WO2023161993A1

WO2023161993A1 - Training data generation program, training data generation method, and information processing device

Info

Publication number: WO2023161993A1
Application number: PCT/JP2022/007230
Authority: WO
Inventors: 啓介後藤
Original assignee: 富士通株式会社
Priority date: 2022-02-22
Filing date: 2022-02-22
Publication date: 2023-08-31

Abstract

This invention suppresses a reduction in interpretation characteristics of a bias correction with respect to training data.　From among a plurality of data pieces each having a plurality of attributes, an information processing device (10) identifies a combination of values for one or more first attributes having a data bias of a threshold value or greater, on the basis of the number of data pieces corresponding to a combination of attribute values. In addition, the information processing device (10) selects an identified attribute value in accordance with the number of combinations in which the value of each attribute is included, among the combinations of values of the one or more first attributes. The information processing device (10) generates training data (2) by changing the attribute values of one or more pieces of data among the plurality of data pieces in accordance with the condition that the data bias is less than the threshold value, where the data bias relates to a combination of values for one or more second attributes including a specific attribute value, among the combinations of the one or more first attribute value combinations.

Description

TRAINING DATA GENERATION PROGRAM, TRAINING DATA GENERATION METHOD, AND INFORMATION PROCESSING APPARATUS

The present invention relates to a training data generation program, a training data generation method, and an information processing device.

　There is a possibility that bias exists in training data in machine learning. Bias in training data causes deterioration of learning accuracy. Inference using a model trained using biased training data may result in unfair inference results.

For example, machine learning may be used to determine whether or not a person can be hired (whether or not he or she is worthy of being hired) according to the attributes of an applicant for a job, based on training data on hiring performance by companies. In this case, the training data includes, for example, an attribute of the applicant such as gender, and a class label indicating whether or not the applicant was determined to be employable. In this training data, when comparing multiple applicants with the same attributes other than gender, the rate of being hired was significantly higher for males than for non-males. shall be Such training data has a bias that if the gender is male, the probability of being hired is very high.

When machine learning is performed using biased training data, a learning result (machine learning model) that reflects the bias is obtained. For example, if the gender is male, a machine learning model is generated that is determined to be employable with a higher probability than other persons. However, gender is not an indicator of job performance. Therefore, the use of such machine learning models leads to gender-unfair results and is not appropriate.

Several techniques have been proposed to remove bias from training data. For example, there has been proposed a learning data generation device that suppresses bias in the attributes of learning data. In addition, an information processing apparatus has been proposed that reduces the influence of the bias of information that is the target of inference in learning data during machine learning of an inference device, and also reduces the influence of the bias of other information related to the information. ing. Systems have also been proposed to detect and mitigate bias in intelligent virtual assistants. Devices have also been proposed that can prevent unethical behavior related to training data.

JP 2018-106216 A JP 2019-45929 A U.S. Patent Application Publication No. 2020/0143794 U.S. Patent Application Publication No. 2019/0147371

When bias correction processing is applied to training data, it is desirable to be able to clearly explain the reason for bias correction. Hereinafter, the ease of explaining the reason for bias correction will be referred to as the interpretability of bias correction.

If the training data contains various biases, removing all the biases will result in various changes to the training data. A wide variety of changes in the training data impairs the interpretability of the bias correction as a whole.

In one aspect, this case aims to prevent the deterioration of the interpretability of bias correction for training data.

In one proposal, a training data generation program is provided that causes a computer to perform the following processes.
Based on the number of data corresponding to a combination of attribute values, among a plurality of data each having a plurality of attributes, the computer determines the first one or more attribute values having a data bias equal to or greater than a threshold value. Identify combinations. The computer selects a particular attribute value according to the number of times each attribute value is included in the first one or more attribute value combinations. Then, the computer selects a plurality of attribute values according to the condition that the data bias is less than a threshold with respect to a second combination of one or more attribute values that includes a specific attribute value among the first combination of one or more attribute values. Training data is generated by changing the value of one or more data attributes of the data.

According to one aspect, it is possible to suppress deterioration in interpretability of bias correction for training data.
The above and other objects, features and advantages of the present invention will become apparent from the following description taken in conjunction with the accompanying drawings which represent exemplary preferred embodiments of the invention.

It is a figure which shows an example of the training data generation method which concerns on 1st Embodiment. It is a figure which shows an example of the hardware of the computer used for 2nd Embodiment. 1 is a block diagram showing functions that a computer has for machine learning; FIG. It is a figure which shows an example of input data. FIG. 11 is a diagram showing an example of generation of change rules; It is a figure which shows an example of bias correction. It is a figure which shows an example of the input data change which impairs interpretability. FIG. 10 is a diagram showing an example of input data change that reduces the bias correction effect; It is a figure which shows an example of the procedure of a bias correction process. FIG. 11 is a flow chart showing an example of a procedure of change rule generation processing; FIG. FIG. 10 is a diagram showing an example of a change rule list; FIG. 9 is a flow chart showing an example of a procedure of sensitive attribute selection processing; FIG. 10 is a diagram showing an example of an applied change rule list; FIG. 11 is a flowchart showing an example of a procedure of sensitive attribute change processing; FIG. It is a figure which shows an example of the training data produced|generated. It is a figure which shows an example of an interpretability management table. FIG. 10 is a flowchart showing an example of a procedure of sensitive attribute selection processing according to importance; FIG. It is a figure which shows an example of sensitive attribute selection according to importance.

Hereinafter, this embodiment will be described with reference to the drawings. It should be noted that each embodiment can be implemented by combining a plurality of embodiments within a consistent range.
[First embodiment]
The first embodiment is a training data generation method capable of suppressing deterioration in interpretability of bias correction for training data in machine learning.

FIG. 1 is a diagram showing an example of a training data generation method according to the first embodiment. FIG. 1 shows an information processing device 10 for implementing the training data generation method. The information processing device 10 can implement the training data generation method by executing, for example, a training data generation program.

The information processing device 10 has a storage unit 11 and a processing unit 12 . The storage unit 11 is, for example, a memory or a storage device that the information processing device 10 has. The processing unit 12 is, for example, a processor or an arithmetic circuit included in the information processing device 10 .

