Examination support system and examination support method

Abstract

To output a fair examination result for a specific attribute in an examination for a person.SOLUTION: An examination support system holds a plurality of attributes including a required fairness attribute and a fairness determination attribute of a past target person, a score indicating the examination result of the past target person, a learning model for using the plurality of attributes as inputs, and outputting the score, and the plurality of attributes of a new target person, inputs the plurality of attributes of a target person for comparison having a value of the fairness determination attribute similar to that of the new target person, to the learning model, acquires a first score, inputs a changed target person obtained by changing the value of the required fairness attribute of the new target person, and the plurality of attributes thereof to the learning model, acquires a second score, and outputs a score to the new target person, on the basis of a position of the second score in the distribution of the first score.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to an examination support system and an examination support method. [Background technology]

[0002] As a background technology of this technical field, there is JP 2021-12593 A (Patent Document 1). This publication states that "a fairness management system that manages a first model created by machine learning, which includes a replacement means that replaces the first input value included in the input with a second input value when a first prediction result, which is a result of a prediction by a first model using an input including a first input value, satisfies a predetermined condition, a storage means that stores a second prediction result, which is a result of a prediction by the first model using the input after the replacement, a receiving means that receives feedback on the second prediction result, and a creation means that creates a second model by performing machine learning using learning data consisting of an input including the first input value and the second prediction result based on the received feedback." (See Abstract). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent Publication No. 2021-12593 Summary of the Invention [Problem to be solved by the invention]

[0004] In order to ensure fairness, the technology described in Patent Document 1 attempts to change the output value of a trained model by rewriting the values ​​of target attributes that require fairness in the input data, such as gender and race. However, even if an attribute is not specified as a target attribute, there are attributes that are highly related to the target attribute.

[0005] For example, suppose that the target attributes for which fairness is required include gender and race, and hobbies, qualifications, alma mater, and residential area are not included in the target attributes but are included in the input attributes of the trained model. In this case, there are many cases where the majority of people with a specific hobby or a specific qualification are of a specific gender, the majority of people who graduated from a specific school are of a specific gender, the majority of people who live in a specific area are of a specific race, etc. In other words, gender, which is included in the target attributes for which fairness is required, is highly related to hobbies, qualifications, and alma mater, which are included in the input attributes but are not included in the target attributes, and race, which is included in the target attributes for which fairness is required, is highly related to residential area, which is included in the input attributes but are not included in the target attributes.

[0006] Therefore, since the trained model in the technology described in Patent Document 1 includes influences from the specific hobby, the specific qualification, the specific school, the specific area of ​​residence, etc., even if the gender or race in the input data is rewritten, there is a risk that bias due to gender or race will remain in the output from the trained model. Therefore, one aspect of the present invention outputs a fair screening result with respect to specific attributes in screening a person. [Means for solving the problem]

[0007] In order to solve the above-mentioned problems, one aspect of the present invention employs the following configuration: A screening support system for supporting screening of persons includes a processor and a memory, and the memory holds log information indicating a plurality of attributes including a fairness requirement attribute and a fairness judgment attribute of past subjects of the screening, and a score indicating a screening result of the screening of the past subjects, a learning model obtained by machine learning on training data in which the plurality of attributes of the past subjects are objective variables and the score of the past subjects is an explanatory variable, input attribute information indicating the plurality of attributes of a new subject of the screening, and fairness judgment attribute information indicating a condition under which the value of the fairness judgment attribute is similar, and the processor The device refers to the conditions indicated by the fairness judgment attribute information, obtains from the log information the multiple attributes of a comparison subject, which is a past subject whose value of the fairness judgment attribute is similar to that of the new subject, inputs the multiple attributes of the comparison subject into the learning model to obtain a first score, inputs the multiple attributes of the new subject and a changed subject in which the value of the fairness-required attribute of the new subject has been changed into the learning model to obtain a second score, and outputs a score indicating the evaluation result for the new subject based on the position of the second score in the distribution of the first scores. Effect of the Invention

[0008] According to one aspect of the present invention, it is possible to output fair screening results regarding specific attributes in screening a person.

[0009] Problems, configurations and effects other than those described above will become apparent from the following description of the embodiments. [Brief description of the drawings]

[0010] [Figure 1] 1 is a block diagram showing a configuration example of a loan examination support system according to a first embodiment. [Diagram 2] FIG. 4 is an explanatory diagram illustrating an example of an overview of a process performed by a learning unit in the first embodiment. [Diagram 3] 2 is an explanatory diagram showing the relationship between the functional units included in the loan examination support information inference unit in the first embodiment. FIG. [Figure 4] 1 is an explanatory diagram showing an example of an overview of processing by functional units other than the overall processing unit included in the loan examination support information inference unit in the first embodiment. FIG. [Diagram 5] FIG. 4 is a diagram illustrating an example of a data configuration of input log information in the first embodiment. [Figure 6] FIG. 11 is a diagram illustrating an example of a data configuration of determination result log information in the first embodiment. [Figure 7] FIG. 13 is a diagram showing an example of the data configuration of repayment result log information in the first embodiment. [Figure 8] 4 is a diagram illustrating an example of a data configuration of fairness-requiring attribute information in the first embodiment. FIG. [Figure 9] 10 is a diagram illustrating an example of a data configuration of fairness determination attribute information in the first embodiment. FIG. [Figure 10] 4 is a diagram illustrating an example of a data configuration of input attribute information in the first embodiment. FIG. [Figure 11] 10 is a diagram illustrating an example of a data configuration of fairness-requiring attribute value combination information in the first embodiment. FIG. [Figure 12] 13 is a diagram illustrating an example of a data configuration of changed input attribute information in the first embodiment. FIG. [Figure 13] FIG. 11 is a diagram illustrating an example of a data configuration of confirmation target result information in the first embodiment. [Figure 14] FIG. 4 is a diagram showing an example of a data configuration of comparison input information in the first embodiment. [Figure 15] FIG. 11 is a diagram illustrating an example of a data configuration of comparative input correspondence information in the first embodiment. [Figure 16] FIG. 11 is a diagram illustrating an example of a data configuration of comparison result information in the first embodiment. [Figure 17] FIG. 4 is a diagram illustrating an example of a data configuration of comparison average information in the first embodiment. [Figure 18] FIG. 11 is a diagram illustrating an example of a data configuration of correction ratio information in the first embodiment. [Figure 19] FIG. 11 is a diagram illustrating an example of a data configuration of correction candidate information in the first embodiment. [Figure 20] 6 is a diagram illustrating an example of a data configuration of correction output information in the first embodiment. FIG. [Figure 21] 4 is a flowchart illustrating an example of a learning process in the first embodiment. [Figure 22] 1 is a flowchart illustrating an example of an overall process in the first embodiment. [Figure 23] 13 is a flowchart illustrating an example of an attribute input and saving process according to the first embodiment. [Figure 24] FIG. 13 is a diagram showing an example of a screen layout for inputting examination subject information in the first embodiment. [Diagram 25] 13 is a flowchart illustrating an example of a post-change input attribute information generation process in the first embodiment. [Figure 26] 13 is a flowchart showing an example of a comparison input information generation process in the first embodiment. [Figure 27] 13 is a flowchart illustrating an example of a process of inputting changed input attribute information according to the first embodiment. [Figure 28] 13 is a flowchart showing an example of a comparison input information input process in the first embodiment. [Figure 29] 11 is a flowchart illustrating an example of an average value calculation process according to the first embodiment. [Diagram 30] 11 is a flowchart illustrating an example of a ratio calculation process according to the first embodiment. [Diagram 31] 5 is a flowchart illustrating an example of an output value correction process in the first embodiment. [Diagram 32] 11 is a flowchart showing an example of a display process in the first embodiment. [Diagram 33] FIG. 13 is a diagram showing an example of a screen configuration of a loan examination support information display screen in the first embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0011] Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings. In this embodiment, the same components are generally designated by the same reference numerals, and repeated explanations are omitted. Note that this embodiment is merely an example for realizing the present invention, and does not limit the technical scope of the present invention.

[0012] In this embodiment, a system for supporting screening of a person based on the attributes of the person will be described. There is a method for supporting the screening using a machine learning model that outputs a value indicating the screening result for the person when a value indicating the attribute of the person is input.

[0013] Machine learning models that support the screening of such individuals are required to comply with fairness requirements. To comply with the fairness requirements, it is required that screening results that are biased against people who belong to a specific category with respect to a certain type of attribute (fairness-requiring attribute) are not output.

[0014] In this embodiment, a person's attributes are classified into three categories: fairness-requiring attributes, fairness judgment attributes, and reference attributes. Fairness-requiring attributes are attributes that are required to satisfy fairness requirements, i.e., attributes that are required not to affect the review results. Fairness judgment attributes are attributes that affect the review results. Reference attributes are attributes that do not fall under either fairness-requiring attributes or fairness judgment attributes, and may be attributes that affect the review results or attributes that do not affect the review results. Therefore, fairness-requiring attributes, fairness judgment attributes, and reference attributes do not overlap with each other.

