Information processing device, method and program

Abstract

To create an optimized schedule for each for a plurality of time slots included in a unit time while considering other time slots.SOLUTION: An information processing apparatus according to the present invention for determining a user as a target of an operation has a target time slot determining unit 21 for determining a target time slot among a plurality of time slots included in a unit time, a candidate user list acquiring unit 22 for acquiring a candidate user list excluding a user who satisfies a condition in a time slot preceding the target time slot, and an operation score outputting unit 23 for, by inputting a user identifier, user attribute data, an action result amount, and a target time slot identifier of each of the plurality of users included in the candidate user list, outputting an operation score in the target time slot with respect to each of the users.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present disclosure relates to a technique for supporting operations for a user. [Background technology]

[0002] Conventionally, an information processing device has been proposed that has a calculation unit that calculates a score, which is the probability that a subject will answer a phone call, using a prediction formula learned by a learning engine based on response information, which is information indicating past phone call response status, and a coefficient corresponding to the subject's attribute information, and an output unit that outputs the score calculated by the calculation unit (see Patent Document 1). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2020-021151 A Summary of the Invention [Problem to be solved by the invention]

[0004] Conventionally, in order to have a user perform a predetermined action (e.g., paying an unpaid amount), an operation (e.g., calling a customer, etc.) is performed to prompt the user to perform the predetermined action, and in order to efficiently utilize resources for the operation, it has been proposed to use machine learning to determine an operation schedule for a predetermined unit of time (e.g., one day). However, simply using machine learning to determine the operation schedule may result in a response in which the operation schedule for the unit of time is determined collectively, and there is room for improvement in terms of optimizing the operation schedule within limited resources.

[0005] In view of the above-mentioned problems, an object of the present disclosure is to create an optimized schedule for each of a plurality of time slots included in a unit time while taking into account the other time slots. [Means for solving the problem]

[0006] One example of the present disclosure is an information processing device for determining a user who is to be the target of an operation to prompt a user to perform a specified action, the information processing device including: a target timeslot determination means for determining a target timeslot from among a plurality of timeslots included in a specified unit time; a candidate user list acquisition means for acquiring a candidate user list from which users who satisfy a specified condition in the preceding timeslot of the target timeslot from among the plurality of timeslots included in the unit time are excluded; and an operation score output means for outputting an operation score indicating the effect of the operation in the target timeslot for each user included in the candidate user list by inputting a user identifier, user attribute data, the amount of action result when the user performs the action, and an identifier of the target timeslot for each of a plurality of users included in the candidate user list.

[0007] The present disclosure can be understood as an information processing device, a system, a method executed by a computer, or a program executed by a computer. The present disclosure can also be understood as such a program recorded on a recording medium readable by a computer or other device, machine, etc. Here, a recording medium readable by a computer, etc. refers to a recording medium that stores information such as data and programs by electrical, magnetic, optical, mechanical, or chemical action and can be read by a computer, etc. Effect of the Invention

[0008] According to the present disclosure, it is possible to create an optimized schedule for each of a plurality of time slots included in a unit time, taking into account the other time slots. [Brief description of the drawings]

[0009] [Figure 1] 1 is a schematic diagram showing a configuration of an information processing system according to an embodiment. [Diagram 2] 1 is a diagram illustrating an outline of a functional configuration of an information processing device according to an embodiment. [Diagram 3] FIG. 2 is a conceptual diagram showing the flow of a process for creating a call list for multiple time slots on a certain day in an embodiment. [Figure 4] FIG. 2 is a schematic diagram showing a flow of a call list creation process according to the embodiment. [Diagram 5] 11 is a flowchart showing a flow of a call list creation process according to the embodiment. [Figure 6] 1 is a diagram illustrating a flow of a call list creation process in an embodiment, using a specific example. FIG. [Figure 7] 1 is a flowchart illustrating a flow of a reinforcement learning process according to an embodiment. [Figure 8] FIG. 11 is a diagram illustrating an example of a process of collecting a log used in the reinforcement learning process according to the embodiment. [Figure 9] FIG. 1 is a diagram showing an overview of a process for creating training data used in a reinforcement learning process according to an embodiment. [Figure 10] FIG. 1 is a diagram illustrating an overview of a reinforcement learning process according to an embodiment. [Figure 11] 1 is a flowchart showing the flow of a machine learning process for a feasibility estimation model according to an embodiment. [Figure 12] FIG. 13 is a diagram showing an outline of the functional configuration of an information processing device according to a variation. [Figure 13] 13 is a flowchart showing the flow of a machine learning process for a variation-related causal score estimation model. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0010] Hereinafter, an embodiment of an information processing device, a method, and a program according to the present disclosure will be described with reference to the drawings. However, the embodiment described below is merely an example of an embodiment, and the information processing device, a method, and a program according to the present disclosure are not limited to the specific configuration described below. In carrying out the present disclosure, a specific configuration according to the embodiment may be appropriately adopted, and various improvements and modifications may be made.

[0011] In this embodiment, an embodiment will be described in which the technology according to the present disclosure is implemented for an operation center management system for collecting receivables by urging payment of overdue credit card usage amounts. However, systems to which the technology according to the present disclosure can be applied are not limited to operation center management systems for urging payment of credit card usage amounts. The technology according to the present disclosure can be widely used as a technology for supporting various operations for users, and the application of the present disclosure is not limited to the examples shown in the embodiment.

[0012] Normally, credit card usage charges are paid by debiting the user's account on the monthly withdrawal date, or by the user depositing the amount by a specified date, but there are cases where the payment of the credit card usage charge is not completed by the specified date due to reasons such as insufficient funds in the user's account, the user not depositing the amount by the specified date, etc. For this reason, conventionally, operations such as calling users (customers) or sending messages from an operation center (call center) are performed to urge users to pay their credit card usage charges and collect receivables.

[0013] Generally, in order to avoid default (non-performance of debt), operations on users are effective, and the more the amount of operations is increased, the higher the amount of debt collection. However, on the other hand, the more the amount of operations is increased, the higher the costs of labor costs for operations, system usage fees, system maintenance costs, etc. are. In view of the above-mentioned problems, the system according to the present disclosure employs a technology for increasing the amount of debt collection while suppressing the costs related to operations, in other words, a technology for improving (preferably maximizing) the effect of debt collection using limited resources as much as possible. Note that in this embodiment, an example will be described in which the predetermined operation is mainly a call to the user, but the content of the predetermined operation is not limited, and may be various operations for encouraging the user to take a predetermined action. For example, message transmission may be adopted as a type of operation on the user. Here, the means for sending the message is not limited, and an e-mail system, a short message service (SMS), or a message sending and receiving service of a social networking service (SNS) may be used.

