Credit learning device, credit learning method, credit estimation device, credit estimation method, and program

Abstract

In order to address the problem of improving the accuracy of estimating an individual's creditworthiness, a learning device 1 is provided with: a first training unit 23a that trains a first model on first training data including combinations of first attribute data groups relating to users who use first service and labels corresponding to a first score that changes with the risk borne by a service provider when the users use the first service; and a second training unit 23b that trains a second model, which has some parameters of the configuration of the trained first model as fixed parameters, on second training data including combinations of second attribute data groups relating to users who use second service and labels corresponding to a second score that changes with the risk borne by the service provider when the users use the second service.

Description

Trust learning device, trust learning method, trust estimation device, trust estimation method, and program

[0001] The present disclosure relates to techniques for estimating a user's trustworthiness.

[0002] Conventionally, a determination device has been proposed that includes an acquisition unit that acquires asset information, which is information that arbitrarily combines individual assets owned by a first user, and the first user's behavioral history on a network, and a determination unit that estimates information regarding a second user's assets based on the asset information and behavioral history on the network of the first user acquired by the acquisition unit (see Patent Document 1).

[0003] Furthermore, in the past, when estimating credit risk, it has been proposed to perform transfer learning from the domain of credit card lending or debt consolidation to the domain of small business lending (see Non-Patent Document 1).

[0004] JP 2018-110017 A

[0005] Hendra Suryanto, et al. , “Transfer Learning in Credit Risk”, Joint European Conference on Machine Learning and Knowledge Discovery in Databases (pp. 483-498), Springer, Cham, 2019.

[0006] Various techniques have been proposed to determine the trustworthiness of users for a given service. However, while these techniques have been effective in accurately determining the trustworthiness of individuals, there is still room for improvement in terms of flexibly improving the accuracy of trustworthiness estimation in specific domains.

[0007] In view of the above-mentioned problems, an object of the present disclosure is to flexibly improve the accuracy of estimating an individual's creditworthiness in a specific domain.

[0008] An example of the present disclosure is a trust learning device comprising: a first learning means for training a first model with first training data including a combination of a first group of attribute data related to a user who uses a first service and a label corresponding to a first score that changes depending on the degree of risk a provider of the first service takes when the user uses the first service; and a second learning means for training a second model having some of the parameters of the trained first model as fixed parameters with second training data including a combination of a second group of attribute data related to a user who uses a second service different from the first service and a label corresponding to a second score that changes depending on the degree of risk a provider of the second service takes when the user uses the second service.

[0009] Another example of the present disclosure is a trust learning device comprising: a first learning means for training a first model with first training data including a combination of a first group of attribute data related to users who are included in a first segment among users using a first service and a label corresponding to a score that changes depending on the degree of risk a provider of a second service that is different from the first service takes when the user uses the second service; and a second learning means for training a second model having some parameters of the trained first model as fixed parameters with second training data including a combination of a second group of attribute data related to users who are included in a second segment that is different from the first segment among users using the first service and a label corresponding to a score that changes depending on the degree of risk a provider of the second service takes when the user uses the second service.

[0010] Another example of the present disclosure is a trust estimation device that includes an estimation means for estimating the second score to be set for a target user by inputting a group of attribute data related to the target user into the second model that has been trained by the above-mentioned learning method.

[0011] 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.

[0012] According to the present disclosure, it is possible to flexibly improve the accuracy of estimating an individual's creditworthiness in a specific domain.

[0013] FIG. 1 is a schematic diagram showing the configuration of an information processing system according to a first embodiment. FIG. 2 is a diagram showing an outline of the functional configuration of an information processing device according to the first embodiment. FIG. 3 is a simplified diagram showing the concept of a decision tree of a machine learning model employed in the first embodiment. FIG. 4 is a diagram showing an outline of a first model generated when a framework based on a neural network is employed in the first embodiment. FIG. 5 is a diagram showing an outline of a second model generated when a framework based on a neural network is employed in the first embodiment. FIG. 6 is a simplified diagram of a score estimation process according to the first embodiment. FIG. 7 is a flowchart showing the flow of a machine learning process according to the first embodiment. FIG. 8 is a flowchart showing the flow of a score estimation process according to the first embodiment. FIG. 9 is a schematic diagram showing the configuration of an information processing system according to a second embodiment. FIG. 10 is a diagram showing an outline of the functional configuration of an information processing device according to the second embodiment. FIG. 11 is a flowchart showing the flow of a machine learning process according to the second embodiment.

[0014] Hereinafter, embodiments of an information processing device, method, and program according to the present disclosure will be described with reference to the drawings. However, the embodiments described below are merely examples, and the information processing device, method, and program according to the present disclosure are not limited to the specific configurations described below. When implementing, specific configurations according to the mode of implementation may be appropriately adopted, and various improvements and modifications may be made. The present invention may appropriately adopt at least a portion of the configurations in the embodiments and variations described below. Note that in the embodiments and variations described below, the trust learning device, trust learning method, trust estimation device, and trust estimation method will simply be referred to as the learning device, learning method, estimation device, and estimation method.

[0015] First Embodiment In this embodiment, a description is given of a case in which the technology according to the present disclosure is implemented to convert a label indicating deferred payment risk into a label capable of estimating risk for a different financial-related service by relearning (retraining) a data set related to a financial service different from the deferred payment service based on a machine learning model generated for calculating a user's default risk for a deferred payment service (the "deferred payment risk" described below). However, the technology according to the present disclosure can be widely used for a technology for calculating a user's score for a different service by utilizing a model generated using a data set related to the service to calculate the user's score for the service. The application of the present disclosure is not limited to the examples shown in the embodiments. For example, the technology according to the present disclosure may be implemented to enable risk estimation for multiple different services belonging to fields other than finance.

