Machine learning program, machine learning method and information processing device

Abstract

To provide a machine learning program that corrects a machine learning model so as to be applicable to various fairness evaluation standards, a machine learning method and an information processing device.SOLUTION: In an AI system, a control unit 115 includes: a classification result obtaining unit that obtains plural classification results on plural pieces of data obtained by inputting each of the plural pieces of data into a machine learning model; an attribute identifying unit that identifies, on the basis of the plural classification results, first plural attributes that have a difference in attribute value between plural pieces of first data and plural pieces of second data satisfying a standard among plural attributes contained in the plural pieces of first data classified in a first group and in the plural pieces of second data classified in a second group; a label determining unit that determines the respective labels of the plural pieces of data on the basis of a first reference obtained by combining the plural first attributes; and a training unit that trains the machine learning model on the basis of the determined labels and of the plural pieces of data.SELECTED DRAWING: Figure 7

Description

[Technical Field]

[0001] The present invention relates to a machine learning program, a machine learning method, and an information processing device. [Background technology]

[0002] In the field of fairness-conscious machine learning, efforts are being made to modify machine learning models that reflect discriminatory biases contained in training data to make them fairer. To flexibly capture fairness, which changes depending on the situation, there are techniques that modify the fairness of machine learning models based on user evaluations. In the following explanation, machine learning models may be referred to simply as "models."

[0003] Metrics such as acceptance rate, false positive rate, and false negative rate based on attributes such as race and gender are sometimes used to measure bias in machine learning models and data.

[0004] On the other hand, what constitutes fairness varies widely, and the standards for fairness change depending on the situation, so there are cases where automatic model correction based on traditional indicators alone cannot adequately address the situation. For example, in the job market, men are generally given preferential treatment over women, but there are also certain occupations that are more difficult for men to obtain than women.

[0005] Therefore, there is a technique to fairly correct machine learning models based on the user's subjective evaluation, which prompts the user to correct the model while displaying fairness and accuracy metrics. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] Patent Publication No. 2021-012593 [Patent Document 2] Special Publication No. 2019-526107 [Patent Document 3] U.S. Patent Publication No. 2014 / 0249872 [Patent Document 4] U.S. Publication No. 2019 / 0102700 Summary of the Invention [Problem to be solved by the invention]

[0007] However, the above-mentioned conventional technologies do not take into account the "criteria discovery process" in which users search for evaluation criteria that they truly want to use as the basis for fair decision-making. As a result, conventional technologies only present existing fairness metrics and encourage users to modify the model, which can lead to users abandoning the system because they are unable to modify the model to a state that they believe is truly fair.

[0008] One aspect is to modify machine learning models to accommodate a variety of fairness metrics. [Means for solving the problem]

[0009] In one aspect, a machine learning program obtains a plurality of classification results for a plurality of data by inputting each of the plurality of data into a machine learning model, and, based on the plurality of classification results, identifies a first plurality of attributes among a plurality of attributes contained in a first plurality of data classified into a first group and a second plurality of data classified into a second group of the plurality of data, the first plurality of attributes having a difference in attribute values ​​between the first plurality of data and the second plurality of data that meets a criterion; determines a label for each of the plurality of data based on a first index combining the first plurality of attributes; and trains the machine learning model based on the determined labels and the plurality of data. [Effects of the Invention]

[0010] On the one hand, machine learning models can be adapted to accommodate a variety of fairness metrics. [Brief explanation of the drawings]