The storage unit 11 stores the input data 1. Input data 1 is data used for supervised learning in machine learning. Input data 1 includes a plurality of data (records) each having a plurality of attributes. The plurality of attributes are, for example, a first attribute value (A1=a, A2=b) that causes an unfair inference result to be output in machine learning, and a second attribute value that does not cause such a factor. It is divided into values (B1=c, B2=d, . . . ). The values of the first attribute are, for example, gender=male, race=white, and the like. In addition, each of the plurality of data is given a label (C) value indicating the correct answer when learning the object (person, animal, object, phenomenon, etc.) represented by the data.

In the example of FIG. 1, a flag of "1" or "0" indicates whether the value of each attribute is a predetermined value. A flag "1" indicates that the value of the attribute is a predetermined value (for example, "male" for the attribute "gender"). A flag "0" indicates that the value of the attribute is a value other than a predetermined value (for example, "female" for the attribute "sex").

Input data 1 may contain bias. For example, for data with the same value of the second attribute, the difference in the value of the first attribute causes an excessive difference in the appearance frequency of the value of the assigned label (e.g., the ratio of values "1"). If so, there is a bias for the value of that first attribute. If the input data 1 is data about whether or not a job can be hired (label value), if the rate of hiring is excessively higher when the gender is male than when the gender is female, there is a bias for "gender = male" There will be

The processing unit 12 generates training data 2 in which the bias contained in the input data 1 is corrected. For example, the processing unit 12 selects a first attribute having a data bias equal to or greater than a threshold value σ based on the number of data corresponding to a combination of attribute values among a plurality of data included in the input data 1. Identify combinations of values for . A combination of values of the first one or more attributes is, for example, a combination of a value of the first attribute and a value of the second attribute.

For example, the processing unit 12 calculates an index (bias value) indicating data bias for each combination of attribute values. Here, the fact that the value of each attribute or label is a predetermined value (the flag "1" is set) is expressed as A=1, B=1, C=1 (A is A1, A2 and B is any one of B1, B2, . . . ). Consider a case where the combination of attribute values is a combination of one first attribute value (A=1) and one second attribute value (B=1). The bias value in this case is, for example, "(ratio of C=1 data out of A=1, B=1 data)/(ratio of C=1 data out of B=1 data)". Desired.

If the bias value calculated for the combination of attribute values is equal to or greater than the threshold σ, the processing unit 12 determines that the combination of attribute values has a bias equal to or greater than the threshold σ. In the example of FIG. 1, among combinations of the first attribute value “A1=a” and the plurality of second attribute values (B1=c, B2=d, . . . ), There are 10 biased combinations. Further, among the combinations of the first attribute value “A2=d” and each of the plurality of second attribute values (B1=c, B2=d, . One set exists.

Furthermore, the processing unit 12 selects a specific attribute value according to the number of values of each attribute included in the first combination of one or more attribute values. A subset of the attribute values included in the first set of attribute value(s) is selected as the value of the particular attribute.

For example, the processing unit 12 selects at least one attribute value in descending order of the number included in the first combination of one or more attribute values having a data bias equal to or greater than the threshold value σ, and selects the value of the specific attribute Select as At this time, the processing unit 12 selects a specific attribute value from, for example, the first attribute values (A1=a, A2=b). In the example of FIG. 1, 10 first attribute values "A1=a" are included in the first combination of one or more attribute values. In addition, one first attribute value “A2=b” is included in the first combination of one or more attribute values. In this case, the processing unit 12 selects "A1=a" as the value of the specific attribute.

Then, the processing unit 12 generates training data 2 according to predetermined conditions. The predetermined condition is a condition that the data bias is less than the threshold value σ with respect to the second combination of one or more attribute values that includes the specific attribute value among the first combination of one or more attribute values. is. The processing unit 12 generates training data 2 that satisfies a predetermined condition by changing attribute values of one or more data out of the plurality of data.

In the training data 2 generated in this way, the data bias is less than the threshold σ with respect to the combination of the values of the second one or more attributes including the value of the specific attribute, and the value of the specific attribute is Biases in value combinations of the second attribute or attributes are corrected. Moreover, the values that are changed for bias correction are limited to the values of specific attributes (e.g. "A1=a") and the values set in common attributes (e.g. "￢A1=a"). The deterioration of interpretability is also suppressed.

In addition, the processing unit 12, for example, among combinations of a plurality of first attribute values (A1=a, A2=b) and second attribute values (B1=c, B2=d, . . . ) , a combination of values of the first one or more attributes having a data bias equal to or greater than the threshold value σ is identified. Accordingly, the combination of the values of the first one or more attributes includes the value of the first attribute that causes an unfair inference result to be output in machine learning. As a result, it is possible to reliably correct the attribute-related bias that causes an unfair inference result to be output.

The processing unit 12 also selects a specific attribute value from among the plurality of first attribute values. Accordingly, it is possible to appropriately prevent an unfair inference result from being output by a model obtained by performing machine learning using the generated training data 2 .

In generating the training data 2, the processing unit 12 generates, for example, change rules corresponding to each combination of values of one or more second attributes. In the change rule, a value different from the value of the specific attribute is set in the common attribute with the value of the specific attribute, and the value of the second attribute indicated in the combination of the values of the second one or more attributes Data in which a value is set and a predetermined label is assigned is specified as a change target. The change rule specifies that the second value of a specific attribute in the data to be changed should be changed to the first value.

For example, assume that the combination of the first attribute value "A1=a" and the second attribute value "B1=c" is a combination of the values of one or more second attributes. A change rule corresponding to this combination can be expressed as "(A1=a, B1=c, C=z)→(A1=a)" (a symbol indicating negation). “￢A1=a, B1=c, ￢C=z” is the condition of the data to be changed. Data that satisfies this condition is subject to change. The change rule specifies that for data whose common attribute with the value of a specific attribute is another value "￢A1=a", the other value should be changed to the value of the specific attribute "A1=a". . By performing such a change, data having a specific attribute value of “A1=a” but having a label value of “￢C=z” are increased. As a result, it is possible to correct the bias that the rate of the label value "C=z" being too high when having the specific attribute value "A1=a".