[0016] <<System Configuration>> Fig. 1 is a schematic diagram showing the configuration of an information processing system according to this embodiment. In the information processing system according to this embodiment, an information processing device 1 and one or more service providing systems 5 are connected to each other so that they can communicate with each other. A user is a user of a service provided by the service providing system 5, and receives the service by accessing the service providing system 5 from a user terminal. Here, the information processing device 1 represents a learning device, an estimation device, or a combination thereof.

[0017] The information processing device 1 is a computer including a CPU (Central Processing Unit) 11, a ROM (Read Only Memory) 12, a RAM (Random Access Memory) 13, a storage device 14 such as an EEPROM (Electrically Erasable and Programmable Read Only Memory) or an HDD (Hard Disk Drive), a communication unit 15 such as a NIC (Network Interface Card), etc. However, the specific hardware configuration of the information processing device 1 can be omitted, replaced, or added as appropriate depending on the embodiment. Furthermore, 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.

[0018] The service providing system 5 is a computer including a CPU, ROM, 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] Services provided by the service providing system 5 include, for example, online shopping services, online reservation services, loan services, banking services, credit card services, deferred payment services, electronic money payment services, advertising services, operation center services, subscription services, map information services, etc. Note that "deferred payment" is not limited to services known as Buy Now Pay Later (BNPL), and may include any purchase of goods / services with deferred payment.

[0020] The services provided by the service providing system 5 are not limited to the examples given in this embodiment. The service providing system 5 notifies the information processing device 1 of user-related data when providing a service. Here, the user-related data includes service usage history data by the user. The content of the service usage history data varies depending on the content of the service, and may include, for example, history data of the user's location information, payment history data of credit card usage amounts / post-payment usage amounts, electronic money usage history data, transaction history data (including purchase history data of products, etc.), reservation history data, operation history data for the user from the operation center, etc.

[0021] In this embodiment, among the various services provided by the service providing system 5, financial-related services (i.e., services belonging to the "finance" category in business classifications, domains (segments, categories, and domains)) including deferred payment services, loan services, banking services, and credit card services are described. Based on a machine learning model generated to calculate a user's deferred payment risk for a deferred payment service (corresponding to a "first service" described below), re-learning is performed using a dataset for a different financial-related service (e.g., a loan service, corresponding to a "second service" described below), and a label indicating deferred payment risk is converted into a label corresponding to the risk for the different financial-related service. Here, "re-learning" refers to transfer learning, additional learning, etc., depending on the learning mode. Note that the group of services in this embodiment may be a group of services sharing a membership program (user identifier), a group of services sharing a points program (electronic value such as points that can be awarded or used), or a group of services that grant benefits that can be used in one service upon use of the other service.

[0022] The information processing device according to the present disclosure uses a machine learning model that inputs user attribute data and the like and outputs a credit score (in this embodiment, a score indicating deferred payment risk). The attribute data used as input here may also include data indicating default (non-payment of debt). By using various attribute data as input, the information processing device 1 according to the present disclosure calculates a universal and generalized credit score that uniformly reflects the user's overall attributes. The calculated credit score can be used for credit screening (determination involving approval or rejection of deferred payment) in deferred payment services, etc., but as described above, in this embodiment, the calculated credit score is used to determine whether or not to propose payment method options to the user.

[0023] Here, the user attribute data includes actual attribute data and estimated attribute data. The attribute data includes data expressed in data formats such as a score (e.g., a continuous value between 0 and 1) or a label (e.g., a binary value corresponding to presence / absence or yes / no). However, the format of the attribute data is not limited to the examples in this disclosure. Furthermore, the attribute data may include, for example, online service usage status and electronic value usage status including points. Furthermore, the online service usage status may include at least one of the number of cancellations, cancellation rate, and number of orders for an online shopping service or an online reservation service.

[0024] The factual attribute data is data indicating factual attributes of a user that can be confirmed as factual based on user-provided data obtained by the user himself / herself and historical data collected about the user. Examples of user-provided data include registration data registered by the user himself / herself, including the user's name, email address, telephone number, address, place of employment, place of education, etc., and data obtained as a result of the user's own responses to questionnaires, etc. Examples of historical data include the usage history data of the e-commerce service provided by the service providing system 5 described above. The factual attribute data is preferably data obtained by converting the aforementioned user-provided data and historical data into a data format suitable for marketing and / or analytical purposes. For example, examples of factual attribute data that can be obtained based on the usage history data include genres / categories and brands of products / services frequently used by the user, as well as commercial locations, recreational spots, and tourist destinations frequently visited by the user.

[0025] The estimated attribute data is data indicating an inferred attribute obtained by estimation based on user-provided data, history data, fact attribute data, etc. In this embodiment, the estimated attribute data includes a user's personality estimated or predicted using machine learning technology. The estimated attribute data is preferably data on an attribute that influences user behavior when used for targeting.

[0026] In this embodiment, for example, when a decision tree model is used, an importance is set for each attribute data. The importance indicates the degree of correlation between the attribute data and the credit score when the attribute data is used to calculate the credit score, and each time the appropriateness of the credit score is evaluated by the learning unit 23 described below, the model parameters are adjusted so that the credit score becomes a more appropriate value. The importance corresponding to each attribute data corresponds to, for example, a weight corresponding to each node (each regression tree) in a model for calculating the credit score, such as a decision model described below, and is determined appropriately during the process of calculating the credit score.