[0014] Here, in order to efficiently utilize resources for operations (e.g., the operation costs of a call center), it is conceivable to use machine learning to determine an operation schedule for a predetermined unit time (e.g., one day) and select users to be the targets of the operation. Note that, although the present embodiment describes a case in which the unit time is one day, the unit time is not limited to the example given in the present embodiment, and may be other time periods such as one week or one month.

[0015] However, simply determining the operation schedule using machine learning may result in a response in which the operation schedule for the unit time is determined collectively. For this reason, in this embodiment, instead of determining the operation schedule for the unit time collectively, an optimized schedule is created for each of a plurality of time slots included in the unit time while taking into account the other time slots.

[0016] <System configuration> 1 is a schematic diagram showing the configuration of an information processing system according to this embodiment. In the system according to this embodiment, an information processing device 1, an operation center management system 3, and a credit card management system 5 are connected to each other so that they can communicate with each other. An operation terminal (not shown) is installed in the operation center to perform operations according to instructions from the operation center management system 3, and an operator operates the operation terminal to perform operations for a user. The user is a credit card user who pays the credit card usage amount via a financial institution or the like, and the payment history data of the credit card usage amount is notified to the operation center management system 3 via the credit card management system 5.

[0017] The information processing device 1 is an information processing device for outputting data for controlling operations by the operation center management system 3. The information processing device 1 is a computer including a central processing unit (CPU) 11, a read only memory (ROM) 12, a random access memory (RAM) 13, a storage device 14 such as an electrically erasable and programmable read only memory (EEPROM) or a hard disk drive (HDD), a communication unit 15 such as a network interface card (NIC), and the like. However, the specific hardware configuration of the information processing device 1 can be omitted, replaced, or added as appropriate depending on the embodiment. In addition, the information processing device 1 is not limited to a device consisting of a single housing. The information processing device 1 may be realized by a plurality of devices using so-called cloud or distributed computing technology, etc.

[0018] The operation center management system 3, the credit card management system 5, and the operation terminal are all computers equipped with a CPU, a ROM, a RAM, a storage device, a communication unit, an input device, an output device, etc. (not shown). Furthermore, these systems and terminals are not limited to devices consisting of a single housing. These systems and terminals may be realized by multiple devices using so-called cloud or distributed computing technology, etc.

[0019] 2 is a diagram showing an outline of the functional configuration of the information processing device 1 according to this embodiment. The information processing device 1 functions as an information processing device including a target time slot determination unit 21, a candidate user list acquisition unit 22, an operation score output unit 23, a reinforcement learning unit 24, a target user determination unit 25, a performance index calculation unit 26, a feasibility estimation unit 27, and a machine learning unit 28, by reading a program recorded in the storage device 14 into the RAM 13 and executing it by the CPU 11, and controlling each hardware included in the information processing device 1. Note that in this embodiment and other embodiments described later, each function included in the information processing device 1 is executed by the CPU 11, which is a general-purpose processor, but some or all of these functions may be executed by one or more dedicated processors.

[0020] The information processing device 1 determines users to be targets of an operation for urging the user to perform a predetermined action, and outputs a call list including the determined users to the operation center management system 3. In this embodiment, the operation is to contact the user (for example, to make a call) to urge the user to pay the unpaid amount for which payment is delayed. The call to the user may be an automatic call using a recorded or mechanical voice, or a call in which an operator (human) talks to the user. The predetermined action is to pay the overdue credit card amount. The specific payment method is not limited, and may be a transfer to a designated account, a payment at a designated counter, or the like. When making a reminder call to a user who has unpaid credit card payments, the information processing device 1 determines and optimizes a call destination user for each time slot (time period) using a reinforcement learning algorithm.

[0021] The target time slot determination unit 21 determines a target time slot from among multiple time slots included in a unit time (e.g., a day, a week, a month, etc.) and obtains an identifier (time slot ID, e.g., time slot 1, time slot 2, etc.) of the target time slot.

[0022] The candidate user list acquisition unit 22 selects from the list of unpaid users: (1) A user who has received an operation a predetermined number of times (e.g., three times; however, the predetermined number of times may be set to one time) or more in a preceding time slot preceding the target time slot among a plurality of time slots included in a unit time; (2) A user who has received an operation in the previous time slot and performed an action (in this embodiment, a payment); and (3) Users who have been determined to be the targets of the operation in the preceding time slot (in other words, users who are listed on the call list for the preceding time slot); A list of candidate users from which at least one of the above conditions is removed (hereinafter, "candidate user list") is obtained. At this time, which of the above conditions is used to remove users from the candidate user list may be appropriately determined depending on the embodiment. In addition, users who are unlikely to repay by operations such as telephone contact may be removed in advance from the candidate user list.

[0023] The operation score output unit 23 inputs data related to the candidate user, the target time slot, and the preceding time slot into the reinforcement learning model, and outputs an operation score indicating the effect of the operation in the target time slot for each candidate user. Specifically, the operation score output unit 23 inputs the following data as data related to the candidate user, the target time slot, and the preceding time slot. (i) a user identifier (user ID) for each of a plurality of users included in the candidate user list; (ii) User attribute data (e.g., age, gender, occupation, etc. Details of the data used as user attributes will be described later.); (iii) The probability that the user who received the operation will actually execute the action (action execution probability) (iv) the amount of the action outcome when the user performs the action (in this embodiment, the amount due by the user); (v) Operation result in the preceding time slot (result of the user's response to the operation, including a flag indicating whether the user's action was performed or not (success or failure of recovery) and data indicating the amount of the result (amount of recovery), and (vi) Identifier of the target timeslot (timeslot ID).

[0024] In this embodiment, the reinforcement learning model used by the operation score output unit 23 has a weight generation model 231 (first model) and a score calculation model 232 (second model), and outputs an operation score by linking these models. The weight generation model 231 is a neural network that outputs a weight θ to be applied to user data including the action execution possibility for each user and the action result amount (unpaid amount) for each user. Here, the action execution possibility for each user indicates the possibility that a target user will execute an action when an operation is received in a target time slot (user-based debt collection possibility). In addition, the score calculation model 232 is a linear model that calculates an operation score for each user based on the weight θ, the action execution possibility for each user, and the action result amount output from the weight generation model 231.