Note that when selecting a value of a specific attribute, the processing unit 12 selects at least one attribute value from the largest number included in the first combination of one or more attribute values, for example, as the specific attribute value. Select as the value of This allows more biases to be corrected by changing the value for one attribute. That is, a high bias correction effect can be obtained by changing the value of one type of attribute.

For each attribute, there may be differences in the ease of explanation (interpretability) when the bias for that attribute is corrected. In that case, the processing unit 12 may select a specific attribute value in consideration of the interpretability of each attribute. For example, in the process of selecting a specific attribute value, the processing unit 12 selects the first one or more attribute values for each attribute value included in the combination of the first one or more attribute values. Calculate the importance obtained by weighting the numbers contained in the combination of . Then, the processing unit 12 selects a specific attribute value according to the importance of each attribute value. For example, the processing unit 12 selects at least one attribute value in descending order of importance as a specific attribute value. As a result, biases related to attributes with higher interpretability are more likely to be corrected, and a decrease in interpretability due to correcting biases is suppressed.

[Second embodiment]
In the second embodiment, in a computer that performs machine learning, bias correcting processing is performed as preprocessing of training data used for machine learning.

FIG. 2 is a diagram showing an example of computer hardware used in the second embodiment. A computer 100 is entirely controlled by a processor 101 . A memory 102 and a plurality of peripheral devices are connected to the processor 101 via a bus 109 . Processor 101 may be a multiprocessor. The processor 101 is, for example, a CPU (Central Processing Unit), an MPU (Micro Processing Unit), or a DSP (Digital Signal Processor). At least part of the functions realized by processor 101 executing a program may be realized by an electronic circuit such as ASIC (Application Specific Integrated Circuit) or PLD (Programmable Logic Device).

The memory 102 is used as the main storage device of the computer 100. The memory 102 temporarily stores at least part of an OS (Operating System) program and application programs to be executed by the processor 101 . In addition, the memory 102 stores various data used for processing by the processor 101 . As the memory 102, for example, a volatile semiconductor memory device such as RAM (Random Access Memory) is used.

Peripheral devices connected to the bus 109 include a storage device 103 , a GPU (Graphics Processing Unit) 104 , an input interface 105 , an optical drive device 106 , a device connection interface 107 and a network interface 108 .

The storage device 103 electrically or magnetically writes data to and reads data from a built-in recording medium. A storage device 103 is used as an auxiliary storage device for the computer 100 . The storage device 103 stores an OS program, application programs, and various data. As the storage device 103, for example, an HDD (Hard Disk Drive) or an SSD (Solid State Drive) can be used.

The GPU 104 is an arithmetic unit that performs image processing, and is also called a graphics controller. A monitor 21 is connected to the GPU 104 . The GPU 104 displays an image on the screen of the monitor 21 according to instructions from the processor 101 . Examples of the monitor 21 include a display device using an organic EL (Electro Luminescence), a liquid crystal display device, and the like.

A keyboard 22 and a mouse 23 are connected to the input interface 105 . The input interface 105 transmits signals sent from the keyboard 22 and mouse 23 to the processor 101 . Note that the mouse 23 is an example of a pointing device, and other pointing devices can also be used. Other pointing devices include touch panels, tablets, touchpads, trackballs, and the like.

The optical drive device 106 reads data recorded on the optical disc 24 or writes data to the optical disc 24 using laser light or the like. The optical disc 24 is a portable recording medium on which data is recorded so as to be readable by light reflection. The optical disc 24 includes DVD (Digital Versatile Disc), DVD-RAM, CD-ROM (Compact Disc Read Only Memory), CD-R (Recordable)/RW (ReWritable), and the like.

The device connection interface 107 is a communication interface for connecting peripheral devices to the computer 100 . For example, the device connection interface 107 can be connected to the memory device 25 and the memory reader/writer 26 . The memory device 25 is a recording medium equipped with a communication function with the device connection interface 107 . The memory reader/writer 26 is a device that writes data to the memory card 27 or reads data from the memory card 27 . The memory card 27 is a card-type recording medium.

The network interface 108 is connected to the network 20. Network interface 108 transmits and receives data to and from other computers or communication devices via network 20 . The network interface 108 is a wired communication interface that is connected by a cable to a wired communication device such as a switch or router. Also, the network interface 108 may be a wireless communication interface that communicates with a wireless communication device such as a base station or an access point via radio waves.

The computer 100 can implement the processing functions of the second embodiment with the above hardware. The information processing apparatus 10 shown in the first embodiment can also be realized by hardware similar to the computer 100 shown in FIG.

The computer 100 implements the processing functions of the second embodiment, for example, by executing a program recorded on a computer-readable recording medium. A program describing the processing content to be executed by the computer 100 can be recorded in various recording media. For example, a program to be executed by the computer 100 can be stored in the storage device 103 . The processor 101 loads at least part of the program in the storage device 103 into the memory 102 and executes the program. The program to be executed by the computer 100 can also be recorded in a portable recording medium such as the optical disc 24, memory device 25, memory card 27, or the like. A program stored in a portable recording medium can be executed after being installed in the storage device 103 under the control of the processor 101, for example. Alternatively, the processor 101 can read and execute the program directly from the portable recording medium.

Machine learning using training data can be performed using the computer 100 having such hardware.
FIG. 3 is a block diagram showing functions that a computer has for machine learning. Computer 100 has storage unit 110 , change rule generation unit 120 , sensitive attribute selection unit 130 , sensitive attribute change unit 140 and machine learning unit 150 .

The storage unit 110 stores input data 111 and training data 112 . For example, the input data 111 is a data set prepared for supervised learning. Input data 111 includes a plurality of data. Each piece of data is set with values for sensitive attributes, non-sensitive attributes, and class labels. A sensitive attribute is an example of the first attribute in the first embodiment. A non-sensitive attribute is an example of the second attribute in the first embodiment. Input data 111 may contain multiple kinds of biases. The training data 112 is data obtained by removing some bias from the input data 111 . Training data 112 is the input to the learning phase of machine learning.