[0011] [Figure 1] FIG. 10 is a diagram illustrating a model correction in a related example. [Figure 2] FIG. 10 is a diagram illustrating a problem with model correction in a related example. [Figure 3] FIG. 10 is a diagram illustrating an example of a model correction by a satisfied user in the embodiment. [Figure 4] FIG. 10 is a diagram illustrating an example of a model correction by a dissatisfied user in the embodiment. [Figure 5] 10A and 10B are diagrams illustrating an example of a process for specifying a fairness standard in an embodiment. [Figure 6] FIG. 1 is a block diagram schematically illustrating an example of the hardware configuration of an AI (machine learning) system according to an embodiment. [Figure 7] FIG. 1 is a block diagram illustrating an example of the software configuration of an AI system according to an embodiment. [Figure 8] 10 is a flowchart illustrating an example of a process for specifying a fairness standard according to an embodiment. [Figure 9] FIG. 9 is a diagram illustrating an example of a process for accepting a proposed revision of the machine learning model shown in FIG. 8. [Figure 10] 9 is a diagram illustrating an example of a process for calculating the degree of approach to existing metrics and a process for determining the degree of approach shown in FIG. 8. FIG. [Figure 11] FIG. 9 is a diagram illustrating an example of a calculation process for an index with a large difference between an accepted group and a rejected group shown in FIG. 8. [Figure 12] 9 is a diagram illustrating an example of a process for presenting an index having a large difference between the accepted group and the rejected group shown in FIG. 8. FIG. [Figure 13] 9 is a diagram illustrating an example of a process for calculating and presenting an index that maximizes the difference between the accepted group and the rejected group shown in FIG. 8. FIG. [Figure 14] 9A to 9C are diagrams illustrating an example of a process for correcting the model shown in FIG. 8 and a process for presenting model candidates after the correction. [Figure 15] 9 is a diagram illustrating an example of a process for determining whether a specific model candidate shown in FIG. 8 has been selected. [Figure 16] FIG. 10 is a diagram illustrating a first example of the effect of the correction process of the machine learning model in the embodiment. [Figure 17] FIG. 10 is a diagram illustrating a second example of the effect of the correction process of the machine learning model in the embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0012] [A] Related Examples FIG. 1 is a diagram illustrating model modification in a related example.

[0013] In the example shown in Figure 1, the AI ​​system 6 makes a decision on a loan application, and the user 7 refers to the predicted results output by the AI ​​system 6.

[0014] In the example shown by symbol A1, the prediction result is registered in association with ID, gender, income, loan amount, result label, and correction label. The result label is registered with a value of Accept or Reject. User 7 corrects the result label included in the prediction result and registers the value of Accept or Reject as the correction label. Note that user 7 may also correct not only the result label but also the attributes of the loan application user and the weighting values ​​for the attributes.

[0015] The AI ​​system 6 retrains (re-machines) the machine learning model based on the corrected labels input by the user 7.

[0016] Reference symbol A2 shows the Disparate Impact related to gender. Disparate Impact is the ratio of the acceptance rate between attribute values, and may be determined to be fair if it is 0.8 (in other words, the fairness index) or higher. The fairness index is a threshold above (or below) which a certain fairness metric value is considered fair. In the example shown in reference symbol A2, the current value is 0.5, while the fairness index is 0.8.

[0017] FIG. 2 is a diagram illustrating the problem of model correction in the related example.

[0018] In the graph shown in Figure 2, the horizontal axis shows the fairness standard considered by user 7, and the vertical axis shows the existing fairness standard. Note that in the graph shown in Figure 2, the indicator considered fair by user 7 may not actually be on an axis that is perpendicular to the existing fairness indicator, but it is shown as an axis that is perpendicular to the existing fairness indicator.

[0019] As shown by symbol B1, the AI ​​system 6 displays an existing fairness metric to the user 7. For example, the AI ​​system 6 displays, "According to existing fairness standards, anything above this value is considered fair." The "value" displayed by the AI ​​system 6 is the value shown by symbol B11 in Figure 2. Here, a fairness metric is a scale used to measure fairness. For example, a metric called "Disparate Impact" is a scale expressed as (adoption rate of discriminated groups) / (adoption rate of favored groups).

[0020] As indicated by symbol B2, the user 7 modifies the index values ​​of the model in the direction prompted by the AI ​​system 6. In the example shown in Figure 2, the index values ​​of the original model are indicated by black circles, and the index values ​​of the modified model are indicated by white circles.

[0021] As shown by reference symbol B3, the user 7 thinks, for example, "I don't feel like the results have become fair."

[0022] As shown by symbol B4, user 7 performs a criteria search process. Since user 7 does not know how to move the model's index values, he or she performs a search as appropriate. Here, criteria include both metrics and fairness criteria. Once both metrics and fairness criteria are determined, the criteria are determined.

[0023] As shown by symbol B5, the AI ​​system 6 does not take into account the criteria search process and therefore does not change the information presentation.

[0024] Then, as shown by symbol B6, the model index value does not reach the fairness desired by user 7, and user 7 leaves the AI ​​system 6.

[0025] In the relevant example shown in Figure 2, the fairness criteria considered (used) by the AI ​​system 6 are limited, and are mainly derived from the proportion of positive examples and the difference / ratio of accuracy between discriminated and favored groups.