[0025] The reinforcement learning unit 24 generates and / or updates a reinforcement learning model used by the operation score output unit 23 using training data created for each of the multiple time slots. Here, the training data includes a state, an action, and a reward related to the target time slot. The state related to the target time slot includes a user ID for each of the multiple users included in the candidate user list, user attribute data, an action outcome amount when the user executes an action, an operation result in the preceding time slot, and a time slot ID of the target time slot. The action related to the target time slot includes a user ID (call list) of the target user determined for the target time slot.

[0026] The reward for the target timeslot includes a performance indicator for the target timeslot. Note that the reward for the target timeslot may be a performance indicator based only on the operation result for the target timeslot, but in this embodiment, the reward for the target timeslot includes the sum of one or more action performance indicators obtained in the target timeslot and the subsequent timeslot of the target timeslot among the multiple timeslots included in the unit time.

[0027] Here, the training data may be created based on log data collected in actual operation, or may be created based on data generated by simulation. Specifically, the probability that each user who receives an operation in each time slot will execute an action (action execution possibility) is calculated in advance by simulating actual operation, and if the action execution possibility exceeds a threshold, the user is considered to have executed the action ("action execution"), and if it is equal to or less than the threshold, the user is considered to have not executed the action ("action non-execution"), thereby generating a pseudo operation result, and an expected value based on this operation result and the amount of action outcome (unpaid amount) is calculated, and this can be used as a reward in the training data (a specific method of calculating the reward will be described later). By employing such a simulation, when there is no operational track record, it is also possible to train an initial model using training data created by simulation, and update the model using training data created based on log data collected in actual operation after the start of operation. A feasibility estimation model described later may be used in the simulation.

[0028] The target user determination unit 25 extracts users from the users included in the candidate user list based on the operation score, with the upper limit set for the target time slot, determines the extracted users as target users for the operation in the target time slot, and outputs a call list in which the determined target users are listed. At this time, the upper limit is determined based on the resources that can be secured for the operation in the target time slot. In this embodiment, the upper limit is determined based on the number of times that calls can be made in the target time slot.

[0029] The performance index calculation unit 26 calculates the performance index (time slot recovery rate) of the operation in the time slot based on the amount of action performance of the user who performed the action by receiving the operation in the time slot. A specific method of calculating the performance index will be described later.

[0030] The feasibility estimation unit 27 estimates the action execution possibility of the user who received the operation using a machine learning model (hereinafter referred to as the "feasibility estimation model") that outputs the action execution possibility for the user in response to the input of one or more user attribute data related to the user.

[0031] The machine learning unit 28 generates and / or updates a feasibility estimation model used by the feasibility estimation unit 27. The feasibility estimation model is a machine learning model that outputs an action feasibility indicating the possibility that the user will execute an action when one or more user attribute data related to a target user are inputted, if the user receives an operation in a target time slot. The user attributes inputted to the machine learning model may include, for example, demographic attributes, behavioral attributes, or psychographic attributes. Here, the demographic attributes are, for example, the user's sex (gender), family structure, age, etc., the behavioral attributes are, for example, the presence or absence of cash advance use, the presence or absence of revolving payment use, the deposit and withdrawal history related to a specified account, the commercial transaction history related to some product including gambling or lottery (which may include online transaction history in an online marketplace, etc.), etc., and the psychographic attributes are, for example, preferences related to gambling or lottery. However, the available user attributes are not limited to the examples in this embodiment. For example, "time required for operation (call, etc.)" and "credit card usage amount" may also be used as attributes.

[0032] In generating and / or updating the feasibility estimation model, the machine learning unit 28 creates the feasibility estimation model based on teacher data (machine learning data) in which an index based on a statistical amount related to the execution rate of an action by a user who has received an operation among a plurality of users having a predetermined attribute is defined as the action execution possibility of a user having the attribute. In this embodiment, for example, when the content of the operation is a call to a user, "the possibility of collecting a debt when a user is called" is calculated as the action execution possibility based on the statistical amount, and the calculated action execution possibility is combined with the attribute data of the corresponding user and input to the machine learning unit 28 as the teacher data. For example, when the average value of the debt collection possibility from a group of users having a common attribute who have been called is 80%, the action execution possibility of the group of users is "0.8". Note that in this embodiment, an example in which an average value is used as the statistical amount will be described. However, as the statistical amount, a statistical index such as a mode or a median may be used. Here, the statistical amount related to the action execution possibility may be based on the past debt collection possibility of each user within a predetermined period (for example, a predetermined month).

[0033] A machine learning model generation framework that can be adopted when implementing the technology according to the present disclosure is based on an ensemble learning algorithm, for example. For example, a machine learning framework (for example, LightGBM) based on a gradient boosting decision tree (GBDT) may be adopted as the framework. In other words, the framework may be a machine learning framework based on a decision tree model that transfers the error between the correct answer and the predicted value between previous and next weak learners (weak classifiers). The predicted value here refers to the predicted value of the action execution possibility, for example. In addition to LightGBM, the framework may adopt a boosting method such as XGBoost or CatBoost. According to a framework using a decision tree, a machine learning model having a relatively high performance can be generated with less effort in parameter adjustment compared to a framework using a neural network. However, the machine learning model generation framework that can be adopted when implementing the technology according to the present disclosure is not limited to the example in this embodiment. For example, instead of a gradient boosting decision tree, another learning device such as a random forest may be adopted as a learning device, or a learning device other than a so-called weak learning device such as a neural network may be adopted. In particular, when a learning device other than a so-called weak learning device such as a neural network is adopted, ensemble learning does not need to be adopted.

[0034] When adopting a gradient boosting machine learning framework based on a decision tree algorithm, the branching conditions of each node of the decision tree are optimized. Specifically, in the gradient boosting machine learning framework based on a decision tree algorithm, the action execution possibility is calculated for each user group having attributes indicated by each of two child nodes branched from one parent node, and the branching condition of the parent node is optimized so that the difference in the action execution possibility becomes large (for example, so that the difference becomes maximum or becomes equal to or greater than a predetermined threshold), that is, so that the two child nodes branch neatly. For example, when the attribute indicated as the branching condition of the node is age, the age set as the branching threshold may be changed, or the branching condition may be changed to an attribute other than age. In this way, the branching conditions of all nodes of the decision tree are recursively optimized, thereby improving the estimation accuracy of the action execution possibility.

[0035] <Processing flow> Next, a flow of processing executed by the information processing system according to the present embodiment will be described. Note that the specific contents and processing order of the processing described below are an example for implementing the present disclosure. The specific contents and processing order may be appropriately selected according to the embodiment of the present disclosure.

