Classifications

• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N20/00—Machine learning
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
  • G06Q10/00—Administration; Management
  • G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
  • G06Q10/063—Operations research, analysis or management

Definitions

• the disclosed technology relates to a training data generation program, a training data generation method, and a training data generation device.
• the combination of multiple attributes may affect the prediction results of a machine learning model.
• a device acquires a plurality of data, each of which is labeled with a positive example or a negative example, and calculates the ratio of the number of positive example data to the number of negative example data for each of a plurality of types of pairs of a first attribute and a second attribute associated with each of the plurality of data.
• the device identifies candidate data for modification from among the data in which the first attribute corresponding to the first type is associated with a second attribute based on the ratio for each combination of the first type and each of all other types included in the plurality of types.
• the device also selects first data from among the data in which the first attribute corresponding to the first type is associated with a second attribute based on the candidate data for modification identified for each combination.
• the device then causes a computer to execute a process of generating machine learning data by modifying the label of the first data included in the plurality of data.
• the above conventional technologies aim to mitigate bias between groups for each attribute, that is, to eliminate bias in the rate of positive cases between groups.
• a machine learning model that performs face recognition based on a person's facial image had a low recognition rate for black women, but the recognition rate should be the same regardless of race or gender.
• it is desired to intentionally control the bias of a specific group For example, an IT company wants to prioritize hiring women in science and engineering fields with the goal of promoting the participation of women in the workforce. In this case, it is desired to control the bias towards women in science and engineering fields in the machine learning model that makes hiring decisions so that it reaches a target value.
• bias correction is performed by focusing on a group of attributes, if the groups are not independent, the bias of other groups of attributes may also change, making it difficult to achieve the bias target for each of the multiple groups.
• the disclosed technology aims to control the bias of each group, in which data is classified by attributes, so that it falls within the criteria set for each group.
• the disclosed technology accepts a bias criterion and a priority for each of a plurality of groups defined by the values of one or more attributes of a first plurality of data.
• the disclosed technology also generates a second plurality of data by modifying data from the first plurality of data corresponding to each of the plurality of groups based on the bias criterion and the degree of data modification for each of the plurality of groups according to the priority.
• the disclosed technology then generates training data by executing an optimization problem based on the bias criterion using the second plurality of data as initial values.
• One aspect is that it has the effect of controlling the bias of each group, in which data is classified by attribute values, so that it falls within the criteria set for each group.
• FIG. 2 is a functional block diagram of a training data generating device.
• FIG. 2 is a diagram illustrating an example of an input data set.
• FIG. 4 is a diagram illustrating an example of setting information.
• FIG. 1 is a block diagram showing a schematic configuration of a computer that functions as a training data generating device. 13 is a flowchart illustrating an example of a training data generation process.
• FIG. 11 is a diagram for explaining a training data generation process.
• the training data generation device 10 functionally includes a control unit 12.
• the control unit 12 further includes a reception unit 14, a first generation unit 16, a second generation unit 18, and a training unit 20.
• the reception unit 14 receives the input data set 30 and the setting information 32 input to the training data generation device 10.
• the example in FIG. 2 is an example of a dataset for generating training data used to train a machine learning model 38 that determines whether a test is passed or failed.
• each row is one piece of data, and each piece of data includes attribute values for multiple attributes.
• the attributes include protected attributes that need to be prevented from being treated discriminatory, attributes other than the protected attributes, and label attributes that indicate whether the data is a positive example or a negative example.
• gender and race are examples of protected attributes
• grades A and B are examples of attributes other than the protected attributes
• pass/fail is an example of a label attribute.
• the input dataset is an example of the "first multiple data" of the disclosed technology.
• the setting information 32 includes criteria and priorities for each group, as shown in FIG. 3.
• a group is defined by the attribute values of one or more attributes among the multiple attributes of the data.
• data is classified for each combination of the attribute values of one or more protected attributes to form a group.
• the group name of each group is expressed using the attribute value of the data belonging to that group. For example, as a group based on the attribute value of one attribute, a group with a gender attribute value of "female" is expressed as a "female group” or simply "female".
• a group with a gender attribute value of "female” and a race attribute value of "white” is expressed as a "female x white group” or simply "female x white”. From the input dataset in FIG. 2, the following groups are formed: female, male, white, non-white, white x female, non-white x female, white x male, and non-white x male.
• the standard for each group is the range indicated by the upper and lower limits of the bias that each group must achieve.
• the average of the upper and lower limits of the standard is referred to as the "target value.”
• the bias is expressed using DI (Disparate Impact), which is an example of a fairness index, as shown in the following formula (1).
• the numerator on the right side of equation (1) is the positive example rate of the target group, and the denominator is the positive example rate of the group other than the target group.
• the positive example rate is the number of positive example data items relative to the total number of data items.
• a positive example is, for example, data in which the attribute corresponding to the objective variable of binary classification has one of the values, whereas a negative example is, for example, data in which the attribute has the other value.
• the index representing the bias is not limited to DI, and any index representing the bias in the positive example rate of each group may be used.
• the priority for each group is the priority when correcting data (hereinafter referred to as "bias control") so that the bias for each group falls within a standard.
• bias control when performing bias control, there are cases where it is desired to correct the bias of a certain group so that it exceeds a target value, and cases where it is not necessary to reach the target value as long as the bias falls within the standard. For example, there are cases where it is desired to ensure that the "non-white x female" group achieves the target value, while keeping the bias of the "white x male” group within the standard.
