JP7582506B2 - Data correction program, device, and method - Google Patents

Description

The disclosed technology relates to a data correction program, a data correction device, and a data correction method.

The values of certain attributes contained in the training data used to train a machine learning model may become biased, causing the judgment results of the machine learning model to be discriminatory. For example, consider a case where a machine learning model that predicts the pass/fail outcome from a person's attributes is trained using training data in which the attribute values of a person's gender, age, place of origin, etc. are used as explanatory variables, and the pass/fail outcome of the person in a recruitment or test, etc., is used as the objective variable. In this case, if the training data is a past history in which being female has resulted in unfavorable treatment in terms of pass/fail outcomes, the machine learning model trained using that training data will make discriminatory predictions that judge women unfavorably.

As mentioned above, to prevent machine learning models from making discriminatory predictions, there is a technology called DIR (Disparate Impact Remover) that removes factors in training data that bias discriminatory predictions. DIR reduces the correlation between the attribute to be protected from discriminatory treatment ("gender" in the above example) and the values of other attributes, and modifies the values of the other attributes to suppress bias in the data.

Michael Feldman, Sorelle A. Friedler, John Moeller, Carlos Scheidegger, Suresh Venkatasubramanian, "Certifying and Removing Disparate Impact", Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, August 2015, Pages 259-268.

However, DIR is a method for making corrections to attributes whose variables are numerical values that have superiority or inferiority, such as large and small values, so-called numerical features. Therefore, there is a problem in that DIR cannot be applied to correct attributes whose variables are limited numbers or categories, so-called categorical variables.

In one aspect, the disclosed technology aims to remove factors that bias machine learning models toward discriminatory predictions from data containing categorical variables.

In one aspect, the disclosed technology generates a second plurality of data by changing a first type of categorical variable contained in each of the first plurality of data to a one-hot representation. The disclosed technology also sets a first component of the one-hot representation of the first type of categorical variable contained in each of the second plurality of data as a response variable, and sets a portion of the second plurality of data other than the one-hot representation of the first type of categorical variable as an explanatory variable. The disclosed technology inputs the portion of the second plurality of data other than the one-hot representation of the first type of categorical variable to a machine learning model generated based on training data including the above-mentioned response variable and explanatory variable. The disclosed technology ranks each of the second plurality of data based on the estimation result of the machine learning model in that case. The disclosed technology then generates a third plurality of data by correcting the bias of each attribute of the first type of categorical variable in the second plurality of data based on the result of the ranking process.

One aspect is that it has the effect of removing factors that bias machine learning models toward discriminatory predictions from data containing categorical variables.

FIG. 2 is a functional block diagram of a data correction device. FIG. 13 is a diagram illustrating an example of a data set. FIG. 13 is a diagram showing an example of a transformed data set. FIG. 1 is a diagram for explaining problems that arise when applying DIR to a data set that includes categorical variables. FIG. 11 is a diagram for explaining calculation of a certainty factor. FIG. 13 is a diagram for explaining setting of ranks. FIG. 13 is a diagram showing an example of a combination of a value of a component to be corrected and a certainty factor for each rank, and a median value of the rank. FIG. 13 is a diagram for explaining correction of inconsistency in one-hot representation. FIG. 13 is a diagram for explaining correction of inconsistency in one-hot representation. FIG. 1 is a block diagram showing a schematic configuration of a computer functioning as a data correction device. 13 is a flowchart illustrating an example of a data correction process.

Below, an example of an embodiment of the disclosed technology is described with reference to the drawings.

As shown in FIG. 1, a data set including a plurality of data is input to the data correction device 10. Then, the data correction device 10 corrects factors (hereinafter simply referred to as "bias") that are included in the input data set and that bias the machine learning model toward discriminatory predictions, and outputs a corrected data set. The data set is an example of the "first plurality of data" of the disclosed technology, and the corrected data set is an example of the "third plurality of data" of the disclosed technology. FIG. 2 shows an example of a data set. In the example of FIG. 2, each row (each record) is one piece of data, and each piece of data is assigned an "ID" that is identification information for that data. In addition, in the example of FIG. 2, each piece of data has values (variables) for each of attributes such as "gender", "age", "place of origin", and "work type" as features. The value of the attribute "age" is a numerical feature. In addition, the values of the attributes "gender", "place of origin", and "work type" are categorical variables.