[0036] FIG. 3 is a conceptual diagram showing the flow of processing for creating a call list for multiple time slots on a certain day in this embodiment (however, the processing up to time slot 2 is shown in the figure, and processing from time slot 3 onwards is omitted). The processing shown in this figure is executed at a preset timing each month. More specifically, the execution timing of the processing is set to be after the specified date for payment of the credit card usage amount and before the scheduled date for execution of an operation for an unpaid user. Note that it is assumed here that each reinforcement learning model and machine learning model have already been trained.

[0037] In this embodiment, for each time slot, a candidate user list, which is a list of candidate users to be called in the time slot, attribute data of users in the candidate user list, the time slot ID of the time slot for which the call list is to be created, and the operation result in the preceding (e.g., the previous) time slot are input to a reinforcement learning model to generate a call list for the target time slot. Here, the "candidate users" are users who may be called in the time slot, excluding users who have already made a call or have received a predetermined reaction before the time slot, in other words, "remaining users". Note that in this embodiment, the call list may be generated by further inputting the amount of action results (each user's unpaid amount) related to each user in the candidate user list into the reinforcement learning model.

[0038] In this embodiment, by adopting the operation results in the preceding time slot as input to the reinforcement learning model, a call list reflecting the operation results in the preceding time slot is created, enabling real-time optimization of call destination users. In this embodiment, an example in which the actual operation results in the preceding time slot are input to the reinforcement learning model will be described, but the operation results in the preceding time slot input to the reinforcement learning model may be substituted with those obtained by simulation. Also, a reinforcement learning model that does not use the operation results in the preceding time slot may be adopted. When adopting the operation results obtained by simulation or when not using the operation results in the preceding time slot, a call list for one unit time (one day in this embodiment) may be created in one go.

[0039] Fig. 4 is a schematic diagram showing the flow of the call list creation process according to this embodiment. The process shown in this figure shows the process by the reinforcement learning model in more detail among the process flows of creating call lists for multiple time slots shown in Fig. 3, and is executed after the end of the previous time slot and before the target time slot when the actual operation result in the previous time slot is used as an input, and is executed at any timing before the target time slot when the actual operation result in the previous time slot is not used as an input.

[0040] As described above, in this embodiment, the process by the reinforcement learning model includes two models, the weight generation model 231 (first model) and the score calculation model 232 (second model). The weight generation model 231 is a neural network that receives the candidate user list, the user attribute data, the time slot ID, and the operation result from the previous time slot, and outputs the weight θ used in the score calculation model 232. The score calculation model 232 is a linear model that receives the user data for each user (action execution possibility, action outcome amount (unpaid amount), etc.) and the weight θ set by the weight generation model 231, and outputs the call priority score of each user. The user data used here may be acquired as a part of the user attribute data.

[0041] 5 is a flowchart showing the flow of the call list creation process according to the present embodiment. The process shown in this flowchart is executed for each target time slot that is the target of the operation in the process of creating a call list for multiple time slots of a certain day shown in FIG.

[0042] In steps S101 to S103, data to be input to the reinforcement learning model is acquired. The target time slot determination unit 21 acquires the time slot ID of the time slot to be the target for creating the call list (step S101). In this embodiment, there are five time slots: 8:00-10:00, 10:00-12:00, 12:00-14:00, 14:00-16:00, and 16:00-18:00, and each time slot is assigned a time slot ID (time slot 1 to time slot 5).

[0043] The information processing device 1 also acquires operation result data indicating the result of the call in the preceding time slot (step S102). The operation result data includes at least the presence or absence of a reaction obtained from the user to whom the call was made as a result of the call, and may further include the content of the reaction (for example, a response to the call, a promise to pay, or execution of payment, etc.). In this embodiment, an embodiment in which a call list for the next time slot is created at a timing between a certain time slot and the next time slot is described, and an aspect in which operation result data is acquired to create a call list reflecting the operation result of the preceding time slot is described, but the acquisition of operation result data and input to the reinforcement learning model may be omitted. In this way, it is also possible to create call lists for multiple time slots at once.

[0044] The candidate user list acquisition unit 22 also acquires a candidate user list to be the target of the current call list creation process (step S103). The candidate user list acquired here includes users who may be included in the current call list and excludes users who should not be included in the current call list, and at least users who have obtained a predetermined reaction (e.g., a response to a call, a promise to pay, or the execution of a payment, etc.) in the previous time slot are excluded from the candidate user list. Furthermore, at least any of users who have made a call a predetermined number of times or more (e.g., three or more times) in the previous time slot and users listed in the call list for the previous time slot may be excluded from the candidate user list. In this embodiment, an example is described in which a process of excluding users listed in the call list for the previous time slot from the candidate user list is adopted. When data to be input to the reinforcement learning model is acquired, the process proceeds to step S104.

[0045] In steps S104 and S105, an operation score for each user in the candidate user list is acquired based on the acquired data. The operation score output unit 23 inputs the data acquired in steps S101 to S103 to the weight generation model 231, and obtains, as an output, a weight θ for calculating an operation score suitable for the target candidate user list (step S104). In this embodiment, the weight θ is output in the form of a matrix corresponding to each parameter (in this embodiment, the possibility of performing an action, the amount of outstanding payment, etc.) included in the user data to be considered by the linear model. In this embodiment, the weight generation model 231 is a model generated and updated by reinforcement learning, but the type of model used for weight generation is not limited to the example in this embodiment.

[0046] When the weight θ is obtained, the operation score output unit 23 calculates the operation score of each user included in the candidate user list obtained in step S103 based on each parameter (possibility of performing an action, unpaid amount, etc.) included in the user data and the weight θ obtained in step S104 (step S105). After that, the process proceeds to step S106.

[0047] Fig. 6 is a diagram illustrating the flow of the call list creation process in this embodiment with a specific example. According to the example shown in this figure, the score calculation model 232 performs weighting calculated for each parameter (here, the action execution possibility and the unpaid amount) included in the user data, and calculates the operation score for each user by summing up the weighted parameters. That is, the score calculation model 232 in this embodiment can be expressed by the following formula. Operation score = (weight for parameter 1 * parameter 1) + (weight for parameter 2 * parameter 2) + (weight for parameter n * parameter n)

[0048] However, the specific weighting method is not limited to the example in the present disclosure. For example, in the present embodiment, an example in which a linear model is used as the score calculation model 232 has been described, but the model used for operation score calculation is not limited to the example in the present embodiment. In addition, the example of parameters and weight θ shown in FIG. 6 is simplified to explain the flow of the process according to the present embodiment, and the type and number of parameters used can be changed as appropriate, and the size of the weight matrix can vary depending on the type and number of parameters. When implementing the technology according to the present disclosure, a larger weight matrix and more types of parameters may be used.