The change rule generation unit 120 generates bias change rules based on the input data 111 . For example, the change rule generation unit 120 calculates a numerical value that serves as an index for bias determination using a predetermined formula for the sensitive attribute included in the input data 111, and if the calculated numerical value is equal to or greater than a predetermined value, the sensitive attribute value Generate change rules that change the .

The sensitive attribute selection unit 130 selects a sensitive attribute to be changed from among the generated change rules. For example, the sensitive attribute selection unit 130 selects the sensitive attribute included in the largest number of change rules.

The sensitive attribute changing unit 140 changes the value of the selected sensitive attribute so that the bias is corrected. For example, the sensitive attribute changing unit 140 changes the value of the selected sensitive attribute for one or more pieces of data in the input data 111 so that the numerical value used as the index for bias determination is less than the threshold. Then, the sensitive attribute changing unit 140 stores the changed input data 111 as the training data 112 in the storage unit 110 .

The machine learning unit 150 performs machine learning using the training data 112. For example, the machine learning unit 150 receives the training data 112 and generates an inference model. Thereafter, when data to be inferred is input, the machine learning unit 150 performs inference processing using the generated model. For example, when a model is generated for judging whether or not an applicant can be hired for recruitment, the machine learning unit 150 uses the attributes of the applicant as input to make inferences using the model, and determines whether the applicant is worthy of being hired. output the inference result.

It should be noted that the lines connecting each element shown in FIG. 3 indicate part of the communication paths, and communication paths other than the illustrated communication paths can also be set. Also, the function of each element shown in FIG. 3 can be realized by causing a computer to execute a program module corresponding to the element, for example.

Next, the input data 111 will be specifically described.
FIG. 4 is a diagram showing an example of input data. The input data 111 shown in FIG. 4 is data indicating the result of determination of whether or not to hire an applicant for recruiting. In the data (record) for each applicant registered in the input data 111, values are set for multiple sensitive attributes, multiple non-sensitive attributes, and class labels.

In the input data 111, the value in the data corresponding to each attribute or class label is indicated by a "1" or "0" flag. When the flag of each attribute or class label of each data is "1", it indicates that the value of that attribute or class label is a predetermined value (for example, the value of the attribute "gender" is "male"). When the flag of each attribute or class label of each data is "0", it indicates that the value of that attribute or class label is other than the predetermined value (for example, the value of the attribute "gender" is other than "male"). .

In FIG. 4, each attribute and class label name area of the input data 111 shows the conditions under which the area or class label flag is "1". For example, if "sex=male" (the character string to the left of the symbol is the attribute name), a flag "1" is set for the data whose value of the sensitive attribute "sex" is "male".

Sensitive attributes include "gender" and "race". In the column of the sensitive attribute "gender", a flag "1" is set for data of male applicants (sex=male), and a flag "0" is set for data other than male applicants. In the column of the sensitive attribute “race”, a flag “1” is set if the applicant is Caucasian (race=Caucasian), and a flag “0” is set if the applicant is not Caucasian.

Non-sensitive attributes include "hometown", "annual income", and "age". In the non-sensitive attribute "hometown" column, for example, if the applicant is from the Kanto region (hometown=Kanto), a flag "1" is set, and if the applicant is not from the Kanto region, a flag "0" is set. In the non-sensitive attribute "annual income" column, for example, if the applicant's annual income exceeds 1.3 million (annual income > 1.3 million), a flag "1" is set, and if the annual income does not exceed 1.3 million, a flag "0" is set. be. In the non-sensitive attribute "age" column, a flag "1" is set if the age of the applicant is over 20 (age>20), and a flag "0" is set if the age is not over 20.

The class label is "recruitment". In the column of the class label "employment", a flag "1" is set when the applicant is worthy of being hired (employment=possible), and a flag "0" is set when not being worthy of being hired. The value of the class label indicates the result determined by the hiring manager based on the results of interviews, practical tests, and the like.

It should be noted that which attribute in the input data 111 is set as the sensitive attribute is preset by the user. In the example of FIG. 4, "sex" and "race" are designated as sensitive attributes in order to correct biases due to sex discrimination and racial discrimination. For example, if discrimination based on place of birth is rampant and correction of bias due to place of birth discrimination is required, "hometown" may be designated as a sensitive attribute.

Assume that the input data 111 shown in FIG. 4 has biases in both the sensitive attributes "gender" and "race". For example, out of all the data, the ratio of "male" (sex = male) for the sensitive attribute "sex" and "acceptable" (adopted = acceptable) for the class label "employment" is "2/3". shall be It is also assumed that the value of the sensitive attribute "sex" is other than male (￢sex = male) and the rate of "acceptable" (acceptable = acceptable) of the class label "employment" is "1/3". . In this case, the probability of “employment = acceptable” varies greatly depending on gender, and it is considered that there is a bias related to the sensitive attribute “gender”.

In this way, the input data 111 from which the training data 112 is generated may have a bias. In that case, it is inappropriate to directly use the training data 112 for learning by machine learning. For example, a model trained with biased input data 111 may exhibit discriminatory behavior.

The computer 100 then generates the training data 112 based on change rules that change the values of the sensitive attributes in the input data 111 so as to correct the bias. When a bias value calculated for a set of a predetermined value of one sensitive attribute and a predetermined value of one insensitive attribute is equal to or greater than a predetermined threshold value, a change rule is generated in association with the set. For example, let A be the sensitive attribute and B be the non-sensitive attribute of the pair to be subjected to bias calculation. Let C be the class label in the input data 111 . Further, let σ be the threshold value of the bias value. At this time, if the bias value “elift (A, B→C)” of the selected set is equal to or greater than the threshold value σ, it is determined that the selected set is biased. The formula is as follows.

sup(x, y, . . . ) is the number of data in the input data with the characteristics x=1, y=1, . x=1, y=1 indicates that the values of attributes or class labels set to x and y are predetermined values (flag "1"). For example, sup(A, B) is the number of data in which the value of the sensitive attribute of the set to be calculated is a predetermined value and the value of the non-sensitive attribute is a predetermined value. sup(A, B, C) is the number of data in which the value of the sensitive attribute of the set to be calculated is a predetermined value, the value of the non-sensitive attribute is a predetermined value, and the value of the class label is a predetermined value. be. sup(B) is the number of data in which the value of the non-sensitive attribute of the set to be calculated is the predetermined value. sup(A, C) is the number of data in which the value of the sensitive attribute of the set to be calculated is a predetermined value and the value of the class label is a predetermined value.