• the priority is an index for determining the degree of data correction for each group, taking such cases into consideration.
• the setting information 32 includes an implementation probability (described in detail later) that indicates the degree of correction as a priority. The implementation probability is set to be larger the higher the priority, and smaller the lower the priority.
• the first generation unit 16 generates a modified data set by modifying data corresponding to each of the multiple groups from the input data set based on the bias criterion and the degree of data modification for each of the multiple groups according to the priority.
• the modified data set is an example of the "second multiple data" of the disclosed technology.
• the first generation unit 16 selects two groups from multiple groups that are independent of each other as the target for data correction. If bias control of one group affects the bias of another group, the first generation unit 16 determines that the two groups are non-independent. For example, when bias control of "non-white x female" is performed, it also affects the bias of "female.” In this case, the bias of the group that is affected by the bias control of the other group may fall outside the standard. Therefore, the first generation unit 16 first performs bias control on the independent groups.
• the first generation unit 16 determines whether all pairs, which are combinations of two groups selected from the multiple groups, are independent or non-independent. For example, if the two groups in a pair contain the same attribute value, the first generation unit 16 determines that the pair is non-independent. For example, the first generation unit 16 determines that "non-white x female" and “female” are non-independent, and "non-white x female” and “white x male” are independent.
• the first generation unit 16 calculates an evaluation score indicating the need for correction for each data item included in each group of independent pairs. For example, the first generation unit 16 attempts bias control for each pair, and identifies the data item to be corrected in the attempt. Note that, in this step, the data item to be corrected is only identified, and no actual correction is made. The first generation unit 16 tally up the results of the attempts for each pair for each group, and calculates an evaluation score for each data item in the group based on the number of times it was identified as data to be corrected. For example, the first generation unit 16 calculates the evaluation score as the ratio of the number of times it was identified to the number of attempts.
• the first generation unit 16 also calculates an evaluation score threshold based on the difference between the target value and the current DI for each group (hereinafter also referred to as an "independent group") in which the evaluation score of each data item is calculated. For example, when an evaluation score is calculated in which a larger value indicates that the data is more likely to be selected as a data item to be corrected, as described above, the first generation unit 16 calculates a threshold according to the difference between the target value and the current DI, which decreases as the difference increases. A large difference between the target value and the current DI means that more data needs to be corrected to correct the bias.
• the first generation unit 16 selects a group to be corrected from among the independent groups and performs bias control. As the group to be corrected, the first generation unit 16 selects a group whose current DI does not fall within the standard and whose evaluation score is equal to or greater than the calculated threshold value among the data not selected as a target for data correction. If there are multiple groups to be corrected, the first generation unit 16 may select the groups in descending order of the difference between the target value and the current DI.
• the first generation unit 16 selects, from the data included in the selected group and having an evaluation score equal to or higher than a threshold, a number of data according to the magnitude of the difference between the target value multiplied by the implementation probability and the current DI, in descending order of evaluation score.
• the first generation unit 16 then performs corrections, such as changing the label attribute of negative example data to the value of positive examples, so that the DI, which indicates bias, falls within the standard.
• corrections such as changing the label attribute of negative example data to the value of positive examples, so that the DI, which indicates bias, falls within the standard.
• the first generation unit 16 performs the above bias control until there are no more groups that can be selected as groups to be corrected, and stores the data set at the time when there are no more selectable groups in a specified storage area as a corrected data set 34.
• the corrected data set contains independent groups whose bias has reached the target value, independent groups whose bias has not reached the target, and non-independent groups for which bias control has not been performed.
• the second generation unit 18 generates training data by executing an optimization problem based on a bias criterion using the modified dataset 34 generated by the first generation unit 16 as an initial value. Executing the optimization problem based on the bias criterion involves correcting the selected data by solving a mathematical optimization problem so that the bias of each of the multiple groups falls within a criterion that takes into account the priority. The second generation unit 18 performs corrections such as changing the label attribute of negative example data to the value of positive examples so that the DI indicating the bias falls within the criterion. The second generation unit 18 stores the dataset obtained by solving the optimization problem in a specified storage area as the training dataset 36.
• the training unit 20 trains the machine learning model 38 using the training dataset 36 and outputs the trained machine learning model 38.
• the training data generation device 10 may be realized, for example, by a computer 40 shown in FIG. 4.
• the computer 40 includes a CPU (Central Processing Unit) 41, a GPU (Graphics Processing Unit) 42, a memory 43 as a temporary storage area, and a non-volatile storage device 44.
• the computer 40 also includes an input/output device 45 such as an input device and a display device, and an R/W (Read/Write) device 46 that controls the reading and writing of data from and to a storage medium 49.
• the computer 40 also includes a communication I/F (Interface) 47 that is connected to a network such as the Internet.
• the CPU 41, GPU 42, memory 43, storage device 44, input/output device 45, R/W device 46, and communication I/F 47 are connected to each other via a bus 48.
• the storage device 44 is, for example, a hard disk drive (HDD), a solid state drive (SSD), flash memory, etc.
• the storage device 44 which serves as a storage medium, stores a training data generation program 50 for causing the computer 40 to function as the training data generation device 10.
• the training data generation program 50 has an acceptance process control instruction 54, a first generation process control instruction 56, a second generation process control instruction 58, and a training process control instruction 60.
• the storage device 44 also has an information storage area 70 in which information constituting each of the modified dataset 34 and the training dataset 36 is stored.
• the CPU 41 reads the training data generation program 50 from the storage device 44, expands it in the memory 43, and sequentially executes the control instructions of the training data generation program 50.
• the CPU 41 operates as the reception unit 14 shown in FIG. 1 by executing the reception process control instruction 54.