[0049] In step S106, a call list is created. The target user determination unit 25 extracts users included in the candidate user list obtained in step S103 in order of effectiveness of calling up to the upper limit n of the number of users who can be called in the target time slot according to the operation score for each user obtained in step S105, and generates a call list by adding the extracted users to the call list. Then, the target user determination unit 25 outputs the generated call list as a list of targets to be called in the target time slot. According to the example shown in FIG. 6, it can be seen that the top four users are extracted in order of the calculated operation score, and a call list including the extracted users F, G, H, and A is generated. Here, the number n of users that can be included in the call list may be set according to the processing capacity of the operation center. Note that in this embodiment, an example is described in which an operation score for each user is calculated by a combination of the weight generation model 231 and the score calculation model 232, and users are ranked, but the specific method of ranking users is not limited to the example in this embodiment. Thereafter, the process shown in this flowchart ends.

[0050] When the call list is output, the operation center management system 3 manages operations for the target users according to the call list, and the operation terminal executes the operations according to the instructions output by the operation center management system 3. The execution results of the operations are acquired by the information processing device 1 as operation results.

[0051] 7 is a flowchart showing the flow of the reinforcement learning process according to this embodiment. The process shown in this flowchart is executed at a timing designated by an administrator of the information processing device 1. Note that in this embodiment, the reinforcement learning model is not updated between time slots belonging to the same day, but the reinforcement learning model may be updated between time slots.

[0052] Before the process shown in this flowchart is executed, the necessary logs are collected and stored in the log database. The information processing device 1 acquires a log including a list of users who are candidates for calling (candidate user list), user attribute data, the outstanding amount of each user, a time slot ID, a call list created for the target time slot, and the success or failure of collection by calling and the amount collected (operation result) for each user included in the call list (see FIG. 3). The acquired log is used to create the state S, action A, and reward R included in the training data for reinforcement learning. As described above, the operation results for creating the training data may be created by simulation instead of log data obtained in actual operation.

[0053] In steps S201 and S202, a target time slot is determined and a log is acquired. The reinforcement learning unit 24 determines the target time slot in the current reinforcement learning process automatically or based on the administrator's designation (step S201), and acquires the log related to the target time slot from the log database (step S202). Hereinafter, the training data T, state S, action A, and reward R corresponding to the target time slot t are referred to as the training data T. t , state S t , action A t and reward R t When the log is acquired, the process proceeds to step S203.

[0054] In step S203, state S t , action A t and reward R t The reinforcement learning unit 24 uses the log acquired in step S201 to generate a state S for the target time slot t. t , action A t and reward R tAmong the logs acquired in step S201, the data including the candidate user list, the user attribute data, the unpaid amount of each user, and the time slot ID are logs for creating the state S. The reinforcement learning unit 24 creates the state S by converting the time slot ID of the target time slot t and the candidate user table corresponding to the time slot t. t Here, the candidate user table is a table in which the user ID of the candidate user, the attribute data of the candidate user, and the unpaid amount of the candidate user are associated with each other.

[0055] In addition, the reinforcement learning unit 24 performs the following operations: t In addition, among the logs acquired in step S201, the success or failure of collection by calls and the collected amount for each user included in the call list are logs for creating the reward R, and the performance indicator calculation unit 26 calculates the ratio of the amount collected by calls in the target time slot t to the total amount of unpaid amounts of users included in the candidate user list for the target time slot t (hereinafter referred to as the "time slot collection rate"), and calculates this as the reward R. t Let us assume that.

[0056] FIG. 8 is a diagram showing an example of a process of collecting logs used in the reinforcement learning process according to this embodiment. According to the example shown in this figure, it can be seen that the state S2 for time slot 2 is a candidate user table (a table in which attribute data of candidate users and the unpaid amounts of the candidate users are associated) corresponding to time slot ID: 2 and time slot 2, the action A2 is a call list generated for time slot 2, and the reward R2 is a time slot recovery rate obtained as a result of a call in time slot 2. Here, the ratio (0.288 in the example shown in FIG. 8) of the amount recovered by the call in time slot 2 (the total of the recovered amounts of users D, A, and C who successfully recovered, 650) to the total amount of unpaid amounts of users included in the candidate user list for time slot 2 (the total of the unpaid amounts of users A to G, 2260 in the example shown in FIG. 8) is calculated as the time slot recovery rate (performance index). However, the specific calculation method of the performance index is not limited to the example in this embodiment. State S t , action A t and reward R t Once created, the process proceeds to step S204.

[0057] In step S204, the reward G t In the case where there are n time slots in one day, from time slot 1 to time slot n, the reinforcement learning unit 24 calculates the reward G t The reward R for time slots t to n is t From R n The sum of the calculated reward G t is the training data T for the reinforcement learning process according to this embodiment. t That is, in this embodiment, as a reward for training data related to a certain time slot t, a reward R t In addition, the reward G is the sum of the rewards for time slots t to n after the time slot t among the time slots 1 to n on the same day. t is used.

[0058] FIG. 9 is a diagram showing an overview of the process of creating training data used in the reinforcement learning process according to this embodiment. According to the example shown in this figure, if there are five time slots in a day, from time slot 1 to time slot 5 in order from the earliest time, the reward G1 for time slot 1 is the sum of rewards R1 to R5, the reward G2 for time slot 2 is the sum of rewards R2 to R5, and the reward G5 for time slot 5 is the same as reward R5. By using such a reward G, it is possible to obtain a reinforcement learning model that can create a call list in which users are assigned to each time slot so as to maximize the effect of calls throughout the day, rather than simply maximizing the effect of calls in the target time slot. Reward G t Once calculated, the process proceeds to step S205.

[0059] In step S205, a reinforcement learning model is generated or updated. t and Action A t and the reward G obtained in step S204. t training data T including t The reinforcement learning model is generated or updated by updating the parameters of the reinforcement learning model (neural network) according to the gradient calculated based on the above.

[0060] FIG. 10 is a diagram showing an outline of the reinforcement learning process according to this embodiment. In this embodiment, an example is given in which the policy gradient method is used for the reinforcement learning to maximize the reward for a policy function πθ with θ as a parameter. However, the specific method of the reinforcement learning is not limited to the example in this embodiment. After that, the process shown in this flowchart ends.