Formula (2) indicates the ratio of data with a predetermined value for the class label among the data for which both the values of the sensitive attribute and the non-sensitive attribute are predetermined values. Equation (3) indicates the ratio of data with a predetermined value of the class label to data with a predetermined value of the non-sensitive attribute.

The bias value shown on the left side of Equation (1) indicates how much the probability that the value of the class label will be the predetermined value changes depending on whether or not the sensitive attribute is taken into consideration. For example, the ratio of the class label value when the sensitive attribute is taken into consideration (the value obtained by the formula (2)) is the ratio of the class label when the sensitive attribute is not taken into account (the value obtained by the formula (3)). The larger the value obtained), the larger the bias value. Then, if the bias value for a pair of sensitive attributes and non-sensitive attributes is greater than or equal to the threshold σ, a change rule for that pair is generated.

FIG. 5 is a diagram showing an example of generation of change rules. For example, it is assumed that the bias value is calculated for a set of the value "male" of the sensitive attribute "gender" and the value "20" of the non-sensitive attribute "age". The input data 41 includes four data as shown in FIG. The threshold value σ of the bias value is "1.2".

In the example of FIG. 5, "conf (gender = male, age = 20→employment = acceptable) = 2/2 = 1". "conf (age=20→employment=possible)=3/4". Then, "elift (sex=male, age=20→recruitment=possible)=4/3=1.3 .gtoreq..sigma.=1.2". That is, the bias value becomes equal to or greater than the threshold value σ.

Therefore, a change rule "(￢ sex = male, age = 20, ￢ adoption = acceptable) → (sex = male)" is generated to correct the bias regarding the value "male" of the sensitive attribute "age". This change rule applies to data whose sensitive attribute "gender" has a value other than "male," non-sensitive attribute "age" has a value of "20," and class label "employment" has a value of "impossible." ” is changed to “male” (flag “1”). Bias is corrected by changing the input data 41 using such a change rule.

FIG. 6 is a diagram showing an example of bias correction. In the example of FIG. 6, the fourth data of the input data 41 matches the conditions of the data to be changed (￢sex=male, age=20, ￢adoption=permissible) in the change rule. Therefore, the change rule is applied to the corresponding data, and the value of the sensitive attribute "gender" is changed to "male" (flag is changed from "0" to "1"). The training data 42 is data in which the value of the sensitive attribute has been changed.

In the training data 42, "conf (gender = male, age = 20→employment = acceptable) = 2/3". "conf (age=20→employment=possible)=3/4". Then, "elift (sex=male, age=20→recruitment=possible)=8/9=0.88...<σ=1.2". That is, the bias value becomes less than the threshold value σ. Note that the change of the value of the sensitive attribute of data according to the change rule is repeatedly executed until the bias value becomes less than the threshold.

Here, the input data may contain various biases. At that time, applying change rules that correspond to all biases impairs the interpretability of bias correction.

FIG. 7 is a diagram showing an example of input data change that impairs interpretability. For example, it is assumed that the sensitive attributes “sex=male” and “race=white” in the input data 111 both cause bias. In that case, a change rule is generated to change the value of the sensitive attribute "gender" to "male" for the data whose value of the sensitive attribute "gender" is other than "male" and whose value of the class label "employment" is "impossible". (conditions for non-sensitive attributes are omitted). Similarly, for data whose sensitive attribute "race" has a value other than "white" and whose class label "employment" has a value of "impossible", a change rule is generated to change the value of the sensitive attribute "race" to "white". (conditions for non-sensitive attributes omitted).

The training data 112a generated by applying all generated change rules to the input data 111 has changed values for many sensitive attributes. Changing the values of a large number of sensitive attributes in this manner removes many of the biases in the input data 111, but reduces the ease of explanation of the bias correction process. That is, the interpretability of the bias correction process is lowered.

In this way, the more the number of types of sensitive attributes changed by bias correction processing, the lower the interpretability. In machine learning, not only is prediction accuracy by inference high, but a clear explanation to the user as to why such a prediction was made is also required. Similar to the interpretability of such prediction results, even if bias correction is performed on the input data during learning, clearly indicate the reason for the change and the impact on the prediction result (interpretability of bias correction) ) is required. Therefore, it is necessary to correct as many biases as possible without impairing interpretability.

Therefore, it is conceivable to limit the sensitive attributes to be changed for bias correction processing to one. If only one sensitive attribute is changed, deterioration in interpretability of bias correction processing is suppressed. However, in this case, if the sensitive attribute to be changed is selected incorrectly, the bias correcting effect will be weak.

FIG. 8 is a diagram showing an example of input data change that reduces the bias correction effect. In the example of FIG. 8, a change rule for changing the value of the sensitive attribute "gender" to "male" for data whose value of the sensitive attribute "gender" is other than "male" and whose value of the class label "employment" is "impossible" is generated 10 times. The conditions (B1, B2, . . . , B10) for non-sensitive attributes in each change rule are different. In addition, there is one change rule that changes the value of the sensitive attribute "race" to "white" for data whose value of the sensitive attribute "race" is other than "white" and whose class label "recruitment" value is "impossible". generated.

At this time, in order not to impair interpretability, the sensitive attribute to be changed is limited to "race". In this case, training data 112b is generated in which the value of the sensitive attribute "race" is changed from other than "white" to "white" for some data.

When such bias correction processing is performed, only one bias among many biases is corrected. Therefore, the effect of bias correction is small.
Therefore, in the computer 100 according to the second embodiment, the sensitive attribute selection unit 130 selects the value of the sensitive attribute with the highest appearance frequency among the generated change rules as a change target in the bias correcting process. This makes it possible to correct more biases by changing the value of one type of sensitive attribute. In other words, a large bias correcting effect can be obtained by changing the value of a small number of sensitive attributes.