[0027] Here, the attribute data group may include demographic attributes, behavioral attributes, or psychographic attributes. Demographic attributes include, for example, the user's gender, family structure, age, etc. Behavioral attributes may be based on service usage history data, such as whether or not a user has used cash advances, whether or not they have used revolving credit, deposit and withdrawal history for a specific account, commercial transaction history for some product / service including gambling or lotteries (which may include online transaction history in online marketplaces, etc.), and the user's movement history using location information or place information. Psychographic attributes include, for example, preferences for gambling or lotteries. However, available user attributes are not limited to the examples in this embodiment. For example, "time required for operation (such as a call)" and "credit card usage amount / postpaid payment usage amount" from an operation center service, etc. may also be used as attributes. Demographic attributes and behavioral attributes may be treated as factual attributes. Psychographic attributes may be treated as estimated attributes. Note that attributes similar to demographic attributes may be estimated based on factual attributes based on user-provided data or historical data. Similarly, attributes similar to behavioral attributes may be estimated based on factual attributes based on user-provided data or historical data. Psychographic attributes may be factual attributes based on user-provided data, including, for example, the results of user input.

[0028] 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 fact attribute determination unit 21, an estimated attribute determination unit 22, a learning unit 23, and a score estimation unit 24, by a program recorded in the storage device 14 being read into the RAM 13 and executed by the CPU 11, which controls each piece of hardware included in the information processing device 1. Note that in this embodiment and other embodiments described below, 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.

[0029] The fact attribute determination unit 21 determines fact attribute data that can be confirmed as fact for the user based on user-provided data provided by the user himself / herself and / or the user's history data. In this embodiment, the fact attribute determination unit 21 determines the fact attribute data for the user by using a method such as aggregating the user-provided data and / or history data, determining corresponding attributes by referring to other data such as a map, or using the user-provided data and / or history data as is. Note that, although this embodiment employs a method of determining the fact attribute data for the user based on the user-provided data and / or the user's history data, the fact attribute data for the user may also be obtained by other methods.

[0030] The estimated attribute determination unit 22 determines estimated attribute data for the target user based on user-related data including at least one or more fact attribute data determined for the target user by the fact attribute determination unit 21. In this embodiment, the estimated attribute determination unit 22 determines an estimated attribute based on an output value obtained by inputting user-related data including one or more fact attribute data related to the target user into an attribute estimation model, which is a machine learning model. In this embodiment, the output value from the attribute estimation model is a value indicating the probability that the target user has a predetermined estimated attribute, and the estimated attribute determination unit 22 determines that the target user has the estimated attribute if the output value obtained from the attribute estimation model is within a predetermined range. If it is determined that the target user has the predetermined estimated attribute, the estimated attribute determination unit 22 sets a label for the attribute data estimated for the target user to a value indicating the presence or absence of the attribute or the type of the attribute. Furthermore, the estimated attribute data may be represented by a score instead of a label. In this case, the estimated attribute determination unit 22 sets a value indicating the degree (probability) to which the estimated attribute is applicable to the score of the attribute data estimated for the target user. The degree may be the output value of the attribute estimation model.

[0031] The estimated attribute determination unit 22 can also estimate missing (insufficient) attribute data based on other attribute data. Therefore, when attribute data is missing or insufficient for a certain user, the attribute data estimated by the estimated attribute determination unit 22 may be used instead of the actual attribute data. The attribute data complementation by the estimated attribute determination unit 22 can also be applied to complementing attribute data of a user (e.g., a light user of the first service) whose attribute data quantity or quality is insufficient in the second embodiment described below.

[0032] The learning unit 23 generates and / or updates a score estimation model by the score estimation unit 24. The score estimation model is a machine learning model that, when one or more attribute data (attribute data group) related to a target user are input, outputs a credit score indicating the degree of risk that the target user will not properly settle the deferred payment if the target user uses deferred payment. By using such a machine learning model, in this embodiment, it is possible to obtain a credit score that takes into account factors specific to deferred payment, rather than just a credit score. Furthermore, in this embodiment, an example is shown in which a credit score is adopted in which the larger the credit score value, the higher the creditworthiness (lower the risk) and the smaller the credit score value, the lower the creditworthiness (higher the risk), but the relationship between the size of the score value and the level of risk may be reversed.

[0033] When generating and / or updating the score estimation model, the learning unit 23 creates training data for each user, including a combination of the user's attribute data group and credit scores of users with the same attribute data group. The learning unit 23 then generates and / or updates the score estimation model based on the training data. As described above, the attribute data group input to the score estimation model includes fact attribute data determined by the fact attribute determination unit 21 and estimated attribute data estimated by the estimated attribute determination unit 22 based on user-related data including the fact attribute data. The attribute data is combined with labels indicating the risk and corresponding score of the corresponding user and input to the learning unit 23 as training data. In this embodiment, the credit score set in the training data is a credit score determined based on payment history data for deferred payment of a user corresponding to the combination of user attributes. Here, the payment history data for deferred payment includes data indicating the presence or absence of default (non-performance of debt) in deferred payment and the default amount. In this case, the credit score may be a rule-based credit score or a manually set (annotated) credit score. It may also be a credit score that was previously output by a score estimation model and then corrected by an administrator or the like.

[0034] A machine learning model generation / update framework that can be employed as a score estimation model or the like when implementing the technology disclosed herein is based on an ensemble learning algorithm, for example. For example, a machine learning framework (e.g., LightGBM) based on a gradient boosting decision tree (GBDT) may be employed 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 subsequent weak learners (weak classifiers). The predicted value here refers to, for example, a predicted value of a credit score. In addition to LightGBM, the framework may also employ boosting methods such as XGBoost and CatBoost. A framework using a decision tree allows for the generation / update of a machine learning model with relatively high performance with less effort in parameter adjustment compared to a framework using a neural network. However, the framework for generating / updating a machine learning model that can be used to implement the technology disclosed herein is not limited to the example described in this embodiment. For example, instead of a gradient boosting decision tree, a random forest or other learning device may be used as the learning device. Alternatively, a learning device that is not a so-called weak learning device, such as a neural network, may be used. Furthermore, particularly when a learning device that is not a so-called weak learning device, such as a neural network, is used, ensemble learning may not be used.