[0015] It is expected that the same (similar) screening results will be obtained for any group of people with the same (or similar) value of the fairness judgment attribute. In addition, if there are multiple groups of people with the same (or similar) value of the fairness judgment attribute, and the values ​​of the fairness judgment attribute are different (or dissimilar) among the multiple groups, it is expected that different (diverging) screening results will be obtained among the multiple groups.

[0016] However, in order to comply with the fairness requirement, it is required that the same (similar) screening results be obtained for any group consisting of individuals with the same (or similar) value of the fairness judgment attribute, regardless of the value of the fairness attribute. Also, in the case of any multiple groups consisting of individuals with the same (or similar) value of the fairness judgment attribute, where the values ​​of the fairness judgment attribute are different (or dissimilar) among the multiple groups, if the same (similar) screening results are obtained within each of the multiple groups regardless of the value of the fairness attribute, it can be said that the fairness requirement is complied with, even if different (diverging) screening results are obtained among the multiple groups.

[0017] Furthermore, it is not necessary to consider equality between groups with different values ​​of the reference attribute. For example, for subgroups included in an arbitrary group consisting of people with the same (or similar) values ​​of the fairness judgment attribute, and for people in the subgroups grouped by the value of the fairness-requiring attribute, the screening results may be different between people with different values ​​of the reference attribute, or the screening results may be different between people with the same value of the reference attribute.

[0018] The fairness-requiring attributes are determined, for example, by social reasons or legal regulations. Therefore, for example, which attributes correspond to the fairness-requiring attributes may change depending on the region or the era. In addition, which attributes correspond to the fairness judgment attributes or the reference attributes may change depending on the type of screening, past screening results, screening policy, region, era, etc.

[0019] The screening support system of this embodiment outputs a score indicating the screening result in order to support a screening that satisfies the fairness requirement based on the Equalized Odds. If the positive rate and false positive rate of the result of an event that occurs based on the action of a subject who has benefited from the approval of the screening are similar regardless of the value of the fairness attribute, it can be said that the fairness requirement based on the Equalized Odds is satisfied.

[0020] For example, in loan screening, a positive score is when a person to whom a loan has been granted defaults (fails to repay), and a negative score is when the person does not default (successfully repays), and the screening support system of this embodiment outputs a score to assist in screening so that the proportion of people who received loans among those who did not default (positive rate) is equal to the proportion of people who received loans among those who defaulted (false positive rate). EXAMPLES

[0021] In this embodiment, a system for supporting loan screening, which is an example of screening of a person, will be described. Fig. 1 is a block diagram showing an example of the configuration of the loan screening support system. The loan screening support system 1 is, for example, a computer having a CPU (Central Processing Unit) 11, a memory 12, an auxiliary storage device 13, and a communication device 14, and is connected to an input device 2 and an output device 3.

[0022] The CPU 11 includes a processor and executes programs stored in the memory 12. The memory 12 includes a ROM (Read Only Memory), which is a non-volatile storage element, and a RAM (Random Access Memory), which is a volatile storage element. The ROM stores immutable programs (e.g., a Basic Input / Output System (BIOS)). The RAM is a high-speed, volatile storage element such as a DRAM (Dynamic Random Access Memory), and temporarily stores programs executed by the CPU 11 and data used when the programs are executed.

[0023] The auxiliary storage device 13 is a large-capacity non-volatile storage device such as a magnetic storage device (HDD (Hard Disk Drive)) or a flash memory (SSD (Solid State Drive)), and stores the programs executed by the CPU 11 and data used when the programs are executed. That is, the programs are read from the auxiliary storage device 13, loaded into the memory 12, and executed by the CPU 11.

[0024] The communication device 14 is a network interface device that controls communication with other devices in accordance with a predetermined protocol, and includes a serial interface such as a Universal Serial Bus (USB).

[0025] The input device 2 is a device such as a keyboard or a mouse that receives input from an operator. The output device 3 is a device such as a display device or a printer that outputs the results of program execution in a format that can be viewed by the operator.

[0026] Some or all of the programs executed by the CPU 11 may be provided to the loan examination support system 1 from a removable medium (CD-ROM, flash memory, etc.) which is a non-transitory storage medium, or from an external computer equipped with a non-transitory storage device via a network, and stored in a non-volatile auxiliary storage device 13 which is a non-transitory storage medium. For this reason, the loan examination support system 1 should preferably have an interface for reading data from removable media.

[0027] The loan screening support system 1 is a computer system configured on a single physical computer, or on multiple logically or physically configured computers, and may operate in separate threads on the same computer, or may operate on a virtual computer constructed on multiple physical computer resources.

[0028] The CPU 11 includes a learning unit 151 and a loan examination support information inference unit 152, both of which are functional units. The learning unit 151 includes, for example, a learning processing unit 161, which is also a functional unit, and generates the learning model 103.

[0029] The loan examination support information inference unit 152 generates information for supporting loan examination for the examinee. The loan examination support information inference unit 152 includes, for example, an overall processing unit 171, an attribute input saving processing unit 172, a changed input attribute information generating processing unit 173, a comparison input information generating processing unit 174, a changed input attribute information input processing unit 175, a comparison input information input processing unit 176, an average value calculation processing unit 177, a ratio calculation processing unit 178, an output value correction processing unit 179, and a display processing unit 180, all of which are functional units.

[0030] The processes executed by the above-mentioned functional units will be described in detail below with reference to FIGS. 2 to 4 and FIGS.

[0031] For example, CPU 11 functions as learning unit 151 by operating according to a learning program loaded into memory 12, and functions as loan examination support information inference unit 152 by operating according to a loan examination support information inference program loaded into memory 12. The relationship between the programs and CPU 11 is similar for the functional units included in learning unit 151 and loan examination support information inference unit 152.

[0032] In addition, some or all of the processes performed by the functional units included in the loan examination support system 1 may be realized by hardware such as an application specific integrated circuit (ASIC) or a field programmable gate array (FPGA).

[0033] The auxiliary storage device 13 includes, for example, a log information DB (DataBase) 101, a setting information DB 102, and a learning model 103. The log information DB 101 stores log information indicating the results of subjects (hereinafter also referred to as past subjects) who were subject to loan screening in the past, and specifically stores, for example, input log information 111, judgment result log information 112, and repayment result log information 113.

[0034] The setting information DB 102 stores setting information related to fairness attributes and fairness attributes, and specifically stores, for example, fairness attribute information 121 and fairness judgment attribute information 122. The learning model 103 is a model that outputs the loanable amount and repayment possibility, which are scores indicating the screening results, when the attributes of the screening subject are input.

[0035] In addition, the auxiliary storage device 13 stores, for example, input attribute information 131, fairness-required attribute value combination information 132, changed input attribute information 133, confirmation target result information 134, comparison input information 135, comparison input correspondence information 136, comparison result information 137, comparison average information 138, correction ratio information 139, correction candidate information 140, and correction output information 141.

[0036] The above-mentioned various types of information will be described in detail later with reference to FIGS.

[0037] It should be noted that some or all of the information stored in the auxiliary storage device 13 in FIG. 1 may be stored in the memory 12, or in an external database connected to the loan examination support system 1.

[0038] In this embodiment, the information used by the loan examination support system 1 does not depend on the data structure and may be expressed in any data structure. In this embodiment, the information is expressed in a table format, but the information can be stored in a data structure appropriately selected from, for example, a list, a database, or a queue.

[0039] 2 is an explanatory diagram showing an example of an outline of processing by the learning unit 151. A learning processing unit 161 of the learning unit 151 is started by an input to the input device 2 from a system administrator, which is an example of a user. The learning processing unit 161 acquires values ​​of feature amounts, which are objective variables, from input log information 111 included in the log information DB 101, acquires values ​​of feature amounts, which are explanatory variables, from judgment result log information 112 included in the log information DB 101, and generates a learning model 103 based on training data generated from the acquired values ​​of the objective variable and explanatory variables and a predetermined machine learning algorithm.

[0040] 3 is an explanatory diagram showing the relationship between the functional units included in the loan examination support information inference unit 152. The processes by the attribute input save processing unit 172, the changed input attribute information generation processing unit 173, the comparison input information generation processing unit 174, the changed input attribute information input processing unit 175, the comparison input information input processing unit 176, the average value calculation processing unit 177, the ratio calculation processing unit 178, the output value correction processing unit 179, and the display processing unit 180 are called in order by the overall processing unit 171, for example.