Here, we consider a method of applying DIR to remove bias from a dataset containing categorical variables as shown in Figure 2. As described above, DIR modifies numerical features and cannot be applied to categorical variables. Therefore, the following two simple methods are considered for applying DIR to a dataset containing categorical variables. The first method is to modify only the numerical features without modifying the categorical variables. The second method is to remove the categorical variable column from the dataset, leaving only the numerical feature column, and apply DIR. However, in the first method, bias remains in the categorical variables, so it is not possible to completely remove bias from the dataset. In addition, in the second method, information useful for prediction in a machine learning model is missing from the dataset, which may significantly reduce the accuracy of the machine learning model trained using the dataset.

In addition, in order to apply DIR to a dataset that contains categorical variables, it is possible to simply convert the categorical variables into numerical features by replacing them with numerical values. However, in this case, even though there is no superiority or inferiority between the categories, the converted numerical values will have superiority or inferiority, such as large or small, which will affect the accuracy of the machine learning model being trained.

Therefore, it is possible to convert the categorical variable into a one-hot representation and apply DIR. A one-hot representation is an expression in which one of multiple components has a value of 1 and the other components have a value of 0. A one-hot representation converts one column of categorical variables into multiple columns of numerical features. In the example of Figure 2, Figure 3 shows an example of converting the categorical variable "work style" into a one-hot representation. The categorical variable "work style", which includes the categories "full-time" and "part-time", is converted into a one-hot representation with multiple components (columns) of "full-time" and "part-time", as shown in the bold frame in Figure 3. In this example, in each data, one of the components of the one-hot representation, "full-time" and "part-time", has a value of 1, and the other has a value of 0. Focusing on each component (column) of the one-hot representation, each component is a numerical feature.

The DIR algorithm also generates rankings for groups based on the values of attributes that are protected from discriminatory treatment, such as gender and nationality (hereafter referred to as "protected attributes"), based on the values of numerical features that are the values of attributes other than the protected attributes. The DIR algorithm then corrects the values of numerical features other than the protected attributes to the median of the values of numerical features that belong to the same rank. When focusing on a numerical feature that is one component of a one-hot representation, the only possible values are 0 or 1, so when the DIR algorithm described above is applied, there are only two ranks: a rank with a value of 0 and a rank with a value of 1. Therefore, the median of numerical features that belong to a rank with a value of 0 is 0, and the median of numerical features that belong to a rank with a value of 1 is 1, and because the values do not change before and after the correction, bias is not removed.

A more detailed explanation will be given with reference to FIG. 4. As shown in FIG. 4, in a data set in which the categorical variable "working style" is converted to a one-hot representation, the protected attribute is "gender" and the numerical feature to be corrected is "full-time", which is one of the components of the one-hot representation of the categorical variable "working style". The information processing device that executes DIR extracts the "full-time" value of each of the data included in the group of data in which the value of the protected attribute "gender" is "male" (hereinafter also referred to as the "male group"), as shown in A in FIG. 4. Similarly, the information processing device extracts the "full-time" value of each of the data included in the group of data in which the value of the attribute "gender" is "female" (hereinafter also referred to as the "female group"). The information processing device obtains a set of values extracted from each group without duplication. In this case, the information processing device obtains a set of (0, 1) for each group.

In addition, the information processing device sets the size of the ranking to, for example, the smaller of the sizes of the sets obtained from each group, i.e., the number of values contained in the set. Here, the size of the sets for all groups is 2, so the size of the ranking is also 2. For each group, the information processing device sets a rank to the value contained in the obtained set depending on the size of the value. For example, as shown in B in FIG. 4, for the male group, the information processing device sets rank 1 to the value "1" and rank 2 to the value "0". For the female group, the information processing device sets rank 1 to the value "1" and rank 2 to the value "0". Note that if the size of the set is larger than the size of the ranking, the same rank may be set to two or more values.