[0035] FIG. 3 is a simplified diagram illustrating the concept of a decision tree in a machine learning model employed as a score estimation model or the like in this embodiment. When employing a gradient boosting machine learning framework based on a decision tree algorithm, the branching conditions of each node in the decision tree are optimized. Specifically, in the gradient boosting machine learning framework based on a decision tree algorithm, credit scores are calculated for each user group having attributes indicated by two child nodes branching from a single parent node, and the branching conditions of the parent node are optimized so that the difference between these credit scores is large (e.g., so that the difference is maximized or exceeds a predetermined threshold), i.e., so that the two child nodes branch cleanly. For example, if the attribute indicated as the branching condition of a 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, recursively optimizing the branching conditions of all nodes in the decision tree can improve the accuracy of credit score estimation based on attribute data groups.

[0036] In this embodiment, the learning unit 23 includes at least a first learning unit 23a and a second learning unit 23b.

[0037] The first learning unit 23a trains a first model (first score estimation model) with first training data including a combination of a first attribute data group related to a user who uses a first service (in this embodiment, a deferred payment service) and a label corresponding to a score (the user's credit score in the first service) that changes depending on the degree of risk (deferred payment risk) that the provider of the first service will bear if the user uses the first service. As described above, in this embodiment, the first service is a deferred payment service. The risk that the provider of the first service will bear is a deferred payment risk, which is determined based on the payment history in the deferred payment service and is expressed as some kind of index (in this embodiment, a label that changes depending on the magnitude of the deferred payment risk) of the risk that the user will not properly settle the deferred payment if the user uses the deferred payment service.

[0038] The second learning unit 23b trains a second model (second score estimation model) having, as fixed parameters, part of the configuration (parameters such as intermediate layers, nodes, and other weights) of the first model trained by the first learning unit 23a, using second training data including a combination of a second attribute data group related to a user using a second service (a loan service in this embodiment) and a label corresponding to a second score that changes depending on the degree of risk assumed by the provider of the second service when the user uses the second service. The second service is a service different from the first service, and in this embodiment, an example in which the second service is a loan service will be mainly described. However, the second service may be any service for which a correlation is confirmed or inferred between the risk assumed by the provider of the first service when a user uses the first service and the risk assumed by the provider of the second service when a user uses the second service, and is not limited to the example described in this embodiment. For example, a combination of services may be employed in which, for the same user group, a score corresponding to the risk in the first service and a score corresponding to the second service are correlated.

[0039] When a machine learning framework based on gradient boosting decision trees is adopted for the first model and the second model, the second learning unit 23b causes the second model, which has parameters of some of the nodes of the first model as fixed parameters, to learn the second training data, and updates the parameters (weights) of the nodes that were not fixed. In addition, at this time, the second learning unit 23b may cause the second model, in which part of the tree structure of the first model is fixed, to learn the second training data.

[0040] Here, when determining the parameters of nodes to be fixed, a method may be adopted in which, for example, nodes with high importance, i.e., nodes corresponding to branching conditions (attributes) that most affect the estimation (output) by the model, or nodes that have an influence on the output that is greater than a predetermined standard, are preferentially fixed (see FIG. 3 ). Furthermore, when updating the parameters, for example, parameters (weights) excluding frozen layers and nodes can be updated so as to minimize loss in the data set and network output of the target domain (repurposed domain; in this case, the second service).

[0041] When a machine learning framework based on a neural network is adopted for the first model and the second model, the second learning unit 23b causes a second model having some parameters of an intermediate layer (hidden layer) of the first model as fixed parameters to learn the second teacher data. In this case, the second learning unit 23b may cause the second model, in which multiple copies of the trained first model are arranged in parallel and have a common output layer, to learn the second teacher data using the second model having some parameters of the intermediate layer (hidden layer) fixed.

[0042] 4 is a diagram illustrating an overview of a first model generated by the first learning unit 23a when a machine learning framework based on a neural network is adopted in this embodiment. Using the machine learning framework based on a neural network, the first learning unit 23a trains the first model with first training data including a combination of a first attribute data group related to a user who uses a first service and a label that changes depending on the degree of risk the provider of the first service assumes when the user uses the first service. In the example illustrated in FIG. 4, an example of a neural network including intermediate layers (hidden layers) a to c is illustrated as the first model.

[0043] 5 is a diagram illustrating an overview of a second model generated by the second learning unit 23b when a neural network-based machine learning framework is employed in this embodiment. The second learning unit 23b generates the second model by connecting multiple copies of a neural network trained on a dataset of a source domain (the source of the repurposed data; in this case, the first service) in parallel and retraining some of the intermediate layers. Referring to the example illustrated in the drawing, for example, the second learning unit 23b copies multiple neural networks (four neural networks, NN1 to NN4 in the illustrated example) including intermediate layers a to c included in a first model (see FIG. 4) trained on a dataset (first training data) related to the first service, which is the source domain, and connects these multiple neural networks NN1 to NN4 in parallel, with the input layer and output layer shared. Then, a second model is created by freezing some of the intermediate layers (in the example shown in FIG. 5, intermediate layers a to c of NN1, intermediate layers a and b of NN2, and intermediate layer a of NN3, shown by solid lines).The second learning unit 23b then executes a learning process on the second model using a dataset (second training data) related to the target domain (second service), thereby updating the intermediate layers that were not fixed (in the example shown in FIG. 5, intermediate layer c of NN2, intermediate layers b and c of NN3, and intermediate layers a to c of NN4, shown by dashed lines).