• the CPU 41 also operates as the first generation unit 16 shown in FIG. 1 by executing the first generation process control instruction 56.
• the CPU 41 also operates as the second generation unit 18 shown in FIG. 1 by executing the second generation process control instruction 58.
• the CPU 41 also operates as the training unit 20 shown in FIG. 1 by executing the training process control instruction 60.
• the computer 40 that has executed the training data generation program 50 functions as the training data generation device 10.
• the CPU 41 that executes the program is hardware. Also, part of the program may be executed by the GPU 42.
• the functions realized by the training data generation program 50 may be realized, for example, by a semiconductor integrated circuit, more specifically, an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), etc.
• ASIC Application Specific Integrated Circuit
• FPGA Field-Programmable Gate Array
• the training data generation device 10 When the input data set 30 and the setting information 32 are input to the training data generation device 10 and an instruction is given to generate a training data set 36, the training data generation process shown in FIG. 5 is executed in the training data generation device 10. Note that the training data generation process is an example of a training data generation method of the disclosed technology.
• step S10 the reception unit 14 receives the input data set 30 input to the training data generation device 10 and the setting information 32 including the bias criteria and priority (in this embodiment, the implementation probability) for each group.
• step S12 the first generation unit 16 determines whether all pairs, which are combinations of two groups selected from the multiple groups formed from the input dataset 30, are independent or non-independent.
• the first generation unit 16 also attempts bias control for each pair of independent groups, and identifies data to be modified in the attempt.
• the first generation unit 16 then aggregates the results of the pair-by-pair attempt for each group, and calculates an evaluation score for each data item in the group based on the number of times that each data item in the group was identified as data to be modified.
• step S14 the first generation unit 16 calculates an evaluation score threshold for each independent group based on the difference between the target value and the current DI.
• step S16 the first generation unit 16 determines whether or not there is a group to be corrected among the independent groups. The first generation unit 16 determines that, among the independent groups, a group whose current DI does not fall within the criteria and which contains data whose evaluation score is equal to or greater than the calculated threshold among data not selected as a target for data correction is a group to be corrected. If there is a group to be corrected, the process proceeds to step S18.
• step S18 the first generation unit 16 selects a group to be corrected that is determined to exist in step S16. If there are multiple groups to be corrected, the first generation unit 16 may select the groups in descending order of the difference between the target value and the current DI.
• step S20 the first generation unit 16 selects a number of data items corresponding to the difference between the target value multiplied by the implementation probability and the current DI from the data items included in the selected group and having an evaluation score equal to or greater than a threshold, in descending order of evaluation score.
• the first generation unit 16 executes bias control by, for example, modifying the label attribute of the negative example data to the value of the positive example so that the DI indicating the bias falls within the standard, stores the modified data set 34 in a specified storage area, and returns to step S14.
• the first generation unit 16 updates the modified data set 34 each time it executes the bias control in step S20.
• step S22 the second generation unit 18 generates training data by executing an optimization problem based on the bias criterion, using the corrected data set 34 stored in a specified storage area as the initial value.
• the second generation unit 18 stores the generated training data set 36 in a specified storage area, and the training data generation process ends.
• the training unit 20 trains the machine learning model 38 using the generated training dataset 36 and outputs the trained machine learning model 38.
• the training data generation device receives bias criteria and priorities for each of multiple groups defined by the values of one or more attributes of multiple attributes of an input dataset.
• the training data generation device also generates a modified dataset in which data corresponding to each of multiple groups in the input dataset is modified based on the bias criteria and the degree of data modification for each of the multiple groups according to the priority.
• the training data generation device then generates training data by executing an optimization problem based on the bias criteria using the modified dataset as an initial value. This makes it possible to control the bias of each group in which data is classified by attribute values so that it falls within the criteria set for each group. In other words, rather than relaxing each group, bias control is possible in accordance with the purpose of the task using the machine learning model.
• FIG. 6 shows an outline of the above-mentioned training data generation process using a specific example.
• the numerical value in each quadrant in FIG. 6 is the DI of the group corresponding to that quadrant.
• the training data generation device first performs bias control for independent groups based on the bias criterion and priority (execution probability in the above embodiment) for each group.
• the training data generation device sets the corrected data set generated by performing bias control for the independent groups as the initial value of the optimization problem.
• the training data generation device performs bias control based on the bias criterion and priority for each group by solving the optimization problem including non-independent and non-target-attaining groups.
• the setting information includes the implementation probability as a priority, but this is not limiting.
• the setting information may include priorities such as high, medium, and low, and a ranking in order of priority.
• the training data generation device may determine the implementation probability according to the received priority. For example, the training data generation device may determine the implementation probability as follows: when the priority is "high”, the implementation probability is "1.0", when the priority is "medium”, the implementation probability is "0.8", and when the priority is "low”, the implementation probability is "0.6".
• the training data generation device may also determine an implementation probability that is higher the higher the priority ranking and lower the lower the ranking.
• the training data generation program is pre-stored (installed) in the storage device, but this is not limited to this.
• the program related to the disclosed technology may be provided in a form stored in a storage medium such as a CD-ROM, DVD-ROM, or USB memory.
• Training data generation device 12
• Control unit 14
• Reception unit 16
• First generation unit 18
• Second generation unit 20
• Training unit 30
• Input data set 32
• Setting information 34
• Corrected data set 36
• Training data set 38
• Machine learning model 40
• Computer 41
• GPUs 43
• Memory 44
• Input/output device 46
• R/W device 47
• Communication I/F 48
• Bus 49
• Training data generation program 54
• Acceptance process control command 56
• First generation process control command 58
• Second generation process control command 60
• Training process control command 70