[0061] 11 is a flowchart showing the flow of machine learning processing for a feasibility estimation model according to this embodiment. The processing shown in this flowchart is executed at a timing designated by the administrator of the operation center management system 3.

[0062] The machine learning unit 28 calculates the action execution possibility for each of a plurality of user attributes based on the user attribute data, operation history data, and credit card usage payment history data previously accumulated in the operation center management system 3 or the credit card management system 5, and creates teacher data including a combination of the user attributes and the action execution possibility (step S301). Here, the operation history data includes data that can be used to grasp whether an operation has been performed for each user, and the payment history data includes data that can be used to grasp whether the user has paid the credit card usage amount (whether or not there is a default) for each user. Then, the machine learning unit 28 inputs the created teacher data into a machine learning model, and generates or updates a machine learning model (feasibility estimation model) used for feasibility estimation by the feasibility estimation unit 27 (step S302). After that, the process shown in this flowchart ends.

[0063] <Effects> According to this embodiment, rather than simply determining a schedule of operations for a unit time (one day in this embodiment) all at once, weights are calculated taking into consideration two factors for each user: which time slot is most effective for calling and how much payment is required. By using the operation scores calculated based on these weights to extract users to be called, it is possible to create a schedule that improves results over the entire unit time while taking into account other time slots and optimizes the distribution of resources.

[0064] Also, according to this embodiment, it is possible to generate a dynamic call list by considering the operation results of the preceding time slot. Furthermore, according to this embodiment, the reinforcement learning model is updated using the actual operation results during system operation, and the parameters are adjusted according to the user's attribute data, so that it is possible to easily respond to changes in user behavior (however, if the results obtained are better than expected, the model in operation may be fixed). In addition, according to this embodiment, the policy gradient method is introduced to create a call list for each time slot, so that it is possible to create a schedule of operations for a unit time with less computational resources (computational time or memory usage).

[0065] <Variations> In the above-described embodiment, an example was described in which an operation score was calculated based on the action execution possibility for each user, and a call list was created, but instead of the action execution possibility for each user, an operation score may be calculated based on a causality score indicating the effect of an operation on a user on the action execution possibility for that user, and a call list may be created. Hereinafter, regarding the information processing device 1b according to the variation, a description of the parts common to the above-described embodiment will be omitted, and differences from the above-described embodiment will be described.

[0066] FIG. 12 is a diagram showing an outline of the functional configuration of the information processing device 1b according to this variation. The hardware configuration of the information processing device 1b is the same as that of the information processing device 1 described with reference to FIG. 1, and therefore will not be described again. The information processing device 1b functions as an information processing device including a target time slot determination unit 21, a candidate user list acquisition unit 22, an operation score output unit 23b, a reinforcement learning unit 24b, a target user determination unit 25, a performance index calculation unit 26, a feasibility estimation unit 27, a machine learning unit 28b, and an effect estimation unit 29, by a program recorded in the storage device 14 being read into the RAM 13 and executed by the CPU 11, and each hardware included in the information processing device 1 is controlled. Note that in this embodiment and other embodiments described later, each function included in the information processing device 1 is executed by the CPU 11, which is a general-purpose processor, but some or all of these functions may be executed by one or more dedicated processors.

[0067] That is, the information processing device 1b according to this variation further includes an effect estimation unit 29. Also, the operation score output unit 23b and the machine learning unit 28b have partially different processing contents from the operation score output unit 23 and the machine learning unit 28 according to the above embodiment.

[0068] Like the operation score output unit 23 according to the above embodiment, the operation score output unit 23b inputs data related to the candidate user, the target time slot, and the preceding time slot into a reinforcement learning model, and outputs an operation score indicating the effect of the operation in the target time slot for each candidate user. However, instead of the reinforcement learning model of the operation score output unit 23 having the weight generation model 231 (first model) and the score calculation model 232 (second model), the operation score output unit 23b uses a reinforcement learning model having a weight generation model 233 (third model) and a score calculation model 234 (fourth model).

[0069] The weight generation model 233 is a neural network that outputs a weight θ to be applied to user data including a causal score for each user and an action outcome amount (unpaid amount) for each user. The score calculation model 234 is a linear model that calculates an operation score for each user based on the weight θ, the causal score for each user, and the action outcome amount output from the weight generation model 231. Here, the causal score is a score that indicates the effect of an operation on a user in a target time slot t on the possibility of the user performing an action.

[0070] The effect estimation unit 29 estimates the effect of a predetermined operation for a user in a target time slot t for encouraging the user to perform a predetermined action on whether or not the user performs the action. In this embodiment, the effect estimation unit 29 estimates the effect of the operation using a machine learning model (hereinafter referred to as an "effect estimation model") that outputs a causality score indicating the effect of the operation on the user in response to input of one or more user attributes related to the target user.

[0071] The machine learning unit 28b generates and / or updates an effect estimation model used for effect estimation by the effect estimation unit 29. The effect estimation model is a machine learning model that outputs a causal score indicating the effectiveness of an operation on a target user when one or more user attribute data related to the target user are input. The user attributes input to the machine learning model are roughly the same as in the above embodiment, and therefore a description thereof will be omitted. In addition, the framework for generating the machine learning model that can be adopted for implementation is roughly the same as in the above embodiment, and therefore a description thereof will be omitted.

[0072] In generating and / or updating the effect estimation model, the machine learning unit 28b creates the effect estimation model based on teacher data (machine learning data) in which a score based on a statistic related to the possibility of action execution (possibility of debt collection) by a user who has received an operation among a plurality of users having a predetermined attribute and a statistic related to the possibility of action execution by a user who has not received an operation among a plurality of users is defined as a causal score indicating the effect of an operation on a user having the attribute. In this embodiment, for example, when the content of the operation is a call to a user, a causal score based on the difference between each statistic is calculated by the formula "(possibility of debt collection when the user makes a call)-(possibility of debt collection when the user does not make a call)", and the calculated causal score is combined with the attribute data of the corresponding user and input to the machine learning unit 28b as teacher data. Note that in this embodiment, an example in which an average value is used as the statistic will be described. However, as the statistic, a statistical index such as a mode or a median may be used. Here, the statistic related to the possibility of action execution may be based on the past debt collection possibility of each user within a predetermined period (for example, a predetermined month). In addition, in this embodiment, a causal score is calculated based on the difference between the statistics, but if the difference does not have a statistically significant difference, the causal score may be set to zero or approximately zero. Here, an existing statistical method may be used to determine whether or not there is a significant difference. For example, the standard error and confidence interval for a set of average collectability of users for each month can be considered, and the statistical significance of the change in collectability due to the presence or absence of a call can be considered. In this way, it is possible to calculate a causal score taking into account the variation in average collectability of users in the same group.

