquiring user data including data of a plurality of users, the user data including classification data indicating to which of a plurality of segments each of the users belongs in each of a plurality of time periods, and attribute data indicating a plurality of attributes possessed by each of the users; an estimation process for estimating one or more transition attributes from the plurality of attributes, the transition attributes being attributes characteristic of a transition user who transitions to another segment over time; an extraction process configured to extract one or more similar users from at least one of the plurality of segments, the similar users being users having a combination of attributes similar to the one or more transition attributes; and further configured to perform An information processing system according to any one of [1] to [7] above.

[0161] [9] The extraction process further includes extracting one or more transition users from at least one of the plurality of segments; The one or more similar users are extracted from a plurality of remaining users of the segment from which the one or more transitional users are extracted. The information processing system according to [8] above.

[0162]

[10] The processor is further configured to perform a target output process to output a target group; The target group includes the plurality of transitional users and the one or more similar users. The information processing system according to [9] above.

[0163]

[11] The extraction process includes a conversion process of converting the attributes of the multiple transition users into a vector representation; The extraction process includes extracting the one or more similar users based on a similarity of the vector representation. The information processing system according to

[10] above.

[0164]

[12] The processor is further configured to perform an estimation process for estimating an economic effect associated with a user transition; The trial calculation process is calculating a transition effect coefficient, the transition effect coefficient indicating an economic effect associated with a transition from one of the plurality of segments to another of the plurality of segments; outputting a transition path having a higher economic effect from among the plurality of transition paths by comparing the transition effect coefficients with each other for the plurality of transition paths different from each other; Including, An information processing system according to any one of [1] to

[11] above.

[0165]

[13] The estimation process includes inputting input data into a learning model; the input data includes first input data and second input data; The first input data includes a number of transition users through a transition path during a period of time; the second input data includes a value indicating a quality of the transition user of the certain transition path; The learning model is configured to output the transition effect coefficient of the certain transition path when the input data is input. The information processing system according to

[12] above.

[0166]

[14] The processor is configured to execute a graph output process to output one or more graphs showing a distribution of users included in each of the plurality of segments. An information processing system according to any one of the above [1] to

[13] .

[0167]

[15] the one or more graphs include a representation of the user's transition history; the transition history includes a transition path indicating a transition from one segment to another segment among the plurality of segments; 14. An information processing system according to claim 13.

[0168]

[16] One or more computers, Classifying a plurality of users into a plurality of segments; Estimating one or more transition attributes characteristic of a transition user who transitions to another segment over time from among a plurality of attributes of the user;

[0033] The said estimation is predicting whether the user will become the transition user; Calculating the importance of each of the attributes in the prediction. Information processing methods.

[0169]

[17] to the one or more computers, outputting a target group from at least one of the plurality of segments, the target group including one or more transitional users having the one or more transition attributes and one or more similar users having a combination of attributes similar to the one or more transition attributes; calculating a transition effect coefficient indicating an economic effect associated with a transition from one of the plurality of segments to another segment; and further comprising: The method further includes, after causing the one or more computers to execute one of the four steps of the classification, the estimation, the output, and the estimation, feeding back the execution result to the remaining one or more of the four steps. The information processing method according to

[16] above.

[0170]

[18] A method for generating a learning model, comprising: When multiple users are classified into multiple segments, a user who transitions to another segment over time is a transition user. The user data includes a plurality of corresponding attribute data for each of the plurality of users; the learning model is configured to, when the user data is input, output a prediction result of whether the corresponding user will become the transition user and an importance of each attribute in the prediction; The method for producing the same comprises: Obtaining training data; and training a pre-trained learning model using the training data; The training data includes the user data and a correct answer label indicating whether each of the plurality of users included in the user data corresponds to the transition user. How to generate a learning model.

[0171]

[19] One or more computers, Classifying a plurality of users into a plurality of segments; Estimating one or more transition attributes characteristic of a transition user who transitions to another segment over time from among a plurality of attributes of the user; A computer program comprising: The said estimation is predicting whether the user will become the transition user; Calculating the importance of each of the attributes in the prediction. Computer program.