As shown in C in Figure 4, the information processing device collects data from each group for each rank that has the value set for that rank as the "full-time" value, and calculates the median of the "full-time" values of data in the same rank. Then, as shown in D in Figure 4, the information processing device corrects the "full-time" value of each data to the median calculated for the rank to which the data belongs. In this case, all "full-time" values belonging to rank 1 are "1", so the median is also "1", and all "full-time" values belonging to rank 2 are "0", so the median is also "0". Therefore, with the above method, the values do not change before and after the correction, so the bias is not removed.

In this embodiment, when creating a ranking of data for each group based on the values of the components to be corrected, the ranking is created by further ranking data that have the same values of the components to be corrected. The functional configuration of the data correction device 10 according to this embodiment is described in detail below.

Functionally, as shown in Figure 1, the data correction device 10 includes a conversion unit 12, a ranking unit 14, and a correction unit 16.

The conversion unit 12 converts the categorical variables to be corrected, which are included in each piece of data included in the data set, into a one-hot representation, and generates a converted data set. The categorical variables to be corrected are an example of a "first type of categorical variable" in the disclosed technology, and the converted data set is an example of a "second plurality of data" in the disclosed technology. For example, in a data set as shown in FIG. 2, if the categorical variable to be corrected is "work style", the conversion unit 12 converts the value of the categorical variable of the attribute "work style" into a one-hot representation, and generates a converted data set as shown in FIG. 3. Note that each row (each record) of the converted data set shown in FIG. 3 corresponds to one piece of converted data.

The ranking unit 14 calculates a proxy value for ranking the values of each component of the one-hot expression of the categorical variable to be corrected in the converted data. The proxy value may be, for example, a confidence level indicating the likelihood of the value of each component of the one-hot expression for the part of the converted data other than the one-hot expression of the categorical variable to be corrected. Specifically, as shown in FIG. 5, the ranking unit 14 sets each component of the one-hot expression of the categorical variable to be corrected contained in each of the converted data included in the converted data set as the objective variable (in the example of FIG. 5, the diagonally shaded columns "Component 1" and "Component 2") as the objective variable. In addition, the ranking unit 14 sets the part of the converted data other than the categorical variable to be corrected (in the example of FIG. 5, "Attribute 1" to "Attribute 6") as the explanatory variable. The ranking unit 14 generates a machine learning model based on training data consisting of the above-mentioned objective variable and explanatory variable. The ranking unit 14 may perform machine learning of the machine learning model using all of the training data. In addition, the ranking unit 14 may perform machine learning of the machine learning model using a portion of the training data (e.g., 4/5 of the total), and perform cross-validation to validate the generated machine learning model using the remaining portion of the training data (e.g., 1/5 of the total).

The ranking unit 14 inputs the portion of the converted data other than the one-hot representation of the categorical variable to be corrected, i.e., the explanatory variable portion of the training data, into the generated machine learning model, and obtains the confidence level for each component of each converted data as the prediction result of the machine learning model. The confidence level for each component is a value between 0.0 and 1.0, and the sum of the confidence levels for each component of each converted data is 1.0.

The ranking unit 14 ranks each piece of transformed data based on the confidence level, which is the prediction result of the machine learning model. Specifically, the ranking unit 14 arranges the transformed data by the value of the protected attribute in the order of the component values of the one-hot expression of the categorical variable to be modified. Furthermore, the ranking unit 14 arranges transformed data having the same value of the component to be modified in the order of decreasing confidence level to set a rank. The protected attribute is an example of a "first attribute" of the disclosed technology.

A more detailed description will be given with reference to FIG. 6. The ranking unit 14 accepts the designation of the protected attribute. For example, if the designation of the protected attribute accepted is the attribute "gender", the ranking unit 14 extracts the converted data in which the value of the component to be corrected is "1" and the converted data in which the value of the component to be corrected is "0" for each of the male group and the female group. The ranking unit 14 arranges the converted data in which the value of the component to be corrected is "1" in descending order of confidence for each group. Similarly, the ranking unit 14 arranges the converted data in which the value of the component to be corrected is "0" in descending order of confidence for each group. The ranking unit 14 concatenates the converted data in which the value of the component to be corrected is "0" in descending order of confidence for each group, after the converted data in which the value of the component to be corrected is "1" in descending order of confidence for each group. As a result, the converted data in which the value of the component to be corrected is "0" in descending order of confidence for each group is arranged in descending order of confidence.