[0026] On the other hand, there are various evaluation criteria that users 7 consider when determining fairness. For example, when deciding on a loan, it is more fair to decide based on the ratio of income to loan amount.

[0027] Therefore, User 7 may feel uneasy about the metrics presented by AI System 6. If User 7 feels uneasy, he or she will explore how to modify the model through a criteria discovery process to make it fair according to the fairness metrics that he or she considers appropriate.

[0028] In the related example, AI system 6 does not consider this criteria search process and only displays the existing fairness metrics, so user 7 ends up making corrections without reaching a satisfactory fairness standard.

[0029] [B] Embodiment An embodiment will be described below with reference to the drawings. However, the embodiment described below is merely an example, and is not intended to exclude various modifications or application of techniques not explicitly stated in the embodiment. For example, this embodiment can be implemented with various modifications within the scope of its purpose. Furthermore, each figure does not intend to include only the components shown in the figure, but may include other functions, etc.

[0030] In the following drawings, the same reference numerals denote similar parts, and therefore the description thereof will be omitted.

[0031] [B-1] Configuration example Fig. 3 is a diagram illustrating an example of a model correction by a satisfied user in the embodiment. Fig. 4 is a diagram illustrating an example of a model correction by a non-satisfied user in the embodiment.

[0032] In the embodiment, the AI ​​system 1 has the user 2 present their fairness criteria in advance and then modifies the machine learning model.

[0033] However, user 2 may not always be aware of the fairness criteria he or she considers to be important from the beginning, and may not be able to specify the fairness criteria in advance.

[0034] Identifying fairness metrics requires understanding how modifications to the machine learning model change the value of the metric.

[0035] Therefore, User 2 is required to know how the metrics change through model modifications (for example, by adjusting the weights on attribute values ​​or reassigning correct labels), and to know what appropriate metrics are feasible. For example, in AI System 1 that makes loan application decisions, it is conceivable that incompatible metrics are specified: it is natural that the acceptance rate should vary depending on income, but it should be equalized by age. Because income is thought to tend to increase as age increases, it is highly likely that equalization by age and changes in acceptance rate due to income are incompatible.

[0036] Therefore, in the embodiment, the AI ​​system 1 identifies metrics that determine fairness evaluation criteria during user 2's interaction with the AI ​​system 1.

[0037] In Figures 3 and 4, black circles indicate the index values ​​of the original model, and white circles indicate the index values ​​of the corrected model.

[0038] When a user (hereinafter referred to as a "satisfied user") 2 who is satisfied with the existing fairness standards shown in Figure 3 modifies the model, as shown by symbol C1, the model's index value approaches the fairness based on the existing metrics each time the model is modified.

[0039] When a user (hereinafter referred to as a "dissatisfied user") 2 who is not satisfied with the existing fairness criteria shown in Figure 4 modifies the model, the result is as shown by symbol D1. For example, even if the model is modified, the index value of the model moves away from the fairness based on the existing metrics or does not approach the fairness based on the existing metrics (in other words, the deviation or metric value does not change).

[0040] Therefore, AI system 1 changes the information presented to user 2 who has entered the criteria search process.

[0041] FIG. 5 is a diagram illustrating an example of a process for specifying a fairness standard according to an embodiment.

[0042] AI system 1 identifies the evaluation criteria that separate the accepted group from the rejected group among the models revised by unconvinced user 2 as the fairness criteria.

[0043] The accept group is a group of instances that perform better in the revised model, such as positive examples or top-ranked instances in binary classification, while the reject group is a group of instances that perform worse in the revised model.

[0044] User 2's fairness basis indicates whether the attributes on which the decision is based appear to be fair. To achieve this, we can identify the attributes that are likely to be the basis for User 2's revision.

[0045] The AI ​​system 1 identifies multiple attributes that show a large difference between the accepted group and the rejected group as candidate criteria attributes. The AI ​​system 1 also combines the candidate criteria attributes to create composite metrics that further differentiate the accepted group from the rejected group.

[0046] AI System 1 then presents potential model modifications and asks User 2 to identify the metrics they actually want to consider.

[0047] In the example shown in Figure 5, as indicated by symbol E1, AI system 1 considers the difference between a discriminated group and a favored group as an existing evaluation criterion. As indicated by symbol E2, user 2 determines whether the attributes used in decision-making are appropriate based on their own evaluation criteria.

[0048] As shown at symbol E3, the AI ​​system 1 identifies the attributes that form the basis of the evaluation criteria for the user 2.