[0044] The method for fixing the intermediate layer described with reference to FIG. 5 is merely an example in this embodiment, and various variations in the method for fixing the intermediate layer may be adopted (see, for example, Non-Patent Document 1). In the example described with reference to FIG. 5, the second model was created by connecting multiple copies of a neural network trained with a data set of the source domain in parallel. However, the second learning unit 23b may fix some of the intermediate layers of a single neural network trained with a data set of the source domain (i.e., multiple copies of the intermediate layer are not created) and perform re-learning. The output layer may also be tuned by the second learning unit 23b during the re-learning process.

[0045] The score estimation unit 24 estimates a score (credit score, creditworthiness) such as a second score of the target user based on a group of attribute data including fact attributes and estimated attributes related to the target user. In this case, the score estimation unit 24 may perform some processing (normalization, ranking, labeling, etc.) on the determined fact attributes and / or estimated attributes to include them as part of the group of attribute data, or may use other types of scores (e.g., so-called credit scores) or labels calculated using the determined fact attributes and / or estimated attributes as all or part of the group of attribute data. Here, other machine learning models may be used to calculate other types of scores or labels. The score estimation unit 24 may estimate a score related to a virtual user by inputting a group of attribute data of the target user, who is a virtual user with no usage history or transaction history in the service, into a second model.

[0046] 6 is a simplified diagram of the score estimation process according to this embodiment. In this embodiment, the score estimation unit 24 estimates (calculates) the score of a user by inputting a group of attribute data of the user into a score estimation model. Here, the output value of the score estimation model is, for example, a score with 0 as the minimum value and 1 as the maximum value. In this embodiment, the score estimation unit 24 inputs a group of attribute data related to the target user into a second model generated by the learning unit 23, thereby estimating a second score that changes depending on the degree of risk assumed by the provider of the second service when the target user uses the second service.

[0047] <<Processing Flow>> Next, a processing flow executed by the information processing device according to this embodiment will be described. Note that the specific content and processing order of the processing described below are an example for implementing the present disclosure. The specific processing content and processing order may be selected as appropriate depending on the embodiment of the present disclosure.

[0048] 7 is a flowchart showing the flow of the machine learning process according to this embodiment. The process shown in this flowchart is executed periodically or at a timing designated by an administrator.

[0049] In steps S101 and S102, a first score estimation model is generated. The first learning unit 23a creates first training data for a group of users of a first service, the first training data including a combination of a group of attribute data for each user accumulated in the past and a label predetermined for the corresponding user (step S101). Then, the first learning unit 23a performs a training process for the first score estimation model based on the created first training data (step S102). Then, the process proceeds to step S103.

[0050] In steps S103 and S104, a second score estimation model is generated. The second learning unit 23b generates second training data for a group of users of the second service, the second training data including a combination of a group of attribute data for each user accumulated in the past and a label predetermined for the corresponding user (step S103). The second learning unit 23b then uses the second training data to re-train the second score estimation model, which has some of the parameters of the first score estimation model generated in step S102 as fixed parameters, thereby constructing a second score estimation model to be used for score estimation by the score estimation unit 24 (step S104). After that, the processing shown in this flowchart ends.

[0051] 8 is a flowchart showing the flow of the score estimation process according to this embodiment. The process shown in this flowchart is executed for each target user periodically or at a specified timing.

[0052] In steps S201 and S202, actual attribute data and estimated attribute data are determined. The actual attribute determination unit 21 determines actual attribute data related to the target user based on the user-provided data and / or history data of the target user (step S201). Then, the estimated attribute determination unit 22 determines estimated attribute data related to the target user based on at least the actual attribute data determined in step S201 (step S202). Then, the process proceeds to step S203.

[0053] In steps S203 and S204, the score of the target user is estimated. The score estimation unit 24 determines a group of attribute data including the actual attribute data determined in step S201 and the estimated attribute data determined in step S202 (step S203). The score estimation unit 24 then inputs the group of attribute data determined in step S203 into a second estimation model and acquires the output value as a score that changes depending on the degree of risk that the target user will not properly settle the deferred payment if the target user uses deferred payment (step S204). However, the score estimation method is not limited to the example described in this embodiment. For example, the score may include a value calculated by inputting the group of attribute data into a predetermined function or statistical model other than a machine learning model. The processing shown in this flowchart then ends.

[0054] <<Variations>> In this embodiment, the learning unit 23 may further include a third learning unit, a fourth learning unit, ..., an nth (n>2) learning unit. Here, the nth learning unit trains an nth model (nth score estimation model) having a configuration of fixed parameters, etc., that is a part of the configuration of parameters, etc., of the n-1th model trained by the n-1th learning unit, using nth training data that includes a combination of an nth attribute data group related to a user who uses an nth service different from the first, second, ..., n-1th service, and a label corresponding to an nth score that changes depending on the degree of risk the provider of the nth service assumes if the user uses the nth service. In this case, the score estimation unit 24 may estimate the score of the target user by inputting a group of attribute data including fact attributes and estimated attributes related to the target user into the nth model trained by the nth learning unit. In this embodiment, a combination of services may be adopted in which, for the same user group, there is a correlation between the score corresponding to the risk in the (n-1)th service and the score corresponding to the nth service.

[0055] In this embodiment, the nth learning unit may determine two or more models from a model group including the (n-1)th model and the (n-2)th model, and may cause the nth model, which has some parameters of each of the determined models as fixed parameters, to learn the nth teacher data. In this case, the nth learning unit may determine two or more models including at least the (n-1)th model from the model group.