The ranking unit 14 obtains a set of combinations of the value of the component to be corrected and the confidence for each group without duplicates, and sets the size of the set, i.e., the number of combinations included in the set, whichever is smaller, as the size of the ranking. The ranking unit 14 then sets a rank for the transformed data arranged in order of the value of the component to be corrected and in order of decreasing confidence, according to the size of the ranking. For example, suppose that for two transformed data in which one component of the one-hot representation of a categorical variable has the same value "1", the confidence for that component is 0.998 for one transformed data and 0.940 for the other transformed data. In this case, the former transformed data is set to a higher rank.

The correction unit 16 generates a corrected data set by correcting the bias of each attribute of the categorical variable to be corrected in the converted data set based on the result of the ranking process by the ranking unit 14. Specifically, the correction unit 16 corrects the value of the component to be corrected in the converted data set having the same rank set by the ranking unit 14 to the median value of the component to be corrected in the converted data set having the same rank. FIG. 7 shows an example of a combination of the value of the component to be corrected and the confidence level for each rank, and the median value of the rank. Even if the value of the component to be corrected is the same, the rank differs depending on the high or low confidence level, so that even if the rank is the same, the value may differ between the male group and the female group (rank 453 in the example of FIG. 7). In this case, the value of the component to be corrected in the converted data of one of the groups is changed from the original value. That is, it becomes possible to apply DIR to the categorical variable. Note that, when the median value is 0.5, the correction unit 16 may set a predetermined value of 1 or 0 as the median, i.e., the corrected value.

Furthermore, the correction unit 16 corrects the one-hot representation of the categorical variable to be corrected so as to maintain the consistency of the one-hot representation after changing the component value to the median value. Specifically, as shown in FIG. 8, there may be a plurality of one-hot values, i.e., "1", in the one-hot representation of the categorical variable to be corrected. In this case, the correction unit 16 leaves the value of one of the multiple "1"s as "1" and corrects the values of the other components to "0" based on the confidence of each component. For example, the correction unit 16 may leave the value of the component with the highest confidence as "1". In the example of FIG. 8, the values of component 1 and component 2 are "1", but since component 1 has a higher confidence than component 2, the correction unit 16 corrects the value of component 2 from "1" to "0" (thick frame in FIG. 8).

Also, as shown in FIG. 9, there may be cases where the one-hot value, i.e., "1", does not exist in the one-hot representation of the categorical variable to be modified. In this case, the modification unit 16 modifies the value of one of the components to "1" based on the confidence of each component. For example, the modification unit 16 may modify the value of the component with the highest confidence to "1". In the example of FIG. 9, since the values of components 1, 2, and 3 are all "0", the modification unit 16 modifies the value of component 1, which has the highest confidence, from "0" to "1" (bold frame in FIG. 9). The modification unit 16 also performs modifications to maintain the consistency of the one-hot representation as described above, and then outputs the modified data set.

The data modification device 10 may be realized, for example, by a computer 40 shown in FIG. 10. The computer 40 includes a CPU (Central Processing Unit) 41, a memory 42 as a temporary storage area, and a non-volatile storage unit 43. The computer 40 also includes an input/output device 44 such as an input unit and a display unit, and an R/W (Read/Write) unit 45 that controls the reading and writing of data to a non-temporary storage medium 49. The computer 40 also includes a communication I/F (Interface) 46 that is connected to a network such as the Internet. The CPU 41, memory 42, storage unit 43, input/output device 44, R/W unit 45, and communication I/F 46 are connected to each other via a bus 47.

The storage unit 43 may be realized by a hard disk drive (HDD), a solid state drive (SSD), a flash memory, etc. The storage unit 43 as a storage medium stores a data correction program 50 for causing the computer 40 to function as the data correction device 10. The data correction program 50 has a conversion process 52, a ranking process 54, and a correction process 56.