[0049] Reference symbol E4 shows the evaluation criteria for the accept group (A) and the reject group (R) for each of the attribute names α, β, and γ in a bar graph. As indicated by reference symbol E41, the AI ​​system 1 identifies attributes that have a larger difference between the accept group and the reject group than others by measuring the p-value of a statistical test, etc. In the example shown in Figure 5, as indicated by reference symbol E42, the AI ​​system 1 identifies attribute names α and β as criteria attribute candidates.

[0050] As shown at symbol E5, the AI ​​system 1 identifies composite metrics. In the example shown at symbol E6, the AI ​​system 1 identifies α+β, α-β, α / β, and α*β as composite metrics.

[0051] As indicated by reference symbol E61, the AI ​​system 1 presents to user 2, among the identified composite metrics, α-β, α / β, which has a larger difference between the accepted group and the rejected group than the others. In the example indicated by reference symbol E62, user 2 selects the composite metric α-β.

[0052] FIG. 6 is a block diagram schematically illustrating an example of the hardware configuration of an AI system 1 according to an embodiment.

[0053] As shown in Figure 6, the AI ​​system 1 is an example of an information processing device and includes a Central Processing Unit (CPU) 11, a memory unit 12, a display control unit 13, a storage device 14, an input interface (IF) 15, an external recording medium processing unit 16, and a communication IF 17.

[0054] The memory unit 12 is an example of a storage unit, and is illustratively a read-only memory (ROM) or a random access memory (RAM). A program such as a basic input / output system (BIOS) may be written to the ROM of the memory unit 12. The software program in the memory unit 12 may be read and executed by the CPU 11 as appropriate. The RAM of the memory unit 12 may be used as a temporary storage memory or a working memory.

[0055] The display control unit 13 is connected to a display device 131 and controls the display device 131. The display device 131 is a liquid crystal display, an organic light-emitting diode (OLED) display, a cathode ray tube (CRT), an electronic paper display, or the like, and displays various information to an operator, etc. The display device 131 may be combined with an input device, such as a touch panel. The display device 131 displays various information to the user 2.

[0056] The storage device 14 is a storage device with high IO performance, and may be, for example, a dynamic random access memory (DRAM), an SSD, a storage class memory (SCM), or an HDD.

[0057] The input IF 15 may be connected to input devices such as a mouse 151 and a keyboard 152, and may control the input devices such as the mouse 151 and the keyboard 152. The mouse 151 and the keyboard 152 are examples of input devices, and an operator performs various input operations via these input devices.

[0058] The external recording medium processing unit 16 is configured to be able to load a non-transitory recording medium 160. The external recording medium processing unit 16 is configured to be able to read information recorded on the recording medium 160 when the recording medium 160 is loaded. In this example, the recording medium 160 is portable. For example, the recording medium 160 is a flexible disk, an optical disk, a magnetic disk, a magneto-optical disk, a semiconductor memory, or the like.

[0059] The communication IF 17 is an interface that enables communication with an external device.

[0060] The CPU 11 is an example of a processor, and is a processing device that performs various controls and calculations. The CPU 11 realizes various functions by executing an operating system (OS) and programs loaded into the memory unit 12. The CPU 11 may be a multiprocessor including multiple CPUs, a multi-core processor having multiple CPU cores, or a configuration having multiple multi-core processors.

[0061] The device for controlling the operation of the entire AI system 1 is not limited to the CPU 11, and may be, for example, any one of an MPU, GPU, APU, DSP, ASIC, PLD, and FPGA. Note that MPU is an abbreviation for Micro Processing Unit, GPU is an abbreviation for Graphics Processing Unit, and APU is an abbreviation for Accelerated Processing Unit. DSP is an abbreviation for Digital Signal Processor, and ASIC is an abbreviation for Application Specific Integrated Circuit. Furthermore, PLD is an abbreviation for Programmable Logic Device, and FPGA is an abbreviation for Field Programmable Gate Array.

[0062] Furthermore, the device for controlling the overall operation of the AI ​​system 1 may be a combination of two or more of a CPU, MPU, GPU, APU, DSP, ASIC, PLD, and FPGA. As an example, the AI ​​system 1 may include an accelerator (for convenience, reference numeral 11 is used) 11 in addition to the CPU 11. The accelerator 11 is hardware that executes arithmetic processing used in neural network calculations such as matrix operations, and performs training of machine learning models, etc. Examples of the accelerator 11 include a GPU, APU, DSP, ASIC, PLD, and FPGA.