[0056] In this embodiment, the second service different from the first service, or the nth service different from the first, second, ... (n-1)th service, may be determined based on the respective attribute data sets of users who have usage histories or transaction histories in each service. For example, the second service different from the first service may be another service that includes more than a predetermined percentage of users whose attribute data sets are similar or identical to those of users of the first service. Here, the similarity between users may be based on the percentage of attribute data shared between the users, whether or not certain attribute data is shared, whether or not there are more than a predetermined number of combinations of similar embedded expressions related to users of each service, etc. Furthermore, for example, the nth service different from the (n-1)th service may be another service that includes more than a predetermined percentage of users whose attribute data sets are similar or identical to those of users of at least the (n-1)th service.

[0057] <<Effects>> According to this embodiment, by re-training a machine learning model that is trained according to the payment status of a deferred payment service and is capable of outputting a credit score using data from a service other than the deferred payment service, it is possible to improve the accuracy of estimating an individual's creditworthiness even for services with a relatively small amount of training data set.

[0058] <Second embodiment> This embodiment relates to a technology in which a model is trained using attributes of users (including the transaction history of the users) included in one of two different user segments that use a common service as a dataset, and the model is retrained using the dataset of the other segment.

[0059] More specifically, in this embodiment, the technology according to the present disclosure is described in a manner in which a machine learning model for calculating the default risk of users of financial-related services (e.g., the deferred payment risk described in the first embodiment) is trained using a dataset of attributes of users included in one of multiple user segments that use commerce-related services, and then retrained based on the machine learning model using datasets of users included in other segments that use the commerce-related services. However, the technology according to the present disclosure can be widely used in techniques for calculating scores for users of different services using models generated using multiple datasets that belong to multiple different segments of a service, and the application of the present disclosure is not limited to the examples shown in the embodiments. For example, the technology according to the present disclosure may be implemented to enable risk estimation for multiple different services in fields other than finance.

[0060] <<System Configuration>> FIG. 9 is a schematic diagram showing the configuration of an information processing system according to this embodiment. In the information processing system according to this embodiment, an information processing device 1b and one or more service providing systems 5 are communicably connected to each other. A user is a user of a service provided by the service providing system 5 and receives the service by accessing the service providing system 5 from a user terminal. Here, the information processing device 1b represents a learning device, an estimation device, or a combination thereof. The configuration of the information processing system according to this embodiment is generally similar to the configuration of the information processing system according to the first embodiment described with reference to FIG. 1, except that the functional configuration of the information processing device 1b, which will be described later with reference to FIG. 10, is different from that of the information processing device 1 according to the first embodiment. Therefore, a description of the common parts will be omitted.

[0061] In this embodiment, a machine learning model is trained to calculate the default risk of users in a financial-related service (corresponding to a "second service" described below), which is a different business classification or domain, using a dataset of attributes of users in one of multiple user segments that use a commercial transaction-related service (corresponding to a "first service" described below) among the various services provided by the service providing system 5. Based on the machine learning model, a dataset of users in another segment that uses the commercial transaction-related service is used to calculate the default risk of users in the other segment. Here, the commercial transaction-related service is a service that belongs to the "commercial transaction" business classification or domain (segment, category, domain), such as an online shopping service or an online reservation service. The financial-related service is a service that belongs to the "finance" business classification or domain, such as a deferred payment service. The term "relearning" refers to transfer learning, additional learning, or the like, depending on the learning mode. In addition, the group of services in this embodiment may be a group of services that share a common membership program (user identifier), a group of services that share a common points program (electronic value such as points that can be awarded or used), or a group of services that grant benefits that can be used in one service when one service is used in the other service.

[0062] The information processing device 1b according to the present disclosure uses a machine learning model that inputs user attribute data, etc., obtained in a first service (in this embodiment, an online shopping service) and outputs a credit score (in this embodiment, a score indicating deferred payment risk) for a second service (in this embodiment, a deferred payment service). Details of the attribute data used as input here are the same as those described in the first embodiment, including the fact that it may include actual attribute data and estimated attribute data, and therefore will not be described again. Furthermore, a framework that can be used to generate / update the machine learning model is also the same as those described in the first embodiment, and therefore will not be described again.

[0063] 10 is a diagram showing an outline of the functional configuration of an information processing device 1b according to this embodiment. The information processing device 1b functions as an information processing device including a fact attribute determination unit 21, an estimated attribute determination unit 22, a learning unit 25, a score estimation unit 24, and a segmentation unit 26, by a program recorded in a storage device 14 being read into a RAM 13 and executed by a CPU 11, which controls each piece of hardware included in the information processing device 1b. Note that in this embodiment and other embodiments described below, each function included in the information processing device 1b 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.

[0064] The fact attribute determination unit 21, the estimated attribute determination unit 22, and the score estimation unit 24 are the same as those described in the first embodiment, and therefore will not be described again. On the other hand, the processing content of the learning unit 25 differs from that of the learning unit 23 described in the first embodiment, and the segmentation unit 26 is a functional unit that does not appear in the first embodiment, so the learning unit 25 and the segmentation unit 26 will be described below.

[0065] The segmentation unit 26 segments users who use a first service (in this embodiment, an online shopping service) to determine users included in the first segment and users included in the second segment. Here, the segmentation unit 26 segments users who use the first service based on user attribute data obtained from the first service.

[0066] Although the type of attribute data referenced for segmentation is not limited, in this embodiment, an example will be described in which the frequency of use of the first service for each user over a predetermined period is used as the attribute data referenced for segmentation. In this embodiment, the segmentation unit 26 acquires and references the frequency of use of the first service over a predetermined period (e.g., the number of times) from the first service, and places users whose frequency of use is equal to or greater than a predetermined threshold into a first segment composed of users who trade frequently (a segment composed of heavy users who trade frequently), and places users whose frequency of use is less than the predetermined threshold into a second segment composed of users who trade infrequently (a segment composed of light users who trade infrequently).