FIG. 9 is a diagram illustrating an example of a procedure for bias correction processing. The processing shown in FIG. 9 will be described below along with the step numbers.
[Step S101 ] The change rule generation unit 120 reads the input data 111 from the storage unit 110 .

[Step S102 ] Based on the input data 111 , the change rule generation unit 120 generates a bias change rule included in the input data 111 . One or more change rules are generated. The change rule indicates the set of sensitive and non-sensitive attributes that cause the bias. The value of the sensitive attribute indicated in the change rule is subject to change when correcting the bias. Details of the change rule generation process will be described later (see FIG. 10).

[Step S103] The sensitive attribute selection unit 130 selects a sensitive attribute to be changed for bias correction from among the sensitive attributes included in any of the generated change rules. Details of the sensitive attribute selection process will be described later (see FIG. 12).

[Step S104] The sensitive attribute changing unit 140 changes the value of the selected sensitive attribute in the input data 111 based on the change rule including the sensitive attribute. Details of the sensitive attribute change processing will be described later (see FIG. 14).

[Step S105 ] The sensitive attribute changing unit 140 stores the input data 111 with changed sensitive attribute values in the storage unit 110 as the training data 112 .
Training data 112 is thus generated based on the input data 111 . Next, change rule generation processing will be described in detail.

FIG. 10 is a flowchart illustrating an example of the procedure of change rule generation processing. The processing shown in FIG. 10 will be described below along with the step numbers.
[Step S111] The change rule generation unit 120 selects one unselected set from the sets that can be generated with one sensitive attribute and one non-sensitive attribute.

[Step S112] The change rule generation unit 120 calculates a bias value for the selected pair. For example, the change rule generator 120 calculates the bias value “elift (A, B→C)” shown in Equation (1).

[Step S113] The change rule generation unit 120 determines whether the calculated bias value is equal to or greater than the threshold value σ. If the bias value is equal to or greater than the threshold σ, change rule generation section 120 advances the process to step S114. Further, if the bias value is less than the threshold value σ, the change rule generation unit 120 advances the process to step S115.

[Step S114] The change rule generation unit 120 generates a change rule for the selected pair. For example, the change rule generation unit 120 determines that the value of the selected sensitive attribute is other than the predetermined value (flag "0"), the value of the selected insensitive attribute is the predetermined value (flag "1"), and the value of the class label is other than the predetermined value. A change rule is generated for the data of (flag "0") to be changed. The generated change rule indicates that the value of the sensitive attribute should be changed to a predetermined value (flag "1"). The change rule generation unit 120 registers the generated change rule in the change rule list.

[Step S115] The change rule generator 120 determines whether there is an unselected combination of a sensitive attribute and a non-sensitive attribute. If there is an unselected pair, change rule generation unit 120 advances the process to step S111. Further, if all pairs have been selected, the change rule generation unit 120 ends the process.

In this way, bias values are calculated for all pairs of sensitive attributes and non-sensitive attributes, and modification rules corresponding to pairs whose bias values are equal to or greater than the threshold σ are generated. That is, change rules for correcting each bias included in the input data 111 are generated. The generated change rule is shown, for example, in a change rule list.

FIG. 11 is a diagram showing an example of a change rule list. In the change rule list 51 shown in FIG. 11, ten change rules including the value "male" of the sensitive attribute "sex" are registered. In the change rule list 51, one change rule including the value "white" of the sensitive attribute "race" is registered.

Based on the change rule list 51, the sensitive attribute selection unit 130 selects a sensitive attribute to be applied to the change for bias correction.
FIG. 12 is a flowchart showing an example of the procedure of sensitive attribute selection processing. The processing shown in FIG. 12 will be described below along with the step numbers.

[Step S121 ] The sensitive attribute selection unit 130 selects one unselected sensitive attribute value included in one of the change rules from the change rule list 51 .
[Step S122 ] The sensitive attribute selection unit 130 counts the appearance frequency of the selected sensitive attribute in the change rule list 51 . In counting the appearance frequency, for example, if the value of the sensitive attribute after being changed by the change rule is the value of the selected sensitive attribute, one occurrence of the value of the sensitive attribute selected by the change rule is counted. For example, in the change rule “(￢sex=male, B1, ￢employment=possible)→(sex=male)”, one sensitive attribute value “sex=male” appears.

[Step S123] The sensitive attribute selection unit 130 determines whether there is an unselected sensitive attribute value in the change rule list 51 or not. If there is an unselected sensitive attribute value, the sensitive attribute selection unit 130 advances the process to step S121. If all sensitive attribute values have been selected, sensitive attribute selection section 130 advances the process to step S124.

[Step S124] The sensitive attribute selection unit 130 identifies the value of the sensitive attribute with the highest appearance frequency as the value of the sensitive attribute to be applied to the change for bias correction.

[Step S125] The sensitive attribute selection unit 130 extracts from the change rule list 51 a change rule that includes the value of the specified sensitive attribute. The sensitive attribute selection unit 130 registers, for example, the extracted change rule in the applicable change rule list.

In this way, the modification rule including the sensitive attribute with the highest frequency of appearance is extracted as an application target in the bias correction processing.
FIG. 13 is a diagram showing an example of an applied change rule list. For example, in the change rule list 51 shown in FIG. 11, the appearance frequency of the sensitive attribute "sex=male" is "10", and the appearance frequency of the sensitive attribute "race=white" is "1". Therefore, “sex=male” is specified as the sensitive attribute with the highest appearance frequency. A change rule including the sensitive attribute “sex=male” is registered in the applied change rule list 52 . Based on the application change rule list 52, the value of the sensitive attribute in the input data 111 is changed.

FIG. 14 is a flowchart illustrating an example of the procedure of sensitive attribute change processing. The processing shown in FIG. 14 will be described below according to the step numbers.
[Step S131] The sensitive attribute changing unit 140 selects an unselected change rule in the applied change rule list 52 as a change rule to be applied.