[0063] FIG. 7 is a block diagram schematically illustrating an example of the software configuration of an AI system 1 according to an embodiment.

[0064] 6 functions as a classification result acquisition unit 111, an attribute specification unit 112, a label determination unit 113, and a training unit 114. The classification result acquisition unit 111, the attribute specification unit 112, the label determination unit 113, and the training unit 114 are examples of a control unit 115.

[0065] The classification result acquisition unit 111 acquires a plurality of classification results of a plurality of data obtained by inputting each of the plurality of data into a machine learning model.

[0066] The attribute identification unit 112 identifies a plurality of attributes included in each of the first plurality of data classified into the first group and the second plurality of data classified into the second group, among the plurality of data, based on the plurality of classification results acquired by the classification result acquisition unit 111. Of the identified plurality of attributes, the attribute identification unit 112 identifies a plurality of first attributes whose difference in attribute value (e.g., p-value) between the first plurality of data and the second plurality of data satisfies a criterion (e.g., top N attributes; N is a natural number). Note that the first group is an example of an accept group, and the second group is an example of a reject group.

[0067] The label determination unit 113 determines a label for each of the plurality of data based on a first index (in other words, a composite metric) that combines a first plurality of attributes.

[0068] The training unit 114 trains the machine learning model based on the labels determined by the label determination unit 113 and a plurality of data.

[0069] [B-1] Example of operation The process of specifying the fairness criteria in the embodiment will be described with reference to the flowchart (steps S1 to S13) shown in FIG.

[0070] At T=t, the prediction result by the machine learning model or the result at T=t-1 is presented to User 2, and an input of a proposed revision of the machine learning model is accepted from User 2 (Step S1).

[0071] It is calculated whether the index value is approaching the existing metrics (step S2).

[0072] It is determined whether the index value is approaching the existing metrics (step S3).

[0073] If the index value is approaching an existing metric (see the Yes route in step S3), existing metrics that are close in distance are displayed and a model modification is suggested to user 2 (step S4). Then, the process proceeds to step S10.

[0074] On the other hand, if the index value is not approaching the existing metrics (see the No route in step S3), the index with the largest difference between the accepted group and the rejected group is calculated (step S5).

[0075] The indices with the largest difference are presented as a ranking, and the selection by user 2 of an index that is likely to be useful in determining fairness is accepted (step S6).

[0076] Using the selected indices, the indices that produce the largest difference are calculated and presented as candidates (step S7).

[0077] The model is corrected based on the registration candidates (step S8).

[0078] The revised model candidates are presented to User 2 (step S9).

[0079] A model selection is accepted from User 2 (step S10).

[0080] It is determined whether a particular candidate has been selected (step S11).

[0081] If it is determined that a specific candidate has been selected (see the Yes route in step S11), it is determined that the fairness standard expected by user 2 has been reached (step S12), and the process of specifying the fairness standard ends.

[0082] On the other hand, if it is determined that no specific candidate was selected (see the No route in step S11), it is determined that the fairness standard expected by user 2 has not been reached, and the process returns to step S1 with T=t+1, and the interaction is repeated (step S13).

[0083] FIG. 9 is a diagram illustrating an example of the process of accepting a proposed revision of the machine learning model shown in FIG. 8 (step S1).

[0084] AI system 1 outputs the prediction results, attribute values, and fairness criteria information to user 2. During the first loop, AI system 1 directly outputs the prediction results based on machine learning. If feedback is received from user 2, AI system 1 outputs the results at time T=t-1.

[0085] Based on the output, user 2 inputs model correction information into AI system 1.

[0086] In the example shown in Figure 9, symbol F1 shows a table that AI system 1 uses to determine whether to lend money to a person based on their income, expenses, and loan amount. The label for T=t-1 indicates 1 for accept and 0 for reject.

[0087] Reference symbol F2 shows the Disparate Impact related to gender. Disparate Impact is the ratio of the acceptance rate between attribute values, and may be determined to be fair if it is a fairness index, for example, 0.8 or higher. The fairness index is a threshold above which (or below which) a certain fairness metric value is considered fair. In the example shown in reference symbol F2, the current value is 0.5, while the fairness index is 0.8.

[0088] As shown in symbol F3, user 2 modifies the labels to modify the model.