[0067] However, as described above, the type of attribute data referenced for segmentation is not limited. For example, the attribute data referenced for segmentation may be attribute data related to service usage history or transaction history. Furthermore, for example, the attribute data referenced for segmentation may refer to the login frequency of each user to the first service over a predetermined period. When the login frequency of each user to the first service is referenced as the attribute data, the segmentation unit 26 acquires and references the login frequency (e.g., number of times) to the first service over a predetermined period from the first service, and places users whose login frequency is equal to or greater than a predetermined threshold in a first segment (a segment composed of active users), and users whose login frequency is less than the predetermined threshold in a second segment (a segment composed of inactive users). Furthermore, the segmentation unit 26 may perform heuristic segmentation, for example, for users who have a service usage history, to uniquely determine a user segment based on rules according to attribute data including actual attribute data and estimated attribute data.

[0068] The segmentation unit 26 may also encode attribute data of each user with a service usage history into an embedded representation (vector representation) and identify similar / close users as segments based on the Euclidean distance or cosine similarity between the embedded representations. In other words, the segmentation unit 26 treats users who are close and similar to each other in some feature space (embedded space, latent space) as users included in one segment. The encoding method may acquire the embedded representation using a trained model (a pre-trained model based on a neural network or BERT, etc.), or may acquire the embedded representation by embedding each user as a node in a graph network. The learning unit 25, described later, may perform training using a dataset of users belonging to a certain segment segmented in this way, and then preferentially select segments that are close to (or highly similar to) the segment, and retrain using a dataset of users belonging to the selected segment.

[0069] The learning unit 25 generates and / or updates a score estimation model by the score estimation unit 24. The score estimation model is a machine learning model that, when one or more attribute data (attribute data group) related to a target user are input, outputs a credit score that indicates the degree of risk that the target user will not correctly settle the deferred payment if the target user uses deferred payment. This embodiment differs from the first embodiment, which uses attribute data and a score for the same service, in that the user's attribute data included in the training data is attribute data obtained in a first service (in this embodiment, an online shopping service) and the score is the same user's credit score in a second service (in this embodiment, a deferred payment service) that is different from the first service.

[0070] In this embodiment, the learning unit 25 includes at least a first learning unit 25a and a second learning unit 25b.

[0071] The first learning unit 25a trains a first model (first score estimation model) with first training data including a combination of a first attribute data group related to users included in a first segment among users who use the first service and a label corresponding to a score (the user's credit score in the second service) that changes depending on the degree of risk the provider of the second service will bear if the user uses a second service different from the first service. The second service is a service different from the first service. As described above, in this embodiment, the first service is an online shopping service and the second service is a deferred payment service. The risk the provider of the second service will bear is a deferred payment risk, which is expressed as some kind of index (in this embodiment, a label that changes depending on the magnitude of the deferred payment risk) of the risk that the user will not properly settle the deferred payment if the user uses the deferred payment service, determined based on the payment history in the deferred payment service.

[0072] The second learning unit 25b trains a second model (second score estimation model) having as fixed parameters part of the configuration of the first model trained by the first learning unit 25a (parameters such as intermediate layers, nodes, and other weights) using second training data including a combination of a second group of attribute data relating to users who are included in a second segment among users who use the first service and a label corresponding to a score that changes depending on the degree of risk the provider of the second service assumes when the user uses the second service.

[0073] The specific framework for training the second teacher data into a second model (second score estimation model) that has some of the parameters of the configuration of the first model trained by the first learning unit 25a as fixed parameters is the same as that described in the first embodiment with reference to Figures 3 to 5, and therefore will not be described here.

[0074] <<Processing Flow>> Next, a processing flow executed by the information processing device according to this embodiment will be described. Note that the specific content and processing order of the processing described below are an example for implementing the present disclosure. The specific processing content and processing order may be selected as appropriate depending on the embodiment of the present disclosure.

[0075] 11 is a flowchart showing the flow of machine learning processing according to this embodiment. The processing shown in this flowchart is executed periodically or at a timing designated by an administrator.

[0076] In step S300, a user group is segmented. The segmentation unit 26 segments users who use the first service based on attribute data of the users who use the first service, thereby determining users included in the first segment and users included in the second segment. Then, the process proceeds to step S301.

[0077] In steps S301 and S302, a first score estimation model is generated. The first learning unit 25a generates first training data including a combination of a first attribute data group related to users who are included in a first segment among users who use the first service and a label corresponding to a score that changes depending on the degree of risk assumed by the provider of a second service different from the first service when the user uses the second service (step S301). Then, the first learning unit 25a performs a training process for the first score estimation model based on the generated first training data (step S302). The process then proceeds to step S303.

[0078] In steps S303 and S304, a second score estimation model is generated. The second learning unit 25b generates second training data including a combination of a second attribute data group related to users who are included in a second segment among users who use the first service and a label corresponding to a score that changes depending on the degree of risk the provider of the second service assumes when the user uses the second service (step S303). The second learning unit 25b then uses the second training data to re-train a second score estimation model that has some of the parameters of the first score estimation model generated in step S302 as fixed parameters, thereby constructing a second score estimation model to be used for score estimation by the score estimation unit 24 (step S304). The processing shown in this flowchart then ends.

[0079] The flow of the score estimation process according to this embodiment is roughly the same as the flow of the process described with reference to FIG. 8 in the first embodiment, and therefore a description thereof will be omitted.