The CPU 41 reads the data correction program 50 from the storage unit 43, expands it in the memory 42, and sequentially executes the processes of the data correction program 50. The CPU 41 operates as the conversion unit 12 shown in FIG. 1 by executing the conversion process 52. The CPU 41 also operates as the ranking unit 14 shown in FIG. 1 by executing the ranking process 54. The CPU 41 also operates as the correction unit 16 shown in FIG. 1 by executing the correction process 56. As a result, the computer 40 that has executed the data correction program 50 functions as the data correction device 10. The CPU 41 that executes the program is hardware.

In addition, the functions realized by the data correction program 50 can also be realized, for example, by a semiconductor integrated circuit, more specifically, an ASIC (Application Specific Integrated Circuit).

Next, the operation of the data correction device 10 according to this embodiment will be described. When a data set is input to the data correction device 10 and an instruction to remove the bias is given, the data correction device 10 executes the data correction process shown in FIG. 11. Note that the data correction process is an example of a data correction method of the disclosed technology.

In step S10, the conversion unit 12 acquires a data set input to the data correction device 10. In addition, the ranking unit 14 accepts designation of protection attributes from the user. Next, in step S12, the conversion unit 12 converts categorical variables to be corrected that are included in the acquired data set into one-hot representations, and generates a converted data set. The categorical variables to be corrected may be categorical variables designated by the user, or all categorical variables included in the data set may be set sequentially as categorical variables to be corrected.

Next, in step S14, the ranking unit 14 generates training data in which each component of the one-hot representation of the categorical variable to be corrected contained in the transformed data set is used as the objective variable, and attributes other than the categorical variable to be corrected are used as the explanatory variables. Based on the generated training data, the ranking unit 14 generates a machine learning model that estimates a confidence level indicating the likelihood of the value of each component of the one-hot representation for the part other than the one-hot representation of the categorical variable to be corrected.

Next, in step S16, the ranking unit 14 inputs the explanatory variable portion of the generated training data into the generated machine learning model. Then, the ranking unit 14 obtains the confidence level for each component of the one-hot representation of the categorical variable to be corrected in each transformed data as the prediction result of the machine learning model.

Next, in step S18, the ranking unit 14 sorts the transformed data by the values of the components of the one-hot representation of the categorical variable to be modified, for each value of the protection attribute received in step S10. Furthermore, the ranking unit 14 sorts the transformed data having the same values of the components to be modified in order of decreasing confidence, and sets a rank.

Next, in step S20, the correction unit 16 corrects the value of the component to be corrected in the transformed data to the median value of the component to be corrected in the set of transformed data having the same rank set in step S18 above. Next, in step S22, for transformed data in which there is an inconsistency in that the one-hot representation of the categorical variable to be corrected has multiple "1s" or no "1s", the correction unit 16 corrects the value of one of the components to be "1" based on the confidence of each component. Then, the correction unit 16 outputs the corrected data set, and the data correction process ends.

As described above, the data correction device according to this embodiment generates a converted data set in which the categorical variable to be corrected included in the data set is changed to a one-hot representation. The data correction device also generates a machine learning model based on training data in which each component of the one-hot representation of the categorical variable to be corrected included in the converted data set is used as the objective variable, and attributes other than the categorical variable to be corrected are used as explanatory variables. The data correction device then ranks each of the converted data based on the prediction result of the machine learning model when a part other than the one-hot representation of the categorical variable to be corrected is input from the converted data set. Furthermore, the data correction device generates and outputs a corrected data set by correcting the bias of the values of each component of the categorical variable to be corrected due to the difference in the value of the protection attribute in the converted data set based on the result of the ranking process. This makes it possible to remove factors that bias the discriminatory prediction of the machine learning model from the data including the categorical variable. Furthermore, by generating a machine learning model using the corrected data set generated by the data correction device according to this embodiment as training data, discriminatory prediction by the machine learning model can be suppressed.

In the above embodiment, the data correction program is pre-stored (installed) in the storage unit, but this is not limiting. The program according to the disclosed technology can also be provided in a form stored in a non-transitory storage medium such as a CD-ROM, DVD-ROM, or USB memory.