[0089] FIG. 10 is a diagram illustrating an example of the process of calculating the degree of proximity to existing metrics (step S2) and the process of determining the degree of proximity (step S3) shown in FIG.

[0090] The AI ​​system 1 modifies, in other words, retrains (recalculates) the model based on the label modification suggestions from the user 2, as indicated by symbols G1 and G2.

[0091] AI system 1 calculates whether the revised model approaches the fairness criteria set by AI system 1.

[0092] In the example shown in Figure 10, the Disparate Impact for gender in terms of fairness is calculated. If the proximity is not above a certain level, it is determined that User 2 is not satisfied with the existing fairness criteria presented and enters the criteria search process.

[0093] In the example shown by symbol G3, the value is closer to the fair value by 0.2, so user 2 is satisfied. On the other hand, in the example shown by symbol G4, the value is not closer at all (only 0), so it is determined that the criteria search process has begun.

[0094] FIG. 11 is a diagram illustrating an example of the calculation process (step S5) of the index with a large difference between the accepted group and the rejected group shown in FIG.

[0095] When AI System 1 enters the criteria discovery process, it considers attributes that User 2 is likely to use in making decisions.

[0096] AI System 1 calculates the statistical difference between reject and accept for each attribute using the corrected model.

[0097] In the example shown in Figure 11, the AI ​​system 1 performs a statistical test, such as a Welch t-test, between the reject and accept groups for income (H1), expenses (H2), and loan amount (H3), and measures the magnitude of the derived p-value. Then, as shown by symbol H4, the AI ​​system 1 registers the averages and p-values ​​of the accept and reject groups for each of the income, expenses, and loan amount in table format. The smaller the p-value, the more likely it is that there is a statistical difference. Note that various methods other than the Welch t-test may be used to measure the magnitude of the p-value.

[0098] FIG. 12 is a diagram illustrating an example of the process of presenting an index (step S6) showing a large difference between the accepted group and the rejected group shown in FIG.

[0099] The AI ​​system 1 determines the top N cases with the largest differences between the rejected and accepted groups calculated in step S5 of Figure 8.

[0100] AI system 1 presents the top N selected results to user 2 as candidate criteria attributes, which are attributes that user 2 is likely to use for fair decision-making.

[0101] AI system 1 accepts the selection of criteria attributes from user 2.

[0102] In the example shown in Figure 12, as indicated by symbol I1, the top two results with small p-values ​​and close to significant differences between groups are presented to user 2. As indicated by symbol I2, AI system 1 presents "income" and "expenses" to user 2 as candidate attributes to use as the basis for fair decisions. As indicated by symbol I3, user 2 determines that "both seem reasonable" and selects both "income" and "expenses." It is also possible to select only one of them.

[0103] FIG. 13 is a diagram illustrating an example of the process of calculating and presenting the index that maximizes the difference between the accepted group and the rejected group shown in FIG. 8 (step S7).

[0104] The AI ​​system 1 generates composite metrics that combine each attribute based on the criteria attributes selected in step S6 of Figure 8.

[0105] In the example shown by reference symbol J1 in Figure 13, the AI ​​system 1 generates multiple composite metrics that cover the four arithmetic operations between criteria attributes. The AI ​​system 1 calculates the difference between the accept and reject groups of the composite metrics. A p-value based on Welch's t-test may be used as a value indicating a statistical difference. The AI ​​system 1 determines the top M results (M is a natural number; in the example shown in Figure 13, M = 4) that have small p-values ​​and whose differences between groups are close to significance.

[0106] As shown by symbol J2, the AI ​​system 1 presents M pieces of data to the user 2 and asks them to select the most relevant one.

[0107] As shown by symbol J3, user 2 selects "income" and "income minus expenses" as likely metrics.

[0108] FIG. 14 is a diagram illustrating an example of the model correction process (step S8) shown in FIG. 8 and the process of presenting the corrected model candidates (step S9).

[0109] AI system 1 modifies the model to maximize the difference between the reject and accept groups for the composite metric selected by user 2. For example, AI system 1 re-labels the data to maximize the difference and retrains the model. In the example shown by symbol K1, the model is modified to maximize the difference between the reject and accept groups for "income" and "income minus expenditure," respectively.

[0110] Then, the AI ​​system 1 presents the results to the user 2 as shown by the symbol K2, and prompts the user 2 to select whether there is a corrected result that is acceptable to the user as shown by the symbol K3.

[0111] FIG. 15 is a diagram illustrating an example of the process of determining whether a specific model candidate shown in FIG. 8 has been selected (step S11).