Landscapes

• Engineering & Computer Science (AREA)
• Business, Economics & Management (AREA)
• Human Resources & Organizations (AREA)
• Theoretical Computer Science (AREA)
• Physics & Mathematics (AREA)
• Software Systems (AREA)
• Economics (AREA)
• Entrepreneurship & Innovation (AREA)
• General Physics & Mathematics (AREA)
• Strategic Management (AREA)
• General Business, Economics & Management (AREA)
• Data Mining & Analysis (AREA)
• Tourism & Hospitality (AREA)
• Operations Research (AREA)
• Marketing (AREA)
• Game Theory and Decision Science (AREA)
• Educational Administration (AREA)
• Artificial Intelligence (AREA)
• Computer Vision & Pattern Recognition (AREA)
• Quality & Reliability (AREA)
• Evolutionary Computation (AREA)
• Medical Informatics (AREA)
• Development Economics (AREA)
• Computing Systems (AREA)
• General Engineering & Computer Science (AREA)
• Mathematical Physics (AREA)
• Management, Administration, Business Operations System, And Electronic Commerce (AREA)
• Complex Calculations (AREA)

Priority Applications (3)

Applications Claiming Priority (1)

Publications (1)

Family

ID=92590263

Family Applications (1)

Country Status (3)

Citations (2)

• 2023
  • 2023-03-02 JP JP2025503557A patent/JPWO2024180775A1/ja active Pending
  • 2023-03-02 EP EP23925329.7A patent/EP4675515A4/en active Pending
  • 2023-03-02 WO PCT/JP2023/007895 patent/WO2024180775A1/ja not_active Ceased

Patent Citations (2)

Non-Patent Citations (2)

Also Published As

Similar Documents

Legal Events