[0172]

[20] at least one memory configured to store computer program code; at least one processor configured to perform operations based on the content of the computer program code, The processor, An acquisition process for acquiring user data including data of a plurality of users, the user data including classification data indicating to which of a plurality of segments each of the users belongs in each of a plurality of time periods, and attribute data indicating a plurality of attributes possessed by each of the users; an estimation process for estimating one or more transition attributes from the plurality of attributes, the transition attributes being attributes characteristic of a transition user who transitions to another segment over time; an extraction process configured to extract one or more similar users from at least one of the plurality of segments, the similar users being users having a combination of attributes similar to the one or more transition attributes; An information processing system configured to execute the

[0173]

[21] One or more computers, Acquiring user data including data of a plurality of users, the user data including classification data indicating to which of a plurality of segments each of the users belongs in each of a plurality of time periods, and attribute data indicating a plurality of attributes possessed by each of the users; estimating one or more transition attributes from the plurality of attributes, the transition attributes being attributes characteristic of a transition user who transitions to another segment over time; and Extracting one or more similar users from at least one of the plurality of segments, the similar users being users having a combination of attributes similar to the one or more transition attributes; and An information processing method, comprising:

[0174]

[22] A method for generating a learning model, comprising: The learning model is configured to output a transition effect coefficient of a transition path when input data is input; The method for producing the same comprises: Obtaining training data; and training a pre-trained learning model using the training data; the training data and the input data include first input data and second input data; The first input data includes a number of transition users through a transition path during a period of time; The second input data includes a value indicating a quality of the transition user of the certain transition path. How to generate a learning model.

[0175]

[23] One or more computers, Acquiring user data including data of a plurality of users, the user data including classification data indicating to which of a plurality of segments each of the users belongs in each of a plurality of time periods, and attribute data indicating a plurality of attributes possessed by each of the users; estimating one or more transition attributes from the plurality of attributes, the transition attributes being attributes characteristic of a transition user who transitions to another segment over time; and Extracting one or more similar users from at least one of the plurality of segments, the similar users being users having a combination of attributes similar to the one or more transition attributes; and A computer program comprising:

[0176]

[24] at least one memory configured to store computer program code; at least one processor configured to perform operations based on the content of the computer program code, The processor, an acquisition process for acquiring campaign data related to one or more marketing campaigns, the campaign data including prediction data regarding to which segment a user belonging to one of a plurality of segments will transition after the one or more marketing campaigns are implemented; A calculation process for calculating an economic effect associated with a user transition based on the policy data; configured to run The trial calculation process is calculating a transition effect coefficient, the transition effect coefficient indicating an economic effect associated with a transition from one of the plurality of segments to another of the plurality of segments; outputting a transition path having a higher economic effect from among the plurality of transition paths by comparing the transition effect coefficients with each other for the plurality of transition paths different from each other; An information processing system comprising:

[0177]

[25] One or more computers, Acquiring campaign data related to one or more marketing campaigns, the campaign data including prediction data regarding to which segment a user belonging to one of a plurality of segments will transition after the one or more marketing campaigns are implemented; Calculating an economic effect associated with a user transition based on the policy data; and Including, The above-mentioned calculation is calculating a transition effect coefficient, the transition effect coefficient indicating an economic effect associated with a transition from one of the plurality of segments to another of the plurality of segments; outputting a transition path having a higher economic effect from among the plurality of transition paths by comparing the transition effect coefficients with each other for the plurality of transition paths different from each other; An information processing method comprising:

[0178]

[26] One or more computers, Acquiring campaign data related to one or more marketing campaigns, the campaign data including prediction data regarding to which segment a user belonging to one of a plurality of segments will transition after the one or more marketing campaigns are implemented; Calculating an economic effect associated with a user transition based on the policy data; and

[0033] The above-mentioned calculation is calculating a transition effect coefficient, the transition effect coefficient indicating an economic effect associated with a transition from one of the plurality of segments to another of the plurality of segments; outputting a transition path having a higher economic effect from among the plurality of transition paths by comparing the transition effect coefficients with each other for the plurality of transition paths different from each other; A computer program comprising: [Explanation of symbols]

[0179] 11...information processing system, 12...network, 13...user terminal, 14...user, 15...analytical database, 20...service server, 21...processor, 22...memory, 23...communication device, 24...communication bus, 25...application, 26...database, 30...information processing device, 31...processor, 32...memory, 33...communication device, 34...communication bus, 35...learning program, 36...training data, 37, 37U, 37T, 37E...machine learning model, 40...analysis device, 41...processor, 42...memory, 43...communication device, 44...communication bus, 45...computer program code, 46...output result.

Claims

1. at least one memory configured to store computer program code; at least one processor configured to perform operations based on the content of the computer program code, The processor, A classification process for classifying a plurality of users into a plurality of segments; an estimation process for estimating one or more transition attributes characteristic of a transition user who transitions to another segment over time from among a plurality of attributes of the user; configured to run The estimation process includes an input process of inputting user data including the plurality of attributes of each of the users into a learning model; The learning model is configured to input the user data and output a calculation result; The calculation result includes a prediction result of whether the user will become the transition user, and an importance of each of the attributes in the prediction. Information processing system.

2. The estimation process further includes extracting a transitional user having the one or more transition attributes from at least one of the plurality of segments. The information processing system according to claim 1 .

3. The user is a subscriber of one or more services provided by the operator; The one or more services include one or more of a points program service, a credit card service, an electronic payment service, a commerce service, a travel business service, a communication service, a banking service, a securities exchange service, and an insurance service; The user data includes a usage history of the one or more services. The information processing system according to claim 1 .