[0112] As shown by symbol L1, the AI ​​system 1 asks the user 2 whether they would be satisfied with a revision of either the model that takes into account "income" or "income-expenses."

[0113] If the selection shown in step S10 of Figure 8 results in a satisfactory correction result, the AI ​​system 1 determines that user 2 has discovered the desired criteria, as shown by symbol L2, and ends use of the system.

[0114] On the other hand, if there is no satisfactory correction result, the AI ​​system 1 considers that the desired criteria have not been discovered, as shown by symbol L3, and repeats the process from step S1 in Figure 1.

[0115] [C] Effect FIG. 16 is a diagram illustrating a first example of the effect of the correction process of the machine learning model in the embodiment.

[0116] As shown by symbol M1, User 2 experimentally moves the indicator values ​​through a metrics exploration process. As shown by symbol M2, AI system 1 determines and presents the fairness desired by User 2.

[0117] This allows User 2 to enter the criteria discovery process and quickly arrive at a fair model that User 2 considers fair in a single interaction.

[0118] FIG. 17 is a diagram illustrating a second example of the effect of the correction process of the machine learning model in the embodiment.

[0119] AI system 1 can present to user 2, as shown by symbol N2, a variety of metrics that could not be identified in the related example shown in Figure 2 even if only a fairness index based on existing metrics, as shown by symbol N1, was used. Then, as shown by symbol N3, user 2 can select from the various metrics presented by AI system 1, thereby identifying appropriate metrics based on user 2's subjective opinion.

[0120] The classification result acquisition unit 111 acquires a plurality of classification results of the plurality of data obtained by inputting each of the plurality of data into a machine learning model. The attribute identification unit 112 identifies a plurality of attributes included in each of a first plurality of data classified into a first group and a second plurality of data classified into a second group, among the plurality of data, based on the plurality of classification results acquired by the classification result acquisition unit 111. The attribute identification unit 112 identifies a first plurality of attributes, among the identified plurality of attributes, for which a difference in attribute value between the first plurality of data and the second plurality of data satisfies a criterion. The label determination unit 113 determines a label for each of the plurality of data based on a first index combining the first plurality of attributes. The training unit 114 trains the machine learning model based on the labels determined by the label determination unit 113 and the plurality of data.

[0121] This allows us to modify machine learning models to adapt to a variety of fairness criteria.

[0122] The label determination unit 113 determines a label based on a first index represented by arithmetic operations between the first plurality of attributes, thereby making it possible to determine a label based on an appropriate first index.

[0123] The label determination unit 113 creates a plurality of first indices and presents them to the user 2, accepts a selection of one or more first indices by the user 2 from the plurality of first indices, and determines a label based on the one or more first indices whose selection has been accepted. This makes it possible to determine a label based on a first indices that satisfies the requirements of the user 2.

[0124] [D] Other The disclosed technology is not limited to the above-described embodiment, and various modifications can be made without departing from the spirit of the present embodiment. The configurations and processes of the present embodiment can be selected or combined as needed.

[0125] In the above-described embodiment, the correction process of the machine learning model for the classification task has been described, but the present invention is not limited to this. The correction process of the machine learning model in the embodiment can be applied to various tasks.

[0126] [E] Supplementary Note The following additional notes are provided regarding the above-described embodiments.

[0127] (Appendix 1) acquiring a plurality of classification results for the plurality of data obtained by inputting each of the plurality of data into a machine learning model; Based on the plurality of classification results, among a plurality of attributes included in a first plurality of data classified into a first group and a second plurality of data classified into a second group among the plurality of data, a first plurality of attributes whose difference in attribute values ​​between the first plurality of data and the second plurality of data satisfies a criterion; determining a label for each of the plurality of data based on a first index that combines the first plurality of attributes; training the machine learning model based on the determined labels and the plurality of data; A machine learning program that lets a computer perform processing.

[0128] (Appendix 2) the process of determining the label includes a process of determining the label based on the first index represented by an arithmetic operation between the first plurality of attributes; The machine learning program described in Appendix 1.

[0129] (Appendix 3) Accepting a selection of one or more of the first metrics from a plurality of the first metrics; causing the computer to execute a process; The process of determining the label includes a process of determining the label based on the one or more first indicators for which the selection is accepted. 1. The machine learning program according to claim 1 or 2.