[Step S132] The sensitive attribute changing unit 140 selects data from the input data 111 that conforms to the selected change rule. For example, it is assumed that the selected change rule is "(￢sex=male, hometown=Kanto, ￢recruitment=possible)→(sex=male)". In this case, the sensitive attribute changing unit 140 selects data whose value of the sensitive attribute “sex” is other than “male”, whose value of the non-sensitive attribute “hometown” is “Kanto”, and whose value of the class label “employment” is “impossible”. to select.

[Step S133] The sensitive attribute changing unit 140 changes the value of the sensitive attribute of the selected data according to the selected change rule. For example, the sensitive attribute change unit 140 changes the value of the sensitive attribute indicated by the change rule in the selected data from a value other than the predetermined value (flag "0") to a predetermined value (flag "1").

[Step S134] The sensitive attribute changing unit 140 determines whether the bias value of the bias to be changed in the selected change rule is less than the threshold σ. If the bias value is less than the threshold value σ, the sensitive attribute changing unit 140 advances the process to step S135. Also, if the bias value is equal to or greater than the threshold σ, the sensitive attribute changing unit 140 advances the process to step S132.

[Step S135] The sensitive attribute change unit 140 determines whether or not there is an unselected change rule in the applied change rule list 52. If there is an unselected change rule, the sensitive attribute change unit 140 advances the process to step S131. Also, if all the change rules in the applied change rule list 52 have been selected, the sensitive attribute change unit 140 ends the sensitive attribute change process.

The input data 111 in which the value of the sensitive attribute has been changed in this manner is stored in the storage unit 110 as the training data 112 .
FIG. 15 is a diagram showing an example of generated training data. Similar to the input data 111, the training data 112 includes a plurality of data having sensitive attributes, non-sensitive attributes, and class label values. The values set for each data are the same as the values of the input data 111 except for the values changed by the sensitive attribute change process. In the example of FIG. 15, for at least some of the data whose value of the sensitive attribute “sex” was other than “male” (flag “0”), the value of the sensitive attribute “sex” was “male” (flag “1”). ) has been changed.

In this way, by applying change rules that change sensitive attributes that occur frequently in change rules, it is possible to apply more change rules by changing one type of sensitive attribute (high bias reduction effect). Moreover, since only one type of sensitive attribute is changed, interpretability is not impaired.

[Third Embodiment]
The third embodiment performs bias correction processing in consideration of the difference in interpretability for each sensitive attribute. Sensitive attributes include those that clearly cause unfairness when biased, and those that are difficult to assert as being unfair.

For example, if the decision result of adoption shown in the input data 111 shows that the adoption is largely different depending on the value "white" of the sensitive attribute "race", the bias of the sensitive attribute causes unfairness. can be easily explained. In other words, it can be said that the interpretability of the bias correction processing for changing the value "white" of the sensitive attribute "race" is high.

As another example, it is possible that in the result of determination of whether or not a person can be hired shown in the input data 111, whether or not the value of the sensitive attribute "sex" is "male" can greatly affect whether or not the person can be hired. At this time, if the work being recruited includes physical labor, there is a possibility that it cannot be said to cause unfairness. For example, in the case of a care worker, it is conceivable that a man will be required to assist the movement of the person requiring care. If the percentage of male applicants for such nursing care jobs is extremely low, the percentage of male applicants who are accepted for employment will be high. In such a case, it is difficult to simply explain whether the bias in the value "male" of the sensitive attribute "gender" should be interpreted as unfair. Therefore, it can be said that the interpretability of the bias change processing for the value "male" of the sensitive attribute "gender" is low.

Therefore, in the third embodiment, the user can set a value indicating the ease of interpretation for each sensitive attribute. For example, a higher value is set for the interpretability of a sensitive attribute that is more interpretable when changed for bias correction. The interpretability for each sensitive attribute is set, for example, in an interpretability management table.

FIG. 16 is a diagram showing an example of the interpretability management table. In the interpretability management table 61, values indicating interpretability are set in association with sensitive attribute values. In the example of FIG. 16, the ease of interpretation of the value "male" of the sensitive attribute "gender" is "w1=1". The interpretability of the value "white" of the sensitive attribute "race" is "w2=20".

When an interpretability value is set for each sensitive attribute value, the sensitive attribute selection unit 130 weights the appearance frequency of the sensitive attribute in the change rule list 51 according to the interpretability value. For example, the sensitive attribute selection unit 130 takes the result of multiplying the appearance frequency of the sensitive attribute by the value (weight value) of the easiness of interpretation of the sensitive attribute as the degree of importance. Then, the sensitive attribute selection unit 130 selects the sensitive attribute with the highest importance as the sensitive attribute to be changed.

FIG. 17 is a flow chart showing an example of the procedure of sensitive attribute selection processing according to importance. The processing shown in FIG. 17 will be described below along with the step numbers.
[Step S201 ] The sensitive attribute selection unit 130 selects one unselected sensitive attribute value included in one of the change rules from the change rule list 51 .

[Step S202] The sensitive attribute selection unit 130 counts the appearance frequency in the change rule list 51 of the value of the selected sensitive attribute.
[Step S203] The sensitive attribute selection unit 130 calculates the importance of the value of the selected sensitive attribute. For example, the sensitive attribute selection unit 130 acquires the interpretability value of the selected sensitive attribute value from the interpretability management table 61 . Then, the sensitive attribute selection unit 130 sets the result of multiplication of “appearance frequency×value of ease of interpretation” as the importance of the value of the selected sensitive attribute.

[Step S204] The sensitive attribute selection unit 130 determines whether there is an unselected sensitive attribute value in the change rule list 51 or not. If there is an unselected sensitive attribute value, the sensitive attribute selection unit 130 advances the process to step S201. If all sensitive attribute values have been selected, the sensitive attribute selection unit 130 advances the process to step S205.

[Step S205] The sensitive attribute selection unit 130 specifies the value of the sensitive attribute with the highest degree of importance as the value of the sensitive attribute to be applied to the change for correcting the bias.