[0041] Fig. 4 is an explanatory diagram showing an example of an overview of processing by functional units excluding the overall processing unit 171 included in the loan examination support information inference unit 152. In Fig. 4, information that is the input and output of each functional unit and relationships between each functional unit are indicated by arrows.

[0042] The attribute input saving processing unit 172 displays on the output device 3 a screen for accepting input of attribute information of a subject for whom a new screening is to be carried out (hereinafter also referred to as a new subject), acquires the attribute information of the new subject in accordance with input from the user (loan officer) to the input device 2, and saves the acquired input attribute information 131.

[0043] The changed input attribute information generation processing unit 173 generates fairness-requiring attribute value combination information 132 indicating a combination of fairness-requiring attribute values ​​with reference to the fairness-requiring attribute information 121. The changed input attribute information generation processing unit 173 also generates changed input attribute information 133 indicating information to which input information other than the fairness-requiring attributes of the new subject is added for each combination of the fairness-requiring attribute values, that is, attribute information of the new subject and a virtual screening subject (also referred to as a changed subject) in which only the fairness-requiring attributes of the new subject have been changed.

[0044] The comparison input information generation processing unit 174 extracts from the input log information 111 past subjects whose values ​​of the fairness judgment attribute indicated by the fairness judgment attribute information 122 are similar to any of the subjects indicated by the changed input attribute information 133, and stores the attribute information of the extracted past subjects in the comparison input information 135. In addition, the comparison input information generation processing unit 174 generates comparison input correspondence information 136 indicating the correspondence between the IDs of the past subjects in the changed input attribute information 133 and the IDs of the past subjects in the input log information 111 and the repayment result log information 113.

[0045] The changed input attribute information input processing unit 175 inputs the attribute information of the changed input attribute information 133 into the learning model 103 to obtain an inference result, and stores the obtained inference result in the confirmation target result information 134. The comparison input information input processing unit 176 inputs the attribute information of the comparison input information 135 into the learning model 103 to obtain an inference result, and stores the obtained inference result in the comparison result information 137.

[0046] The average value calculation processing unit 177 calculates the average value of the inference results indicated by the comparison result information 137 for each combination of the repayment results of past subjects indicated by the repayment result log information 113 and the combination of values ​​of the fairness-requiring attribute indicated by the fairness-requiring attribute value combination information 132, and stores this in the comparison average information 138. The average value calculation processing unit 177 specifies the correspondence relationship between the past subjects corresponding to the repayment results indicated by the repayment result log information 113 and the past subjects corresponding to the inference results indicated by the comparison result information 137 from the comparison input correspondence information 136.

[0047] The ratio calculation processing unit 178 calculates a ratio for a combination of values ​​of the fairness-required attributes of a new subject based on a comparison between the inference result indicated by the confirmation target result information 134 and the average value for each repayment result indicated by the comparison average information 138, and stores the ratio in the corrected ratio information 139.

[0048] The output value correction processing unit 179 corrects the inference result corresponding to a combination different from the combination of values ​​of the fairness-requiring attribute of the past subject indicated by the comparison average information 138, based on the inference result indicated by the confirmation target result information 134 and the correction ratio information 139, and stores the corrected inference result in the correction candidate information 140. The output value correction processing unit 179 selects an inference result to be output from the inference result indicated by the correction candidate information 140 and the inference result indicated by the confirmation target result information 134, and stores the selected inference result in the corrected output information 141.

[0049] The display processing unit 180 displays the inference result indicated by the corrected output information 141 on the output device 3. The user (loan officer) performs loan screening based on the inference result displayed on the output device 3.

[0050] 5 to 20, examples of data configurations of various information held by the loan examination support system 1 will be described below. In the examples of data configurations of various information described below, unless otherwise specified, ID refers to the identifier of the record of the relevant type of information.

[0051] 5 is a diagram showing an example of the data configuration of the input log information 111. The input log information 111 shows past performance values ​​of attributes of a subject. In the example of the input log information 111 in FIG. 5, age, place of employment, annual income, hobbies, licenses, gender, and race are all examples of attributes. The attribute values ​​stored in the input log information 111 are used as values ​​of objective variables in machine learning training data for generating the learning model 103.

[0052] At least a fairness judgment attribute and a fairness requirement attribute are stored in the input log information 111, and a reference attribute may also be stored. That is, at least a fairness judgment attribute and a fairness requirement attribute are used as objective variables (i.e., variables input to the learning model 103) in the machine learning training data for generating the learning model 103, and a reference attribute may also be used. The IDs in the input log information 111 in Figure 5, the judgment result log information 112 in Figure 6 described later, and the repayment result log information 113 in Figure 7 described later indicate IDs that identify past recipients.

[0053] Fig. 6 is a diagram showing an example of the data configuration of the judgment result log information 112. The judgment result log information 112 shows actual scores indicating the screening results for past subjects. The loanable amount and repayment possibility stored in the judgment result log information 112 in Fig. 6 are both examples of scores indicating the screening results. The loanable amount and repayment possibility values ​​stored in the judgment result log information 112 are used as explanatory variable values ​​in machine learning training data for generating the learning model 103.

[0054] 7 is a diagram showing an example of the data configuration of the repayment result log information 113. The repayment result log information 113 shows the past repayment results of the target. The repayment result is an example of the result of an event that occurs based on the action taken by the target who has enjoyed the benefit of having his / her application approved.

[0055] In the example of repayment result log information 113 in Figure 7, a past subject with a result of "Success" indicates that they succeeded in repaying the loan amount that was executed, and a past subject with a result of "Failure" indicates that they failed to repay the loan amount that was executed.

[0056] Fig. 8 is a diagram showing an example of the data configuration of the fairness-requiring attribute information 121. The fairness-requiring attribute information 121 indicates attributes that correspond to fairness-requiring attributes and values ​​that each of the fairness-requiring attributes can take. The example of the fairness-requiring attribute information 121 in Fig. 8 indicates that "gender" and "race" correspond to fairness-requiring attributes, and that the possible value for "gender" is either "male" or "female," and the possible value for "race" is either "A" or "B."

[0057] 9 is a diagram showing an example of the data configuration of the fairness determination attribute information 122. The fairness determination attribute information 122 indicates attributes corresponding to the fairness determination attributes and a range (excerpt range) that is treated as being similar for each of the fairness determination attributes in the comparison input information generation process described later.

[0058] 9 indicates that "age", "place of employment", and "annual income" correspond to the fairness judgment attribute information 122. The example of the fairness judgment attribute information 122 in FIG. 9 also indicates that in the comparison input information generation process described later, information on past subjects who belong to the same "age" range in 5-year increments, the same "industry", and the same "annual income" range in 1 million yen increments as the subjects (new subjects and changed subjects) indicated by the changed input attribute information 133 is extracted from the input log information 111 as past subjects whose fairness judgment attribute values ​​are similar to those of the new subjects. In other words, the extraction range of the fairness judgment attribute information 122 indicates the condition for determining that the corresponding fairness judgment attributes are similar.

[0059] Note that the fairness-requiring attribute information 121 and the fairness-judgment attribute information 122 do not include the same attributes, i.e., the fairness-requiring attributes and the fairness-judgment attributes do not overlap. This is because the same level of screening results should be expected for groups that differ only in the value of the fairness-requiring attribute, while different levels of screening results should be expected for groups that differ in the value of the fairness-judgment attribute.

[0060] Additionally, attributes that are not included in the fairness-requiring attribute information 121 or the fairness judgment attribute information 122, i.e., attributes that are neither fairness-requiring attributes nor fairness judgment attributes, are reference attributes. Therefore, attributes "hobby" and "license" that are included in the example of the input log information 111 in Fig. 5 and / or the example of the input attribute information 131 in Fig. 10 described later, and that are not included in the example of the fairness-requiring attribute information 121 in Fig. 8 or the example of the fairness judgment attribute information 122 in Fig. 9, are reference attributes.

[0061] 10 is a diagram showing an example of the data configuration of the input attribute information 131. The input attribute information 131 indicates the values ​​of the attributes of a new subject. The input attribute information 131 includes values ​​of attributes that are explanatory variables input to the learning model 103. In other words, the input attribute information 131 may include values ​​of the same attributes as the input log information 111.

[0062] 11 is a diagram showing an example of a data configuration of the fairness-requiring attribute value combination information 132. The fairness-requiring attribute value combination information 132 indicates all combinations of values ​​of the fairness-requiring attributes defined in the fairness-requiring attribute information 121.

[0063] In the example of fairness-requiring attribute information 121 in Figure 8, the possible values ​​for the fairness-requiring attribute "gender" are either "male" or "female", and the possible values ​​for the fairness-requiring attribute "race" are either "A" or "B", so in the example of fairness-requiring attribute value combination information 132 in Figure 11, (female, A), (male, A), (female, B), and (male, B) are stored as combinations of fairness-requiring attribute values ​​(gender, race).