[0073] In creating the teacher data, since a user who has been called once cannot become a user who has not been called thereafter, for a user, only one of the possibility of recovery when the user is called and the possibility of recovery when the user is not called can be obtained. For this reason, teacher data on the effect of a call is created for each user group having a common attribute. That is, the causal score indicating the effect of a call on a user group consisting of users having a certain common attribute is obtained by, for example, dividing a plurality of users having the common attribute into a first sub-user group to be called and a second sub-user group to not be called, calculating the average value of the collectability of the claim from the first sub-user group to be called and the average value of the collectability of the claim from the second sub-user group to not be called, respectively, and calculating the difference between these average collectability values ​​based on the above formula. For example, if the average value of the collectability of the claim from the first sub-user group to be called is 80% and the average value of the collectability of the claim from the second sub-user group to not be called is 70%, the causal score indicating the effect of a call on the user group is "10".

[0074] 13 is a flowchart showing the flow of machine learning processing for the causal score estimation model according to this variation. The processing shown in this flowchart is executed at a timing designated by the administrator of the operation center management system 3.

[0075] The machine learning unit 28b calculates a causal score for each of a plurality of user attributes based on the user attribute data, operation history data, and credit card usage payment history data that have been accumulated in the operation center management system 3 or the credit card management system 5 in the past, and creates teacher data including a combination of the user attributes and the causal score (step S401). Here, the operation history data includes data that can be used to grasp whether an operation has been performed for each user, and the payment history data includes data that can be used to grasp whether the user has paid the credit card usage amount (whether or not there is a default) for each user. Then, the machine learning unit 28b inputs the created teacher data into a machine learning model, and generates or updates a machine learning model (causal score estimation model) used for effect estimation by the effect estimation unit 29 (step S402). After that, the process shown in this flowchart ends.

[0076] The flowcharts of other processes in this variation are generally similar except that, as described above, causal scores are used instead of action executable possibilities in some processes, and therefore descriptions thereof will be omitted.

[0077] 1. Information processing device

Claims

1. An information processing device for determining a user who is a target of an operation for prompting a user to perform a predetermined action, a target time slot determining means for determining a target time slot from among a plurality of time slots included in a predetermined unit time; a candidate user list acquisition means for acquiring a candidate user list from which users who satisfy a predetermined condition in a time slot preceding the target time slot among the plurality of time slots included in the unit time have been removed; an operation score output means having a model for outputting weights to be applied to an action execution possibility for each user, which indicates the possibility that the user will execute the action if the user receives the operation in the target time slot, and an action result amount for each user, which indicates the result amount when the user executes the action, in response to input of at least a user identifier for each user, user attribute data, and an identifier for the target time slot; and a model for calculating an operation score indicating the effect of the operation in the target time slot for each user, based on the weights output from the model for outputting the weights, the action execution possibility for each user, and the action result amount for each user, the operation score output means inputting the user identifiers, user attribute data, and identifier for the target time slot for each of a plurality of users included in the candidate user list into the model for outputting the weights to obtain the weights, and inputting the obtained weights, the action execution possibility for each user included in the candidate user list, and the action result amount for each user into the model for calculating the operation score, to output the operation score for each user included in the candidate user list; a target user determination means for determining a target user of the operation in the target time slot from among users included in the candidate user list based on the operation score; An information processing device comprising:

2. The apparatus further includes an execution possibility estimation means for estimating the execution possibility of the action of the user who has received the operation. The information processing device according to claim 1 .

3. the execution possibility estimation means estimates the action execution possibility of the user by using an execution possibility estimation model that outputs the action execution possibility for the user in response to input of one or more attributes related to the user; The information processing device according to claim 2 .

4. the execution possibility estimation model is created based on teacher data in which an index based on a statistical amount relating to an execution rate of the action by a user who has received the operation among a plurality of users having a predetermined attribute is defined as the execution possibility of the action of a user having the attribute; The information processing device according to claim 3 .

5. An information processing device for determining a user who is a target of an operation for prompting a user to perform a predetermined action, a target time slot determining means for determining a target time slot from among a plurality of time slots included in a predetermined unit time; a candidate user list acquisition means for acquiring a candidate user list from which users who satisfy a predetermined condition in a time slot preceding the target time slot among the plurality of time slots included in the unit time have been removed; an operation score output means having a model for outputting weights to be applied to the effect of the operation on the user in the target time slot on the possibility of the user performing the action in the target time slot and the amount of action result for each user, which indicates the amount of result when the user performs the action, in response to input of at least a user identifier for each user, user attribute data, and an identifier for the target time slot; and a model for calculating an operation score indicating the effect of the operation in the target time slot for each user, based on the weights, the effect for each user, and the amount of action result for each user output from the model for outputting the weights, the operation score output means inputting the user identifiers, user attribute data, and identifier for the target time slot for each of a plurality of users included in the candidate user list into the model for outputting the weights to obtain the weights, and inputting the obtained weights, the effect for each user included in the candidate user list, and the amount of action result for each user into the model for calculating the operation score, to output the operation score for each user included in the candidate user list; a target user determination means for determining a target user of the operation in the target time slot from among users included in the candidate user list based on the operation score; An information processing device comprising:

6. The system further includes an effect estimation means for estimating an effect of the operation on the user in the target time slot on whether the user executes the action. The information processing device according to claim 5 .

7. the effect estimation means estimates the effect of the operation using an effect estimation model that outputs a causal score indicating the effect of the operation on the user in response to an input of one or more attributes related to the user; The information processing device according to claim 6.

8. the effect estimation model is created based on training data in which a score based on a statistic relating to an execution rate of the action by a user who has received the operation among a plurality of users having a predetermined attribute and a statistic relating to an execution rate of the action by a user who has not received the operation among the plurality of users is defined as a score indicating an effect of the operation on a user who has the attribute; The information processing device according to claim 7.

9. a performance indicator calculation means for calculating a performance indicator of the operation in the time slot based on the action performance amount of the user who performed the action by receiving the operation in the time slot, The model that outputs the weight is a reinforcement learning model generated and / or updated using training data including the performance indicator in the target time slot as a reward related to the target time slot. The information processing device according to claim 1 .