[0130] (Appendix 4) acquiring a plurality of classification results for the plurality of data obtained by inputting each of the plurality of data into a machine learning model; Based on the plurality of classification results, among a plurality of attributes included in a first plurality of data classified into a first group and a second plurality of data classified into a second group among the plurality of data, a first plurality of attributes whose difference in attribute values ​​between the first plurality of data and the second plurality of data satisfies a criterion; determining a label for each of the plurality of data based on a first index that combines the first plurality of attributes; training the machine learning model based on the determined labels and the plurality of data; A machine learning method in which processing is performed by a computer.

[0131] (Appendix 5) the process of determining the label includes a process of determining the label based on the first index represented by an arithmetic operation between the first plurality of attributes; The machine learning method described in Appendix 4.

[0132] (Appendix 6) Accepting a selection of one or more of the first metrics from a plurality of the first metrics; causing the computer to execute a process; The process of determining the label includes a process of determining the label based on the one or more first indicators for which the selection is accepted. 6. The machine learning method according to claim 4 or 5.

[0133] (Appendix 7) acquiring a plurality of classification results for the plurality of data obtained by inputting each of the plurality of data into a machine learning model; Based on the plurality of classification results, among a plurality of attributes included in a first plurality of data classified into a first group and a second plurality of data classified into a second group among the plurality of data, a first plurality of attributes whose difference in attribute values ​​between the first plurality of data and the second plurality of data satisfies a criterion; determining a label for each of the plurality of data based on a first index that combines the first plurality of attributes; training the machine learning model based on the determined labels and the plurality of data; An information processing device comprising a control unit.

[0134] (Appendix 8) The control unit In the process of determining the label, the label is determined based on the first index represented by an arithmetic operation between the first plurality of attributes. 8. The information processing device according to claim 7.

[0135] (Appendix 9) The control unit Accepting a selection of one or more of the first metrics from a plurality of the first metrics; In the process of determining the label, the label is determined based on the one or more first indicators for which the selection is accepted. 9. The information processing device according to claim 7 or 8. [Explanation of symbols]

[0136] 1,6: AI system 2,7:User 11: CPU 12: Memory section 13: Display control section 14:Storage device 16: External recording medium processing unit 111: Classification result acquisition unit 112: Attribute identification part 113: Label determination unit 114: Training Department 115: Control unit 131:Display device 151: Mouse 152: Keyboard 160: Recording media 15: Input IF 17: Communication IF

Claims

1. acquiring a plurality of classification results for the plurality of data obtained by inputting each of the plurality of data into a machine learning model; Based on the plurality of classification results, among a plurality of attributes included in a first plurality of data classified into a first group and a second plurality of data classified into a second group among the plurality of data, a first plurality of attributes whose difference in attribute values ​​between the first plurality of data and the second plurality of data satisfies a criterion; determining a label for each of the plurality of data based on a first index that combines the first plurality of attributes; training the machine learning model based on the determined labels and the plurality of data; A machine learning program that lets a computer perform processing.

2. the process of determining the label includes a process of determining the label based on the first index represented by an arithmetic operation between the first plurality of attributes; The machine learning program according to claim 1 .

3. Accepting a selection of one or more of the first metrics from the plurality of first metrics; causing the computer to execute a process; The process of determining the label includes a process of determining the label based on the one or more first indicators selected and accepted. The machine learning program according to claim 1 or 2.

4. acquiring a plurality of classification results for the plurality of data obtained by inputting each of the plurality of data into a machine learning model; Based on the plurality of classification results, among a plurality of attributes included in a first plurality of data classified into a first group and a second plurality of data classified into a second group among the plurality of data, a first plurality of attributes whose difference in attribute values ​​between the first plurality of data and the second plurality of data satisfies a criterion; determining a label for each of the plurality of data based on a first index that combines the first plurality of attributes; training the machine learning model based on the determined labels and the plurality of data; A machine learning method in which processing is performed by a computer.

5. acquiring a plurality of classification results for the plurality of data obtained by inputting each of the plurality of data into a machine learning model; Based on the plurality of classification results, among a plurality of attributes included in a first plurality of data classified into a first group and a second plurality of data classified into a second group among the plurality of data, a first plurality of attributes whose difference in attribute values ​​between the first plurality of data and the second plurality of data satisfies a criterion; determining a label for each of the plurality of data based on a first index that combines the first plurality of attributes; training the machine learning model based on the determined labels and the plurality of data; An information processing device comprising a control unit.

Images