[0064] The first record of the fairness-requiring attribute value combination information 132 stores a combination of fairness-requiring attribute values ​​indicated by the input log information 111 (the same applies to the changed input attribute information 133 in FIG. 12 described later).

[0065] 12 is a diagram showing an example of the data configuration of the changed input attribute information 133. The changed input attribute information 133 is information in which attributes other than the fairness-requiring attributes of the input attribute information 131 are merged into each record of the fairness-requiring attribute value combination information 132.

[0066] 13 is a diagram showing an example of the data configuration of the confirmation target result information 134. The confirmation target result information 134 indicates, for each record of the changed input attribute information 133, the values ​​of the explanatory variables (loanable amount and repayment possibility) obtained by inputting the attribute values ​​indicated by the record into the learning model 103. Note that, for each ID of the changed input attribute information 133, the values ​​of the explanatory variables obtained by inputting the objective variable corresponding to the ID into the learning model 103 are stored in a record having the same ID in the confirmation target result information 134. The loanable amount and repayment possibility stored in the confirmation target result information 134 are examples of a second score.

[0067] 14 is a diagram showing an example of the data configuration of comparison input information 135. Comparison input information 135 indicates the values ​​of each attribute (indicated by input log information 111) of past subjects whose values ​​of all fairness judgment attributes indicated by fairness judgment attribute information 122 are in the same excerpt range as subjects (new subjects and changed subjects) indicated by changed input attribute information 133. The subjects indicated by comparison input information 135 are an example of subjects for comparison.

[0068] Fig. 15 is a diagram showing an example of a data configuration of the comparison input correspondence information 136. The comparison input correspondence information 136 shows a correspondence relationship between the ID (log ID in Fig. 15) of the input log information 111 from which the comparison input information 135 is extracted, and the ID (ID in Fig. 15) of the comparison input information 135. That is, for example, the example of the comparison input correspondence information 136 in Fig. 15 shows that records with IDs of "1", "2", "3", "4", "5", and "6" in the input log information 111 from which the comparison input information 135 is extracted have been extracted as records with IDs of "1", "2", "3", "4", "5", and "6" in the comparison input information 135, respectively.

[0069] 16 is a diagram showing an example of the data configuration of the comparison result information 137. The comparison result information 137 indicates, for each record of the comparison input information 135, the values ​​of the explanatory variables (loanable amount and repayment possibility) obtained by inputting the attribute values ​​indicated by the record into the learning model 103. Note that, for each ID of the comparison input information 135, the values ​​of the explanatory variables obtained by inputting the objective variable corresponding to the ID into the learning model 103 are stored in a record having the same ID in the comparison result information 137. The loanable amount and repayment possibility stored in the comparison result information 137 are examples of a first score.

[0070] 17 is a diagram showing an example of the data configuration of the comparison average information 138. The comparison average information 138 indicates the average loanable amount and loan repayment amount for each repayment result indicated by the repayment result log information 113 for past subjects indicated by the comparison input information 135 corresponding to each combination of values ​​of the fairness-requiring attribute indicated by the fairness-requiring attribute value combination information 132. The average loanable amount when successful, average repayment possibility when successful, average loanable amount when unsuccessful, and average repayment possibility when unsuccessful stored in the comparison average information 138 are examples of the distribution of the first score.

[0071] For example, the top record in the example of the comparison average information 138 in Fig. 17 indicates that the average loanable amount (average loanable amount in the case of success) and the average repayment possibility (average repayment possibility in the case of success) of past subjects included in the comparison input information 135, who have the gender of "female", the race of "A", and the repayment result of "success" are "500,000" and "50%, respectively. In addition, the top record in the example of the comparison average information 138 in Fig. 17 further indicates that the average loanable amount (average loanable amount in the case of failure) and the average repayment possibility (average repayment possibility in the case of failure) of past subjects included in the comparison input information 135, who have the gender of "female", the race of "A", and the repayment result of "failure" are "100,000" and "20%, respectively.

[0072] The combination of fairness-requiring attribute values ​​indicated by the record in the first row of the comparison average information 138 matches the combination of fairness-requiring attribute values ​​indicated by the input log information 111 .

[0073] 18 is a diagram showing an example of the data configuration of the correction ratio information 139. The correction ratio information 139 indicates the difference between the loanable amount and repayment possibility for each repayment result indicated by the record in the first row of the comparison average information 138 (i.e., the record corresponding to the combination of values ​​of the fairness-requiring attributes indicated by the input log information 111) and the loanable amount and repayment possibility indicated by the record in the first row of the confirmation target result information 134 (i.e., the record corresponding to the combination of values ​​of the fairness-requiring attributes indicated by the input log information 111).

[0074] Furthermore, the correction ratio information 139 indicates the ratio of the difference for each repayment result. In other words, the ratio stored in the correction ratio information 139 indicates the position of the inference result (second score) by the learning model 103 for the new target in the distribution of first scores of past targets who have the same combination of values ​​of the fairness-requiring attribute as the new target and who have similar values ​​of the fairness judgment attribute.

[0075] 19 is a diagram showing an example of the data configuration of correction candidate information 140. Correction candidate information 140 indicates correction candidates (corrected available loan amount and corrected repayment possibility) of the screening results of a new target indicated by input log information 111. Specifically, correction candidate information 140 indicates the corrected available loan amount and corrected repayment possibility based on the average available loan amount and average repayment possibility for each repayment result indicated by each record from the second row onwards of comparison average information 138 and the ratio indicated by correction ratio information 139. The corrected available loan amount and corrected repayment possibility indicated by correction candidate information 140 are examples of correction scores.

[0076] 20 is a diagram showing an example of the data configuration of the corrected output information 141. The corrected output information 141 is information indicating the screening results for a new subject indicated by the input log information 111. The corrected output information 141 indicates, as the screening result, any record in the confirmation target result information 134 or any record in the correction candidate information 140. Furthermore, the corrected output information 141 indicates whether the screening result it indicates is a corrected screening result or not.

[0077] An example of processing by each processing unit will be described below. Before each processing unit starts processing, it is assumed that the column names of various information generated by the processing unit are set in advance.

[0078] 21 is a flowchart showing an example of the learning process by the learning processing unit 161. Before the learning process starts, values ​​of the input log information 111 and the judgment result log information 112 are set in advance. When a predetermined input from a user to the input device 2 is accepted, the process by the learning processing unit 161 starts (S1600). The learning processing unit 161 reads all records of the input log information 111 and the judgment result log information 112 (S1601).

[0079] The learning processing unit 161 associates the values ​​of each attribute indicated by the input log information 111 with the loanable amount and repayment possibility in the judgment result log information 112 for each record having the same ID, executes machine learning using a predetermined algorithm on training data with each attribute as an explanatory variable and the loanable amount and repayment possibility as objective variables, outputs the learning model 103 (S1602), and ends the learning process. Note that in this embodiment, in order to make it possible to confirm and record whether the fairness requirements are met, the learning model 103 is generated in which not only the fairness judgment attribute and the reference attribute but also the fairness required attribute is input.

[0080] 22 is a flowchart showing an example of overall processing by the overall processing unit 171. The overall processing unit 171 calls the attribute input save processing unit 172 and causes the attribute input save processing unit 172 to execute an attribute input save processing described later (S1710). The overall processing unit 171 calls the post-change input attribute information generation processing unit 173 and causes the post-change input attribute information generation processing unit 173 to execute a post-change input attribute information generation processing described later (S1711).

[0081] The overall processing unit 171 calls the comparison input information generation processing unit 174 and has the comparison input information generation processing unit 174 execute the comparison input information generation processing described later (S1712). The overall processing unit 171 calls the changed input attribute information input processing unit 175 and has the changed input attribute information input processing unit execute the changed input attribute information input processing described later (S1713).

[0082] Overall processing unit 171 calls comparison input information input unit 176 and has comparison input information input unit 176 execute a comparison input information input process described later (S1714). Overall processing unit 171 calls average value calculation unit 177 and has average value calculation unit 177 execute an average value calculation process described later (S1715).

[0083] Overall processing unit 171 calls ratio calculation processing unit 178 and has ratio calculation processing unit 178 execute ratio calculation processing, which will be described later (S1716). Overall processing unit 171 calls output value correction processing unit 179 and has output value correction processing unit 179 execute output value correction processing, which will be described later (S1717). Overall processing unit 171 calls display processing unit 180 and has display processing unit 180 execute display processing, which will be described later (S1718), and ends overall processing.

[0084] 23 is a flowchart showing an example of the attribute input save processing by the attribute input save processing unit 172. The attribute input save processing unit 172 displays an examination subject information input screen on the output device 3 (S1720). Details of the examination subject information input screen will be described later with reference to FIG.