[0080] <<Variations>> In the second embodiment described above, an example was described in which users using the first service were divided into two segments and two-stage learning was performed. However, the number of segments to be divided and the number of stages of learning are not limited to the example described above. As in the first embodiment, the learning unit 25 may further include a third learning unit, a fourth learning unit, ..., and an nth (n > 2) learning unit. Here, the nth learning unit trains an nth model (nth score estimation model) having a configuration of fixed parameters, etc., including some of the parameters of the n-1th model trained by the n-1th learning unit, with nth training data including a combination of an nth attribute data group related to users included in the nth segment among users using the first service and a label corresponding to a score that changes depending on the degree of risk the provider of the second service assumes if the user uses a second service different from the first service. In this case, the score estimation unit 24 may estimate the score of the target user by inputting a group of attribute data including fact attributes and estimated attributes related to the target user into the nth model trained by the nth learning unit. For example, as a variation of the embodiment described above, the segments may be divided into three segments, "heavy users," "middle users," and "light users," and three stages of learning may be performed.

[0081] Also in this embodiment, the nth learning unit may determine two or more models from a model group including the (n-1)th model and the (n-2)th model, and may train the nth teacher data on the nth model having some parameters of each of the determined models as fixed parameters. In this case, the nth learning unit may determine two or more models including at least the (n-1)th model from the model group.

[0082] <<Effects>> For users belonging to a certain segment (e.g., light users of a first service), the accuracy of score estimation may be insufficient because the quantity or quality of attribute data is insufficient (e.g., a relatively small amount, missing attribute data, etc.). Therefore, in this embodiment, a first score estimation model is created using attribute data of users in a first segment who have a sufficient quantity or quality of attribute data (e.g., heavy users of the first service), and a second score estimation model is created by fixing a portion of the first score estimation model and then re-learning it using attribute data of users in a second segment who are insufficient in quantity or quality (e.g., light users of the first service). This makes it possible to improve the accuracy of estimating individual creditworthiness even for users in a segment with a relatively small amount of training dataset (here, light users).

[0083] 1, 1b Information processing device

Claims

1. a first learning means for training a first model to learn first teacher data including a combination of a first attribute data group related to a user included in a first segment among users who use a first service, and a label corresponding to a score that changes according to the degree of risk assumed by a provider of a second service different from the first service when the user uses the second service; a second learning means for learning second teacher data including a combination of a second attribute data group related to a user who is included in a second segment different from the first segment among users who use the first service and a label corresponding to a score that changes according to the degree of risk assumed by a provider of the second service when the user uses the second service, into a second model having some parameters of the configuration of the learned first model as fixed parameters; A trust learning device comprising:

2. The system further includes a segmentation means for determining users included in the first segment and users included in the second segment by performing segmentation of users who use the first service. The trust learning device according to claim 1 .

3. The segmentation means performs segmentation of users who use the first service based on attribute data of the users who use the first service. The trust learning device according to claim 2.

4. The segmentation means segments users who use the first service based on a frequency of use of the first service in a predetermined period.

4. The trust learning device according to claim 3.

5. The second service is a deferred payment service, The risk borne by the provider of the second service is a deferred payment risk that the settlement of the deferred payment will not be properly fulfilled by the user when the user uses the deferred payment service, the deferred payment risk being determined based on the payment history in the deferred payment service. The trust learning device according to claim 1 .

6. The second service is a service for which a correlation is confirmed or inferred between a risk borne by a provider of the first service when the user uses the first service and a risk borne by a provider of the second service when the user uses the second service. The trust learning device according to claim 1 .

7. the first model and the second model are gradient boosting decision trees; the second learning means causes the second model, which has parameters of some of the nodes of the first model as fixed parameters, to learn the second teacher data; The trust learning device according to claim 1 .

8. the first model and the second model are neural networks; The second learning means causes the second model having some parameters of an intermediate layer of the trained first model as fixed parameters to learn the second teacher data. The trust learning device according to claim 1 .

9. the second learning means causes the second model, in which a plurality of copies of the trained first model are arranged in parallel, to learn the second teacher data, the second model having some parameters of an intermediate layer included in the second model as fixed parameters; The trust learning device according to claim 8.

10. a fact attribute determining means for determining a fact attribute that can be confirmed as a fact about the user based on user-provided data provided by the user himself or history data of the user; and determining an estimated attribute of the user based on at least the factual attribute of the user. The trust learning device according to claim 1 .

11. On the computer, a first learning step of training a first model to learn first teacher data including a combination of a first attribute data group related to users who are included in a first segment among users who use a first service, and a label corresponding to a score that changes according to the degree of risk assumed by a provider of a second service different from the first service when the user uses the second service; a second learning step of training a second model having some parameters of the trained first model as fixed parameters to learn second teacher data including a combination of a second attribute data group related to users who use the first service and are included in a second segment different from the first segment, and a label corresponding to a score that changes according to the degree of risk a provider of the second service assumes when the user uses the second service; A trust learning method that allows the user to execute the following:

12. Computer, a first learning means for training a first model to learn first teacher data including a combination of a first attribute data group related to a user included in a first segment among users who use a first service, and a label corresponding to a score that changes according to the degree of risk assumed by a provider of a second service different from the first service when the user uses the second service; a second learning means for learning second teacher data including a combination of a second attribute data group related to a user who is included in a second segment different from the first segment among users who use the first service and a label corresponding to a score that changes according to the degree of risk assumed by a provider of the second service when the user uses the second service, into a second model having some parameters of the configuration of the learned first model as fixed parameters; A program that functions as a

13. A credit estimation device comprising an estimation means for estimating the second score to be set for a target user by inputting a group of attribute data relating to the target user into the second model that has been trained by the learning method described in claim 11.

14. On the computer, A credit estimation method, comprising: executing an estimation step of estimating the second score to be set for a target user by inputting a group of attribute data relating to a target user into the second model that has been trained by the learning method described in claim 11.

15. Computer, A program that functions as an estimation means for estimating the second score to be set for a target user by inputting a group of attribute data relating to a target user into the second model that has been trained by the learning method described in claim 11.