[Step S206] The sensitive attribute selection unit 130 extracts from the change rule list 51 a change rule that includes the value of the specified sensitive attribute. The sensitive attribute selection unit 130 registers, for example, the extracted change rule in the applied change rule list 52 .

In this way, the modification rule including the sensitive attribute with the highest importance is extracted as an application target in the bias correction processing.
FIG. 18 is a diagram showing an example of sensitive attribute selection according to importance. For example, the appearance frequency of the sensitive attribute "sex=male" in the change rule list 51 is "10", and the value of the ease of interpretation is "1". In this case, the importance of the value "male" of the sensitive attribute "sex" is "10×1=10". On the other hand, in the change rule list 51, the appearance frequency of the value "white" of the sensitive attribute "race" is "1", and the ease of interpretation is "20". In this case, the importance of the value "white" of the sensitive attribute "race" is "1×20=20".

In the example of FIG. 18, the value of the sensitive attribute with the highest importance is "race=white". As a result, "race=white" is selected as the sensitive attribute to which the bias correction process is applied.

In this way, rather than changing the value "male" of the sensitive attribute "gender", changing the value "white" of the sensitive attribute "race" may make it easier for the user to understand why the bias has been corrected. be. In that case, the bias correcting process can be applied to the value "white" of the sensitive attribute "race". As a result, it is possible to prevent deterioration of interpretability caused by changing the value of the sensitive attribute, which is difficult to explain, in the bias correcting process.

[Other embodiments]
In the second and third embodiments, only one value of the sensitive attribute to which bias correction is applied is selected. to apply modification rules that include the values of those sensitive attributes. In this case, the sensitive attribute selection unit 130 selects, for example, a predetermined number of sensitive attribute values in descending order of appearance frequency in the change rule list 51 as sensitive attribute values to which bias correction processing is applied.

In addition, in the second and third embodiments, when there is a bias for "gender = male", changes such as "(￢ sex = male, ..., ￢ adoption = acceptable) → (sex = male)" Although the rule corrects the bias, it is also possible to correct the bias in other ways. For example, the change rule generation unit 120 may set the change rule as “(sex=male, . As a result, the ratio of the data whose value of the sensitive attribute "gender" is other than "male" to the value of "acceptable" for the class label "acceptable" increases, and the bias is corrected.

Note that the change rule generation unit 120 may generate a change rule for changing the class label. For example, when there is a bias for "sex=male", the change rule generation unit 120 may set the change rule as "(￢sex=male, . In this case, the ratio of data whose value of the sensitive attribute "gender" is other than "male" and whose value of the class label "acceptable" is "acceptable" increases, and the bias is corrected. Further, when there is a bias for “sex=male”, the change rule generation unit 120 may set the change rule as “(sex=male, . In this case, the ratio of the data whose value of the sensitive attribute "sex" is "male" and whose value of the class label "acceptable" is "acceptable" is reduced, and the bias is corrected.

The above merely shows the principle of the present invention. Furthermore, many variations and modifications will occur to those skilled in the art, and the present invention is not limited to the precise construction and applications shown and described above, and all corresponding variations and equivalents are and the equivalents thereof.

REFERENCE SIGNS LIST 1 input data 2 training data 10 information processing device 11 storage unit 12 processing unit

Claims

Identifying a first combination of one or more attribute values having a data bias equal to or greater than a threshold based on the number of data corresponding to a combination of attribute values among a plurality of data each having a plurality of attributes death,
Selecting a specific attribute value according to the number of values of each attribute included in the first combination of one or more attribute values;
According to the condition that data bias is less than the threshold with respect to a second combination of one or more attribute values including the specific attribute value among the first combination of one or more attribute values, the plurality of generating training data by changing the value of one or more data attributes of the data;
A training data generation program that causes a computer to execute processing.
In the process of identifying the combination of values of the first one or more attributes, the value of the first attribute that causes an unfair inference result to be output in machine learning, and the second attribute that does not become the factor Identifying a combination of values of the first one or more attributes with a data bias equal to or greater than the threshold from among the combinations of the values of
The training data generation program according to claim 1.
In the process of selecting the value of the specific attribute, the value of the specific attribute is selected from among the first attribute values included in the combination of the first one or more attribute values;
3. The training data generation program according to claim 2.
In the process of generating the training data, a value different from the value of the specific attribute is set to an attribute common to the value of the specific attribute, and a combination of values of the second one or more attributes is set with the value of the second attribute shown in and the other value in the data assigned a predetermined label is changed to the value of the specific attribute;
4. The training data generation program according to claim 3.
In the process of selecting the value of the specific attribute, at least one attribute value in descending order of the number included in the combination of the first one or more attribute values is selected as the value of the specific attribute. select,
A training data generation program according to any one of claims 1 to 4.
In the process of selecting the specific attribute value, for each attribute value included in the first combination of the one or more attribute values, the first combination of the one or more attribute values is selected. calculating the importance obtained by weighting the numbers contained therein, and selecting the value of the specific attribute according to the importance of each value of the attribute;
A training data generation program according to any one of claims 1 to 4.
Identifying a first combination of one or more attribute values having a data bias equal to or greater than a threshold based on the number of data corresponding to a combination of attribute values among a plurality of data each having a plurality of attributes death,
Selecting a specific attribute value according to the number of values of each attribute included in the first combination of one or more attribute values;
According to the condition that data bias is less than the threshold with respect to a second combination of one or more attribute values including the specific attribute value among the first combination of one or more attribute values, the plurality of generating training data by changing the value of one or more data attributes of the data;
A training data generation method in which processing is performed by a computer.
Identifying a first combination of one or more attribute values having a data bias equal to or greater than a threshold based on the number of data corresponding to a combination of attribute values among a plurality of data each having a plurality of attributes and selecting a specific attribute value according to the number of values of each attribute included in the combination of values of the first one or more attributes, and selecting the value of the first one or more attributes one or more of the plurality of data according to the condition that the data bias is less than the threshold with respect to a second combination of the values of one or more attributes that includes the value of the specific attribute among the combinations of a processing unit that generates training data by changing values;
Information processing device having