[0085] The attribute input save processing unit 172 acquires values ​​of the attributes (age, place of employment, annual income, hobbies, licenses, sex, and race) of the new subject to be registered in the input attribute information 131 by input via the input device 2 (S1721). The attribute input save processing unit 172 accepts an input indicating that the values ​​of the attributes of the new subject to be registered in the input attribute information 131 have been finalized via the input device 2 (S1722).

[0086] The attribute input save processing unit 172 generates and outputs input attribute information 131 based on the values ​​of each attribute of the new subject input in step S1721 and confirmed in step S1722 (S1723), and ends the attribute input save processing.

[0087] In step S1723, the attribute input save processing unit 172 specifically, for example, adds a one-row record to the input attribute information 131, and registers the value "1" in the ID column of the added record, and registers the values ​​of the corresponding attributes of the new subject that were input in step S1721 and confirmed in step S1722 in the columns indicating each attribute of the added record.

[0088] 24 is a diagram showing an example of the screen configuration of the screening target information input screen. The screening target input screen 31 includes, for example, an age input area 311, an employer input area 312, an income area 313, a hobby input area 314, a license input area 315, a gender input area 316, and a race input area 317, all of which are areas for inputting attributes in step S1721. The screening target input screen 31 also includes, for example, a button 318. When the button 318 is selected, the attribute values ​​input in each input area are finalized in step S1722.

[0089] 25 is a flowchart showing an example of the changed input attribute information generation process by the changed input attribute information generation processing unit 173. Before the changed input attribute information generation process starts, the value of the fairness-requiring attribute information 121 is set in advance. The changed input attribute information generation processing unit 173 reads all records (only one row) of the input attribute information 131 and all records of the fairness-requiring attribute information 121 (S1730).

[0090] The changed input attribute information generation processor 173 generates a first row record of the fairness-requiring attribute value combination information 132 (S1731). Specifically, for example, the changed input attribute information generation processor 173 adds a first row record to the fairness-requiring attribute value combination information 132, and registers a value of "1" in the ID column of the added record, and registers the values ​​of the corresponding attributes of the input attribute information 131 read in step S1730 in the columns indicating each fairness-requiring attribute of the added first row record.

[0091] The changed input attribute information generation processor 173 generates records from the second row onwards of the fairness-requiring attribute value combination information 132 (S1732). Specifically, for example, the changed input attribute information generation processor 173 adds records from the second row onwards of the fairness-requiring attribute value combination information 132, registers values ​​from "2" onwards as consecutive numbers in the ID column of each of the added records, and registers all combinations of values ​​of the fairness-requiring attribute indicated by the fairness-requiring attribute information 121 in the column indicating the fairness-requiring attribute of each of the added records without duplication.

[0092] The post-change input attribute information generation processor 173 reads all rows of the fairness-requiring attribute value combination information 132 (S1733).

[0093] The post-change input attribute information generation processing unit 173 executes the processes of steps S1735 to S1737 for each record of the fairness-requiring attribute value combination information 132 read in step S1733 (for records selected one by one starting from the record with the smallest ID), executes a loop process (S1734), and terminates the post-change input attribute information generation process.

[0094] The post-change input attribute information generation processor 173 adds one record to the post-change input attribute information 133, and registers a value in the ID column of the added record so that the value becomes a consecutive number starting from "1" (S1735).

[0095] The changed input attribute information generation processing unit 173 registers the value of the corresponding fairness-requiring attribute indicated by the record of the fairness-requiring attribute value combination information 132 selected in the most recent step S1734 in the column of each fairness-requiring attribute (in the example of Figure 12, gender and race correspond to fairness-requiring attributes) of the record of the changed input attribute information 133 added in the most recent step S1735 (S1736).

[0096] The changed input attribute information generation processing unit 173 registers the value of the corresponding attribute indicated by the record of the input attribute information 131 read in step S1730 in the column of each attribute other than the fairness required attribute of the record of the changed input attribute information 133 added in the most recent step S1735 (S1737).

[0097] By the processing of Figure 25, changed input attribute information 133 is generated, and the changed input attribute information 133 is information in which attributes other than the fairness-requiring attributes of the new target persons indicated by the input attribute information 131 are assigned to all combinations of the values ​​of the fairness-requiring attributes.

[0098] 26 is a flowchart showing an example of the comparison input information generating process by the comparison input information generating processor 174. Before the comparison input information generating process starts, the value of the fairness determination attribute information 122 is set in advance.

[0099] The comparison input information generation processor 174 reads all records (only one row) of the input log information 111, all records of the fairness determination attribute information 122, and all records of the post-change input attribute information 133 (S1740).

[0100] The comparison input information generation processing unit 174 executes a loop process (S1741) of steps S1742 to S1743 for each record of the fairness judgment attribute information 122 read in step S1740 (for records selected one by one starting from the record with the smallest ID).

[0101] The comparison input information generation processor 174 acquires (S1742) the value of the fairness judgment attribute indicated by the record selected in the most recent step S1740 of the fairness judgment attribute information 122 from the changed input attribute information 133. Note that, since the values ​​of all attributes except for the fairness requirement attribute are common to all records of the changed input attribute information 133, the comparison input information generation processor 174 only needs to acquire one of the values ​​of the fairness judgment attribute.

[0102] The comparison input information generation processing unit 174 generates a condition that must be satisfied by the value of the fairness judgment attribute of the past subject registered in the comparison input information 135, based on the value of the fairness judgment attribute obtained in the most recent step S1742 and the excerpt range indicated by the record selected in the most recent step S1741 of the fairness judgment attribute information 122 (S1743).

[0103] Specifically, for example, if the fairness judgment attribute indicated by the record selected in step S1741 is “age” and the value acquired in step S1742 is “24 years old”, the excerpt range corresponding to “age” in the fairness judgment attribute information 122 in FIG. 9 is “every 5 years old”, so in step S1743, the comparison input information generation processing unit 174 generates, for example, “20 to 24 years old” as a condition.

[0104] Also, for example, if the fairness judgment attribute indicated by the record selected in step S1741 is “place of employment” and the value acquired in step S1742 is “X industry”, the excerpt range corresponding to “place of employment” in the fairness judgment attribute information 122 in FIG. 9 is “by industry”, so in step S1743, the comparison input information generation processing unit 174 generates, for example, “X industry” as a condition.

[0105] Also, for example, if the fairness judgment attribute indicated by the record selected in step S1741 is "annual income" and the value acquired in step S1742 is "6.5 million," the excerpt range corresponding to "annual income" in the fairness judgment attribute information 122 in FIG. 9 is "every 1 million yen," so in step S1743, the comparison input information generation processing unit 174 generates a condition, for example, "6 million yen or more and less than 7 million yen."

[0106] For example, for a fairness judgment attribute in which a numerical range is specified as an excerpt range, such as the above-mentioned age or annual income, it is preferable that the minimum value or the maximum value of age is determined in advance. For example, if the minimum value of age is 20 years old, the comparison input information generation processing unit 174 can determine non-overlapping age ranges for "every 5 years" such as 20 years old or older and 24 years old or younger, 25 years old or older and 29 years old or younger, 30 years old or older and 34 years old or younger, ..., and can generate a range including the age value acquired in step S1742 as a condition related to age. Note that these ranges may be determined in advance.

[0107] Next, the comparison input information generation processing unit 174 obtains from the input log information 111 all records that satisfy all the conditions generated in step S1743 (i.e., all records related to past subjects whose fairness judgment attributes are similar to those of the new subject) (S1744).

[0108] The comparison input information generation processing unit 174 executes the processes of steps S1746 to S1748 for each record obtained from the input log information 111 in step S1744 (for example, for records selected one by one starting from the record with the smallest ID), executes a loop process (S1745), and terminates the comparison input information generation process.

[0109] The comparison input information generation processor 174 adds one record to the comparison input information 135, and registers a value in the ID column of the added record so that the value becomes a consecutive number starting from "1" (S1746).

[0110] The comparison input information generation processing unit 174 registers the values ​​of the corresponding attributes indicated by the record of the input log information 111 selected in the most recent step S1745 in the columns of each attribute (in the example of FIG. 14, age, place of employment, annual income, license, gender, and race) of the record of the comparison input information 135 added in the most recent step S1746 (S1747).

[0111] The comparison input information generation processor 174 adds one record to the comparison input correspondence information 136, registers the ID value registered in the most recent step S1746 in the ID column of the added record in the comparison input correspondence information 136, and registers the ID value of the record selected in the most recent step S1745 of the input log information 111 in the log ID column of the added record in the comparison input correspondence information 136 (S1748). This associates the ID in the comparison input information 135 with the ID in the input log information 111.

[0112] 27 is a flowchart showing an example of the post-change input attribute information input process by the post-change input attribute information input processor 175. The post-change input attribute information input processor 175 reads all records of the post-change input attribute information 133 (S1750).