10. The model that outputs the weight is a reinforcement learning model that is generated and / or updated using training data that includes, as a reward for the target time slot, the sum of one or more of the performance indicators obtained in the target time slot and a time slot subsequent to the target time slot among the multiple time slots included in the unit time. The information processing device according to claim 9.

11. The target user determination means extracts users from the users included in the candidate user list based on the operation score, up to a predetermined number of users for the target time slot, and determines the extracted users as target users for the operation in the target time slot. The information processing device according to claim 1 .

12. The model that outputs the weights is a reinforcement learning model that is generated and / or updated using training data that includes a target user determined as a target of the operation in the target time slot as a behavior related to the target time slot. The information processing device according to claim 1 .

13. The model that outputs the weights is a reinforcement learning model that is generated and / or updated using training data that includes a user identifier for each of a plurality of users included in the candidate user list, user attribute data, an action outcome amount when the user performs the action, and an identifier of the target time slot as a state related to the target time slot. The information processing device according to claim 1 .

14. the operation score output means obtains the weight by inputting the result of the operation in the preceding time slot into a model that further outputs the weight. The information processing device according to claim 1 .

15. The candidate user list acquisition means acquires a candidate user list from which at least one of users who have received the operation a predetermined number of times or more in a time slot preceding the target time slot among the multiple time slots included in the unit time, users who have received the operation in the preceding time slot and executed the action, and users who have been determined as targets of the operation in the preceding time slot have been removed. The information processing device according to claim 1 .

16. the operation being contacting the user to prompt the user to pay the outstanding amount; the action outcome amount is an amount owed to the user that is to be paid by the user; The information processing device according to claim 1 .

17. A computer for determining a user to be a target of an operation for prompting a user to perform a predetermined action, a target time slot determination step of determining a target time slot from among a plurality of time slots included in a predetermined unit time; a candidate user list acquisition step of acquiring a candidate user list from which users who satisfy a predetermined condition in a time slot preceding the target time slot among the plurality of time slots included in the unit time have been removed; an operation score output step using a model that outputs weights to be applied to an action execution possibility for each user, which indicates the possibility that the user will execute the action if the user receives the operation in the target time slot, and an action result amount for each user, which indicates the result amount when the user executes the action, in response to input of at least a user identifier for each user, user attribute data, and an identifier for the target time slot; and a model that calculates an operation score that indicates the effect of the operation in the target time slot for each user, based on the weights output from the model that outputs the weights, the action execution possibility for each user, and the action result amount for each user, the operation score output step using a model that outputs the weights to obtain the weights, and an operation score output step that outputs the operation score for each user included in the candidate user list by inputting the user identifiers, user attribute data, and identifiers for the target time slot for each user into the model that outputs the weights, and a target user determination step of determining a target user of the operation in the target time slot from among users included in the candidate user list based on the operation score; How to do it.

18. A computer for determining a user to be a target of an operation for prompting a user to perform a predetermined action, a target time slot determination step of determining a target time slot from among a plurality of time slots included in a predetermined unit time; a candidate user list acquisition step of acquiring a candidate user list from which users who satisfy a predetermined condition in a time slot preceding the target time slot among the plurality of time slots included in the unit time have been removed; an operation score output step using a model that outputs weights to be applied to the effect of the operation on the user in the target time slot on the possibility of the user performing the action in the target time slot and the amount of action result for each user, which indicates the amount of result when the user performs the action, in response to input of at least a user identifier for each user, user attribute data, and an identifier for the target time slot; and a model that calculates an operation score indicating the effect of the operation in the target time slot for each user based on the weights, the effect for each user, and the amount of action result for each user output from the model that outputs the weights, the user identifier, user attribute data, and identifier for the target time slot for each of a plurality of users included in the candidate user list are input to the model that outputs the weights to obtain the weights, and an operation score output step that outputs the operation score for each user included in the candidate user list by inputting the obtained weights, the effect for each user included in the candidate user list, and the amount of action result for each user into a model that calculates the operation score; a target user determination step of determining a target user of the operation in the target time slot from among users included in the candidate user list based on the operation score; How to do it.

19. A computer for determining a user to be a target of an operation for prompting the user to perform a predetermined action, a target time slot determining means for determining a target time slot from among a plurality of time slots included in a predetermined unit time; a candidate user list acquisition means for acquiring a candidate user list from which users who satisfy a predetermined condition in a time slot preceding the target time slot among the plurality of time slots included in the unit time have been removed; an operation score output means having a model for outputting weights to be applied to an action execution possibility for each user, which indicates the possibility that the user will execute the action if the user receives the operation in the target time slot, and an action result amount for each user, which indicates the result amount when the user executes the action, in response to input of at least a user identifier for each user, user attribute data, and an identifier for the target time slot; and a model for calculating an operation score indicating the effect of the operation in the target time slot for each user, based on the weights output from the model for outputting the weights, the action execution possibility for each user, and the action result amount for each user, the operation score output means inputting the user identifiers, user attribute data, and identifier for the target time slot for each of a plurality of users included in the candidate user list into the model for outputting the weights to obtain the weights, and inputting the obtained weights, the action execution possibility for each user included in the candidate user list, and the action result amount for each user into the model for calculating the operation score, to output the operation score for each user included in the candidate user list; a target user determination means for determining a target user of the operation in the target time slot from among users included in the candidate user list based on the operation score; A program that functions as a

20. A computer for determining a user to be a target of an operation for prompting the user to perform a predetermined action, a target time slot determining means for determining a target time slot from among a plurality of time slots included in a predetermined unit time; a candidate user list acquisition means for acquiring a candidate user list from which users who satisfy a predetermined condition in a time slot preceding the target time slot among the plurality of time slots included in the unit time have been removed; an operation score output means having a model for outputting weights to be applied to the effect of the operation on the user in the target time slot on the possibility of the user performing the action in the target time slot and the amount of action result for each user, which indicates the amount of result when the user performs the action, in response to input of at least a user identifier for each user, user attribute data, and an identifier for the target time slot; and a model for calculating an operation score indicating the effect of the operation in the target time slot for each user, based on the weights, the effect for each user, and the amount of action result for each user output from the model for outputting the weights, the operation score output means inputting the user identifiers, user attribute data, and identifier for the target time slot for each of a plurality of users included in the candidate user list into the model for outputting the weights to obtain the weights, and inputting the obtained weights, the effect for each user included in the candidate user list, and the amount of action result for each user into the model for calculating the operation score, to output the operation score for each user included in the candidate user list; a target user determination means for determining a target user of the operation in the target time slot from among users included in the candidate user list based on the operation score; A program that functions as a

Images