4. The classification process includes classifying the users based on a plurality of segment indicators; the plurality of segment indicators include a frequency indicator related to a frequency of use of the service and a quantitative indicator related to sales of the service; A boundary value is set for each of the plurality of segment indicators so that no gap occurs between adjacent segments. The information processing system according to claim 3 .

5. The plurality of segment indicators include indicators relating to a plurality of services different from each other.

5. The information processing system according to claim 4.

6. The plurality of segments includes at least one of a segment of new users who have no history of using the service and a segment of dormant users who have not used the service for a certain period of time. The information processing system according to claim 1 .

7. The importance of each attribute in the prediction is calculated as a SHAP (SHapley Additive exPlanations) value. The information processing system according to claim 1 .

8. The processor, An acquisition process for acquiring user data including data of a plurality of users, the user data including classification data indicating to which of a plurality of segments each of the users belongs in each of a plurality of time periods, and attribute data indicating a plurality of attributes possessed by each of the users; an estimation process for estimating one or more transition attributes from the plurality of attributes, the transition attributes being attributes characteristic of a transition user who transitions to another segment over time; an extraction process for extracting one or more similar users from at least one of the plurality of segments, the similar users being users having a combination of attributes similar to the one or more transition attributes; and further configured to perform The information processing system according to claim 1 .

9. The extraction process further includes extracting one or more transitional users from at least one of the plurality of segments; The one or more similar users are extracted from a remaining plurality of users of the segment from which the one or more transitional users are extracted. The information processing system according to claim 8.

10. The processor is further configured to perform a target output process to output a target group; The target group includes the plurality of transitional users and the one or more similar users. The information processing system according to claim 9.

11. The extraction process includes a conversion process of converting the attributes of the multiple transition users into a vector representation; The extraction process includes extracting the one or more similar users based on a similarity of the vector representation. The information processing system according to claim 10.

12. The processor is further configured to execute an estimation process for estimating an economic effect associated with a user transition; The trial calculation process is calculating a transition effect coefficient, the transition effect coefficient indicating an economic effect associated with a transition from one of the plurality of segments to another of the plurality of segments; outputting a transition path having a higher economic effect from among the plurality of transition paths by comparing the transition effect coefficients with each other for the plurality of transition paths different from each other; Including, The information processing system according to claim 1 .

13. The trial calculation process includes inputting input data into a learning model; the input data includes first input data and second input data; The first input data includes a number of transition users through a transition path during a period of time; the second input data includes a value indicating a quality of the transition user of the certain transition path; The learning model is configured to output the transition effect coefficient of the certain transition path when the input data is input. The information processing system according to claim 12.

14. The processor is configured to execute a graph output process to output one or more graphs showing a distribution of users included in each of the plurality of segments.

14. The information processing system according to claim 1.

15. the one or more graphs include a representation of the user's transition history; the transition history includes a transition path indicating a transition from one segment to another segment among the plurality of segments; 15. The information processing system according to claim 14.

16. On one or more computers, Classifying a plurality of users into a plurality of segments; Estimating one or more transition attributes characteristic of a transition user who transitions to another segment over time from among a plurality of attributes of the user; [0033] The said estimation is predicting whether the user will become the transition user; Calculating the importance of each of the attributes in the prediction. Information processing methods.

17. The one or more computers, outputting a target group from at least one of the plurality of segments, the target group including one or more transitional users having the one or more transition attributes and one or more similar users having a combination of attributes similar to the one or more transition attributes; calculating a transition effect coefficient indicating an economic effect associated with a transition from one of the plurality of segments to another segment; and further comprising: The method further includes, after causing the one or more computers to execute one of the four of the classification, the estimation, the output, and the trial calculation, feeding back the execution result to the remaining one or more of the four. The information processing method according to claim 16.

18. A method for generating a learning model, comprising the steps of: When multiple users are classified into multiple segments, a user who transitions to another segment over time is a transition user. The user data includes a plurality of corresponding attribute data for each of the plurality of users; the learning model is configured to, when the user data is input, output a prediction result of whether the corresponding user will become the transition user and an importance of each attribute in the prediction; The method for producing the same comprises: Obtaining training data; and training a pre-trained learning model using the training data; The training data includes the user data and a correct answer label indicating whether each of the plurality of users included in the user data corresponds to the transition user. How to generate a learning model.

19. On one or more computers, Classifying a plurality of users into a plurality of segments; Estimating one or more transition attributes characteristic of a transition user who transitions to another segment over time from among a plurality of attributes of the user; A computer program comprising: The said estimation is predicting whether the user will become the transition user; Calculating the importance of each of the attributes in the prediction. Computer program.

Images