[0113] The changed input attribute information input processing unit 175 executes a loop process (S1751) of executing the processes of steps S1752 to S1753 for each record of the changed input attribute information 133 read in step S1750 (for example, for records selected one by one starting from the record with the smallest ID), and then terminates the changed input attribute information input process.

[0114] The changed input attribute information input processing unit 175 inputs each attribute of the record selected in the most recent step S1751 of the changed input attribute information 133 into the learning model 103, and obtains values ​​of the loanable amount and repayment possibility (S1752).

[0115] The changed input attribute information input processing unit 175 adds one record to the confirmation target result information 134, registers the ID value of the record selected in the most recent step S1751 of the changed input attribute information 133 in the ID column of the added record, and registers the loanable amount and repayment possibility values ​​obtained in the most recent step S1752 in the loanable amount and repayment possibility columns of the added record, respectively (S1753).

[0116] 28 is a flowchart showing an example of the comparison input information input process by the comparison input information input processor 176. The comparison input information input processor 176 reads all records of the comparison input information 135 (S1760).

[0117] The comparison input information input processing unit 176 executes a loop process (S1761) of executing the processes of steps S1762 to S1763 for each record of the comparison input information 135 read in step S1760 (for example, for records selected one by one starting from the record with the smallest ID), and then terminates the comparison input information input process.

[0118] The comparison input information input processing unit 176 inputs each attribute of the record selected in the most recent step S1761 of the comparison input information 135 into the learning model 103, and obtains values ​​of the loanable amount and repayment possibility (S1762).

[0119] The comparison input information input processing unit 176 adds one record to the comparison result information 137, registers the ID value of the record selected in the most recent step S1761 of the comparison input information 135 in the ID column of the added record, and registers the loanable amount and repayment possibility values ​​obtained in the most recent step S1762 in the loanable amount and repayment possibility columns of the added record, respectively (S1763).

[0120] 29 is a flow chart showing an example of average value calculation processing by average value calculation processing section 177. Before the average value calculation processing starts, the value of repayment result log information 113 is preset.

[0121] The average value calculation processing unit 177 reads all records of the input log information 111, the repayment result log information 113, the fairness-required attribute value combination information 132, the comparison input correspondence information 136, and the comparison result information 137 (S1770).

[0122] The average value calculation processing unit 177 executes the processes of steps S1772 to S1779 for each record of the fairness-requiring attribute value combination information 132 read in step S1770 (for example, for records selected one by one starting from the record with the smallest ID), executes a loop process (S1771), and terminates the average value calculation process.

[0123] The average value calculation processing unit 177 adds one record to the comparison average information 138, registers a value in the ID column of the added record so that it is a consecutive number starting from "1", and registers the value of the corresponding fairness-requiring attribute of the record selected in the most recent step S1771 in the fairness-requiring attribute value combination information 132 in the columns of each fairness-requiring attribute of the added record (S1772).

[0124] The average value calculation processing unit 177 acquires, from among the records of the input log information 111, records in which the values ​​of each fairness-requiring attribute match the combination of fairness-requiring attribute values ​​indicated by the record selected in the most recent step S1771 in the fairness-requiring attribute value combination information 132, and the ID value is included in the log ID column of the comparison input correspondence information 136 (S1773).

[0125] The average value calculation processing unit 177 executes a loop process of sequentially executing the processes of steps S1775 to S1779 for each of the values ​​indicating the repayment results, ie, “success” and “failure” (for example, selecting “success” and “failure” in that order) (S1774).

[0126] The average value calculation processing unit 177 identifies, from the repayment result log information 113, the ID of the record whose result value is the repayment result selected in the most recent step S1774, and obtains the record having the identified ID from the records obtained in the most recent step S1773 of the input log information 111 (S1775).

[0127] The average value calculation processing unit 177 identifies a record in the comparison input correspondence information 136 that has a log ID value that is the same as the ID of the record obtained in the most recent step S1775 of the input log information 111, and obtains all values ​​of the ID column in the identified record (S1776).

[0128] The average calculation processing unit 177 obtains the loanable amount and repayment possibility value of the record having the same ID as the ID obtained in step S1776 from the comparison result information 137 (S1777). The average calculation processing unit 177 calculates the average value of the loanable amount and repayment possibility value obtained in step S1777 (S1778).

[0129] The average value calculation processing unit 177 registers the average loanable amount and repayment possibility calculated in step S1778 in the average loanable amount if successful column and average repayment possibility if successful column if the value of the repayment result selected in the most recent step S1774 for the record added in the most recent step S1772 to the comparison average information 138 if the value of the repayment result selected in the most recent step S1774 is "success", or in the average loanable amount if unsuccessful column and average repayment possibility if unsuccessful column if the value of the repayment result selected in the most recent step S1774 is "failure" (S1779).

[0130] 30 is a flowchart showing an example of the ratio calculation process by the ratio calculation processing unit 178. The ratio calculation processing unit 178 reads the record in the first row of the confirmation target result information 134 (i.e., the output by the learning model 103 for the attributes of the new target in the input attribute information 131) and the record in the first row of the comparison average information 138 (i.e., the average of the screening results for each repayment result of past targets who have the same combination of values ​​of the fairness-requiring attributes as the new target in the input attribute information 131) (S1780).

[0131] The ratio calculation processing unit 178 calculates the difference obtained by subtracting the repayment possibility value of the record read in step S1780 of the confirmation target result information 134 from the average repayment possibility value in case of success of the record read in step S1780 of the comparison average information 138, and registers the calculated difference in the "success-result" column of the record whose information column value is "loanable amount" in the correction ratio information 139 (S1781).

[0132] The ratio calculation processing unit 178 calculates the difference by subtracting the value of the average loanable amount in case of failure of the record read in step S1780 of the comparison average information 138 from the value of the loanable amount of the record read in step S1780 of the confirmation target result information 134, and registers the calculated difference in the ``Result-Failure'' column of the record whose information column value is ``Loanable Amount'' in the correction ratio information 139 (S1782).

[0133] The ratio calculation processing unit 178 calculates the difference obtained by subtracting the repayment possibility value of the record read in step S1780 of the confirmation target result information 134 from the average repayment possibility value in case of success of the record read in step S1780 of the comparison average information 138, and registers the calculated difference in the "success-result" column of the record whose information column value of the corrected ratio information 139 is "repayment possibility" (S1783).

[0134] The ratio calculation processing unit 178 calculates the difference obtained by subtracting the value of the average repayment possibility at failure of the record read in step S1780 of the comparison average information 138 from the value of the repayment possibility of the record read in step S1780 of the confirmation target result information 134, and registers the calculated difference in the "Result-Failure" column of the record whose information column value is "Repayment possibility" in the corrected ratio information 139 (S1784).

[0135] The ratio calculation processing unit 178 registers the ratio between the value in the "Success-Result" column and the value in the "Result-Failure" column of a record whose information column value in the corrected ratio information 139 is "Loanable Amount" in the ratio column of that record (S1785).

[0136] The ratio calculation processing unit 178 registers the ratio between the value in the "Success-Result" column and the value in the "Result-Failure" column of the record in which the value in the information column of the corrected ratio information 139 is "Repayment Possibility" (S1786), and terminates the ratio calculation process.

[0137] 31 is a flowchart showing an example of the output value correction process by the output value correction processing unit 179. The output value correction processing unit 179 reads all records of the confirmation target result information 134, records from the second row onwards of the comparison average information 138 (i.e., the average of the screening results for each repayment result of the new target and past target who have different combinations of the fairness requirement attribute values ​​of the input attribute information 131), and all records of the correction ratio information 139 (S1790).

[0138] The output value correction processing unit 179 executes a loop process (S1791) in which the processes of steps S1792 to S1794 are executed in order for each record from the second row onwards read in step S1790 of the comparison average information 138 (for example, for records selected one by one starting from the record with the smallest ID).

[0139] The output value correction processing unit 179 adds one record to the correction candidate information 140, and registers the ID value of the record selected in the most recent step S1791 of the comparison average information 138 in the ID column of the added record (S1792).

[0140] The output value correction processing unit 179 calculates and registers an output value to be registered in the column of the corrected loanable amount of the record added in the most recent step S1792 of the correction candidate information 140 (S1793). Specifically, the output value correction processing unit 179 calculates the output value so that the ratio of the value obtained by subtracting the output value from the value of the average loanable amount in case of success of the record selected in the most recent step S1791 of the comparison average information 138 and the value obtained by subtracting the value of the average loanable amount in case of failure of the record in the comparison average information 138 from the output value becomes the value of the ratio column in the record whose information column value of the correction ratio information 139 is "loanable amount".