10 Data correction device 12 Conversion unit 14 Ranking unit 16 Correction unit 40 Computer 41 CPU
42 Memory 43 Storage unit 44 Input/output device 45 R/W unit 46 Communication I/F
47 Bus 49 Storage medium 50 Data correction program

Claims (9)

A second plurality of data is generated by converting a first type of categorical variable included in each of the first plurality of data into a one-hot representation;
a first component of a one-hot expression of the first type categorical variable included in each of the second plurality of data is used as a response variable, and a machine learning model is generated based on training data in which a portion of the second plurality of data other than the one-hot expression of the first type categorical variable is used as an explanatory variable, and the portion of the second plurality of data other than the one-hot expression of the first type categorical variable is input to the machine learning model, and each of the second plurality of data is ranked based on an inference result of the machine learning model when the portion of the second plurality of data other than the one-hot expression of the first type categorical variable is input;
generating a third plurality of data by correcting bias of each attribute of the first type of categorical variable in the second plurality of data based on a result of the ranking process;
A data correction program that causes a computer to execute a process. the ranking process includes arranging the data in order of the value of the first component for each value of a first attribute, and arranging data having the same value of the first component in order of the value indicated by the prediction result to set a rank;
The data correction program according to claim 1 . the process of correcting the bias includes correcting a value of the first component of data having the same rank set by the value of the first attribute to a median value of values of the first component in a set of data having the same rank.
The data correction program according to claim 2 . The process of correcting the bias includes correcting the one-hot representation of the first type categorical variable after the correction so as to maintain consistency.
The data correction program according to any one of claims 1 to 3. The process of correcting the one-hot representation so as to maintain the consistency of the one-hot representation includes, when a plurality of one-hot values exist in the one-hot representation, correcting the one-hot representation so as to leave one of the plurality of one-hot values based on the inference result, and, when a one-hot value does not exist in the one-hot representation, changing a value of any component in the one-hot representation to a one-hot value based on the inference result.
The data correction program according to claim 4. The prediction result of the machine learning model is a confidence level of the value of the first component for a portion of the second plurality of data other than the one-hot representation of the first type categorical variable.
The data correction program according to any one of claims 1 to 5. performing machine learning of the machine learning model using all of the training data, or performing machine learning of the machine learning model using a portion of the training data and validating the generated machine learning model using the remainder of the training data;
7. The data correction program according to claim 1, wherein the data correction program causes the computer to execute a process. A second plurality of data is generated by converting a first type of categorical variable included in each of the first plurality of data into a one-hot representation;
a machine learning model generated based on training data in which a first component of a one-hot expression of the first type categorical variable included in each of the second plurality of data is used as a response variable, and a portion of the second plurality of data other than the one-hot expression of the first type categorical variable is used as an explanatory variable; and the machine learning model is generated based on training data in which a portion of the second plurality of data other than the one-hot expression of the first type categorical variable is input to the machine learning model, and each of the second plurality of data is ranked based on an inference result of the machine learning model when the portion of the second plurality of data other than the one-hot expression of the first type categorical variable is input;
generating a third plurality of data by correcting bias of each attribute of the first type of categorical variable in the second plurality of data based on a result of the ranking process;
A data correction device comprising a control unit for executing processing. A second plurality of data is generated by converting a first type of categorical variable included in each of the first plurality of data into a one-hot representation;
a first component of a one-hot expression of the first type categorical variable included in each of the second plurality of data is used as a response variable, and a machine learning model is generated based on training data in which a portion of the second plurality of data other than the one-hot expression of the first type categorical variable is used as an explanatory variable, and the portion of the second plurality of data other than the one-hot expression of the first type categorical variable is input to the machine learning model, and each of the second plurality of data is ranked based on an inference result of the machine learning model when the portion of the second plurality of data other than the one-hot expression of the first type categorical variable is input;
generating a third plurality of data by correcting bias of each attribute of the first type of categorical variable in the second plurality of data based on a result of the ranking process;
A data correction method, the processing of which is carried out by a computer.