[0141] The output value correction processing unit 179 calculates and registers an output value to be registered in the repayment possibility column of the record added in the most recent step S1792 of the correction candidate information 140 (S1794). Specifically, the output value correction processing unit 179 calculates the output value so that the ratio of the value obtained by subtracting the output value from the value of the average repayment possibility at success of the record selected in the most recent step S1791 of the comparison average information 138 and the value obtained by subtracting the average repayment possibility at failure of the record in the comparison average information 138 from the output value becomes the value in the ratio column of the record whose information column value of the correction ratio information 139 is "average repayment possibility".

[0142] The output value correction processing unit 179 registers the maximum value (i.e., the score indicating the most favorable screening result for the new target) between the corrected available loan amount value in the correction candidate information 140 and the available loan amount value in the confirmation target result information 134 in the available loan amount column of the first row record of the corrected output information 141 (S1795). The output value correction processing unit 179 registers the maximum value (i.e., the score indicating the most favorable screening result for the new target) between the corrected repayment possibility value in the correction candidate information 140 and the repayment possibility value in the confirmation target result information 134 in the repayment possibility column of the first row record of the corrected output information 141 (S1796).

[0143] When a loan is extended to a new target, it is unclear whether the new target will succeed or fail in repayment. In other words, when a loan is extended to a new target for loan screening, it is not clear whether the result will be a false positive or a positive. However, in steps S1795 and S1796, the output value correction processing unit 179 adopts the maximum value of the loanable amount and the repayment possibility, so that the difference in the average value for each value of the fairness requirement attribute in the subsequent learning results becomes smaller in the cases where the new target will succeed or fail in repayment in the future. Therefore, it becomes easier to perform screening that satisfies the fairness requirements based on Equalized Odds.

[0144] In the above example, the maximum values ​​were selected for the available loan amount and repayment possibility of the corrected output information 141, but the output value correction processing unit 179 may calculate a weighted sum using a specified weight of the available loan amount and repayment possibility for each record of the correction candidate information 140 and the confirmation target result information 134, and register the combination of available loan amount and repayment possibility indicated by the record with the largest calculated weighted sum (i.e., the combination of available loan amount and repayment possibility that is most advantageous for a new target person) as the available loan amount and repayment possibility of the corrected output information 141.

[0145] If the value registered in step S1795 is different from the loanable amount value in the first record of the confirmation target result information 134 and / or the value registered in step S1796 is different from the repayment possibility value in the first record of the confirmation target result information 134, the output value correction processing unit 179 registers ``YES'' in the correction column of the first record of the corrected output information 141, otherwise it registers ``NO'' (S1797) and terminates the output value correction process.

[0146] 32 is a flowchart showing an example of display processing by the display processing unit 180. The display processing unit 180 reads all records (only one row) of the correction output information 141 (S1800). The display processing unit 180 displays a loan examination support information display screen on the output device 3 (S1801). Details of the loan examination support information display screen will be described later with reference to FIG. 33.

[0147] The display processing unit 180 displays the value of the loanable amount column of the record read in step S1800 on the loan screening support information display screen displayed in step S1801 (S1802). The display processing unit 180 displays the value of the repayment possibility column of the record read in step S1800 on the loan screening support information display screen displayed in step S1801 (S1803).

[0148] The display processing unit 180 judges whether the value of the correction column of the record read in step S1800 is "YES" (S1804). When the display processing unit 180 judges that the value of the correction column is "YES" (S1804: YES), the display processing unit 180 displays the reason for the decision indicating that the screening result to which the correction has been applied (i.e., the screening result in which the screening result output from the learning model 103 has been corrected) has been output on the loan screening support information display screen displayed in step S1801 (S1805), and ends the display process.

[0149] If the display processing unit 180 determines that the value of the correction column is not "YES" (S1804: NO), it displays the reason for the decision indicating that the screening result based on machine learning (i.e., the screening result output from the learning model 103) has been output on the loan screening support information display screen displayed in step S1801 (S1806), and terminates the display processing.

[0150] In addition, the display processing unit 180 may display, for example, as reference information for the loan officer, values ​​of the loanable amount and repayment possibility indicated by the correction candidate information 140 and the confirmation target result information 134 that were not registered in the correction output information 141, as well as comparison average information 138 and correction ratio information 139, etc. on the loan screening support information display screen.

[0151] 33 is a diagram showing an example of the screen configuration of the loan examination support information display screen 32. The loan examination support information display screen 32 includes, for example, a loanable amount display area 321, a repayment possibility display area 322, and a decision reason display area 323.

[0152] The loanable amount display area 321 displays the loanable amount value in step S1802. The repayment possibility display area 322 displays the repayment possibility value in step S1803. The determination reason display area 323 displays the determination reason in step S1805 or step S1806.

[0153] In addition, the string "Corrections applied to ensure fairness" displayed in the decision reason display area 323 of the example loan screening support information display screen 32 in Figure 33 is an example of a decision reason indicating that the screening results to which corrections have been applied, which are displayed when the processing of step S1805 is performed, have been output.

[0154] When the processing of step S1806 is executed, the reason for decision display area 323 displays, for example, a string such as "based on machine learning results" as a reason for decision indicating that the review results based on machine learning have been output.

[0155] By displaying the reason for the decision in the decision reason display area 323, the user can easily confirm whether the inference result output from the learning model 103 has been adopted or whether the inference result output from the learning model 103 has been corrected to satisfy the fairness requirements.

[0156] As described above, the loan screening support system 1 of this embodiment generates confirmation target result information 134 indicating the scores (loanable amount and repayment possibility) based on the learning model 103 for a new target and a changed target in which the value of the fairness requirement attribute of the new target has been changed.

[0157] Furthermore, the loan screening support system 1 calculates the average score of comparison subjects, who are past subjects whose fairness judgment attributes are similar to the new subject, for each combination of the values ​​of the fairness requirement attributes and the repayment results.

[0158] Furthermore, the loan screening support system 1 calculates the ratio between the score of the new target indicated in the confirmation target result information 134, the average score of comparison targets who have the same value of the fairness requirement attribute as the new target and have succeeded in repayment, and the average score of comparison targets who have the same value of the fairness requirement attribute as the new target and have failed in repayment. This ratio indicates the position of the new target's score in the distribution of scores for each repayment result of comparison targets who have the same value of the fairness requirement attribute as the new target (i.e., the distribution of scores of positive and false positive targets for the comparison targets).

[0159] Furthermore, the loan screening support system 1 generates correction candidate information 140 by correcting the score of the changed target indicated in the confirmation target result information 134, the average score of the comparison target who has the same value of the fairness requirement attribute as the changed target and succeeded in repayment, and the average score of the comparison target who has the same value of the fairness requirement attribute as the changed target and failed in repayment so as to satisfy the above-mentioned ratio. In other words, the correction candidate information 140 is corrected so that the score of the changed target is in a similar position in a similar distribution.

[0160] The loan screening support system 1 selects the most favorable score for the new target among the scores indicated by the confirmation target result information 134 and the correction candidate information 140. If a different past target is assumed to have an input value for a fairness judgment attribute that is different from the range of the fairness judgment attribute of the input value for the new target, the score of the past target does not affect the calculation of the confirmation target result information 134 and the correction candidate information 140, and therefore does not affect the score of the new target output by the loan screening support system 1.

[0161] For this reason, for groups of new applicants with different fairness judgment attributes, there may be differences in the average for each group of new applicants in the results output by the loan screening support system 1. In other words, differences in fairness judgment attributes may affect the screening results.

[0162] On the other hand, if the maximum score selected by the loan screening support system 1 is derived from the correction candidate information 140, the selected score is a score that has been corrected so as to be in a similar position in the distribution of scores for attribute values ​​with fairness judgment attributes in the same range as the input value (of the new target) and with different fairness-requiring attributes. Therefore, the score of the new target is corrected to the highest score compared to a group of other targets with different fairness-requiring attributes.

[0163] Therefore, as the loan screening support system 1 continues to output for multiple new subjects, the difference between groups of subjects with different fairness-requiring attributes becomes smaller whether the new subject is a positive subject or a false positive, and the fairness requirements based on Equated Odds can be met over time.

[0164] Furthermore, if the maximum score selected by the loan screening support system 1 is derived from the confirmation target result information 134, the selected score is a score for the same fairness judgment attribute as the input value (of the new target) and for a different or the same attribute value of the fairness requirement attribute. In this case, the fairness requirement between new targets is met by correcting to eliminate the possibility that the loan screening support system 1 will give a higher value for another new target with a different fairness requirement attribute and the same input value of the fairness judgment attribute.

[0165] For the above reasons, the loan screening support system 1 can select a score that is influenced by the fairness judgment attribute and satisfies the fairness requirement based on the Equated Odds with respect to the fairness requirement attribute.

[0166] The loan screening support system 1 of this embodiment inputs attributes including a fairness-requiring attribute into the learning model 103 to generate confirmation target result information 134, then generates correction candidate information 140 by correcting the confirmation target result information 134 based on comparison average information 138 and correction ratio information 139, and selects the most advantageous score for the new target from the scores indicated by the confirmation target result information 134 and the correction candidate information 140. In other words, since the score that best satisfies the fairness requirement based on the Equated Odds is selected from the score output from the learning model 103 reflecting the value of the fairness-requiring attribute and the corrected value of the score, even if the attribute corresponding to the fairness-requiring attribute is changed for social reasons, legal regulations, etc., it is not necessary to reconstruct the learning model 103.

[0167] In this embodiment, the loan screening support system 1 has been described, but by appropriately selecting the fairness requirement attribute, fairness judgment attribute, and reference attribute, replacing the loanable amount and repayment possibility with other scores indicating the screening result, and replacing the repayment result with the result of other events that occur based on the actions taken by the subject who has benefited from the screening being approved, it is possible to provide support for any type of screening for a person.

[0168] When the loan screening support system 1 of this embodiment is applied to support for screening of life insurance, for example, the loanable amount and repayment possibility may be replaced with the insurance payable amount and insurance claim payment possibility, etc., and the repayment result may be replaced with the insurance claim payment result, etc. When the loan screening support system 1 of this embodiment is applied to character screening in employment examinations, for example, the loanable amount and repayment possibility may be replaced with a score for evaluating a person, etc., and the repayment result may be replaced with whether the person quit within a specified period after joining the company, etc.

[0169] The present invention is not limited to the above-mentioned embodiment, and various modifications are included. For example, the above-mentioned embodiment has been described in detail to clearly explain the present invention, and is not necessarily limited to those having all the configurations described. It is also possible to replace a part of the configuration of one embodiment with the configuration of another embodiment, and it is also possible to add the configuration of another embodiment to the configuration of one embodiment. It is also possible to add, delete, or replace a part of the configuration of each embodiment with another configuration.

[0170] In addition, the above-mentioned configurations, functions, processing units, processing means, etc. may be realized in part or in whole by hardware, for example, by designing them as integrated circuits. In addition, the above-mentioned configurations, functions, etc. may be realized in software by a processor interpreting and executing a program that realizes each function. Information such as the program, table, file, etc. that realizes each function can be stored in a memory, a recording device such as a hard disk or SSD (Solid State Drive), or a recording medium such as an IC card, SD card, or DVD.

[0171] In addition, the control lines and information lines shown are those that are considered necessary for the explanation, and not all control lines and information lines in the product are necessarily shown. In reality, it can be considered that almost all components are connected to each other. [Explanation of symbols]

[0172] 1 Loan screening support system, 2 Input device, 3 Output device, 11 CPU, 12 Memory, 13 Auxiliary storage device, 111 Input log information, 112 Judgment result log information, 113 Repayment result log information, 121 Fairness required attribute information, 122 Fairness judgment attribute information, 103 Learning model, 131 Input attribute information, 132 Fairness required attribute value combination information, 133 Changed input attribute information, 134 Confirmation target result information, 135 Comparison input information, 136 Comparison input correspondence information, 137 Comparison result information, 138 Comparison average information, 139 Correction ratio information, 140 Correction candidate information, 141 Correction output information, 151 Learning processing unit, 152 Loan screening support information inference unit

Claims

1. A screening support system for supporting screening of a person, comprising: A processor and a memory, The memory includes: Log information indicating a plurality of attributes including a fairness attribute and a fairness judgment attribute of the past subjects of the review, and a score indicating the review result of the review of the past subjects; a learning model obtained by machine learning on training data in which the attributes of the past subject are set as objective variables and the scores of the past subject are set as explanatory variables; and Input attribute information indicating the plurality of attributes of the new subject of the screening; and fairness judgment attribute information indicating a condition under which the values ​​of the fairness judgment attribute are similar; The processor, acquiring, from the log information, the plurality of attributes of a comparison subject who is a past subject and whose value of the fairness judgment attribute is similar to that of the new subject, by referring to the conditions indicated by the fairness judgment attribute information; inputting a plurality of attributes of the comparison subject into the learning model to obtain a first score; inputting the new subject and a changed subject in which the value of the fairness-required attribute of the new subject has been changed, and the plurality of attributes of each of the new subject and the changed subject, into the learning model to obtain a second score; An examination support system that outputs a score indicating the examination result for the new subject based on the position of the second score in the distribution of the first scores.

2. The examination support system according to claim 1, The log information further indicates the results of events that occur as a result of actions taken by the subject in the past; The processor, Calculating the average value of the first score of a comparison subject having the same value of the fairness attribute as the new subject for each outcome of the event of the comparison subject as the distribution in the comparison subject; An examination support system that calculates a ratio based on the difference between the average value for each result of the event and the second score of the new subject as the position of the second score.

3. The examination support system according to claim 2, The examination is a loan examination, The score indicating the result of the assessment includes the loanable amount and the possibility of repayment, The result of the event indicates the result of repayment by the past subject of the loan amount to the past subject, in the screening support system.

4. The examination support system according to claim 2, The processor, Calculating the average value of the first score of a comparison subject having a value of the fairness attribute different from that of the new subject for each outcome of the event and for each value of the fairness attribute of the comparison subject as the distribution in the comparison subject having the value of the fairness attribute; For each value of the fairness attribute, calculate a correction score for each value of the fairness attribute that is different from that of the new target based on the average value for each outcome of the event corresponding to the value of the fairness attribute and the ratio; An examination support system that selects the most favorable score for the new subject from the corrected score and the second score, as a score that indicates the examination result for the new subject.

5. The examination support system according to claim 4, connected to a display device, The processor, When the score indicating the screening result for the new subject is selected from the corrected scores, a display indicating that the score output from the learning model has been corrected is displayed on the display device; An examination support system that, when a score indicating the examination result for the new subject is selected from the second scores, displays on the display device a message indicating that the score output from the learning model has been adopted.

6. 1. A screening support method using a screening support system that supports screening of a person, comprising: The examination support system includes a processor and a memory, The memory includes: Log information indicating a plurality of attributes including a fairness attribute and a fairness judgment attribute of the past subjects of the review, and a score indicating the review result of the review of the past subjects; a learning model obtained by machine learning on training data in which the attributes of the past subject are set as objective variables and the scores of the past subject are set as explanatory variables; and Input attribute information indicating the plurality of attributes of the new subject of the screening; and fairness judgment attribute information indicating a condition under which the values ​​of the fairness judgment attribute are similar; The examination support method includes: The processor refers to a condition indicated by the fairness judgment attribute information and acquires, from the log information, the multiple attributes of a comparison subject who is a past subject whose value of the fairness judgment attribute is similar to that of the new subject; The processor inputs a plurality of attributes of the comparison subject into the learning model to obtain a first score; The processor inputs the new subject and a changed subject in which a value of the fairness-required attribute of the new subject has been changed, and the plurality of attributes of each of the new subject and the changed subject into the learning model to obtain a second score; A screening support method in which the processor outputs a score indicating the screening result for the new subject based on the position of the second score in the distribution of the first scores.

7. The examination support method according to claim 6, The log information further indicates the results of events that occur as a result of actions taken by the subject in the past; The examination support method includes: The processor calculates an average value of the first score of a comparison subject having the same value of the fairness attribute as the new subject for each outcome of the event of the comparison subject as the distribution of the comparison subject; An examination support method in which the processor calculates a ratio based on the difference between the average value for each outcome of the event and the second score of the new subject as the position of the second score.

8. The examination support method according to claim 7, The examination is a loan examination, The score indicating the result of the assessment includes the loanable amount and the possibility of repayment, The result of the event indicates a repayment result by the past subject of the loan amount to the past subject, in the screening support method.

9. The examination support method according to claim 7, The processor calculates an average value of the first scores of a comparison subject having a value of the fairness attribute different from that of the new subject for each outcome of the event and for each value of the fairness attribute of the comparison subject, as the distribution in the comparison subject having the value of the fairness attribute; The processor calculates, for each value of the fairness-requiring attribute, a correction score for each value of the fairness-requiring attribute that differs from the new subject based on the average value for each outcome of the event corresponding to the value of the fairness-requiring attribute and the ratio; The screening support method, wherein the processor selects the most favorable score for the new subject from the corrected score and the second score, as a score indicating the screening result for the new subject.

10. The examination support method according to claim 9, The examination support system is connected to a display device, The examination support method includes: When the processor selects a score indicating the screening result for the new subject from the corrected scores, a display is displayed on the display device indicating that the score output from the learning model has been corrected; An examination support method in which, when the processor selects a score indicating the examination result for the new subject from the second scores, a display indicating that the score output from the learning model has been adopted is displayed on the display device.

Images