Computer implementation method, computer program product, and computer system for controlling background bias in machine learning modules.

Abstract

To provide a system, method and program for generating a machine learning module.SOLUTION: A method includes the steps of: providing training data including a first data set corresponding to a first feature amount and a second data set corresponding to a second feature amount; generating a further data set depending on the first data set and the second data set; calculating a first correlation metric depending on the further data set and a selected data set; confirming whether the first correlation metric is larger than a first threshold; training an ML module depending on the further data set; defining a first subset of a value of a further feature amount and a second subset of the value of the further feature amount; determining a bias metric of an ML module corresponding to the first subset or the second subset; confirming bias constraint; and releasing the ML module.SELECTED DRAWING: Figure 9

Claims

1. A computer implementation method, wherein the method is To provide training data comprising a first dataset corresponding to at least a first feature and a second dataset corresponding to a second feature, wherein one of the first and second features is a selected feature, and the dataset corresponding to the selected feature is a selected dataset. The method involves automatically generating a further dataset that depends on at least one of the first and second datasets, wherein the further dataset corresponds to additional features. A first correlation metric is calculated as a measure of the correlation between the selected feature and the further feature, depending at least on the further dataset and the selected dataset. To confirm whether the first correlation metric is greater than the first threshold, This involves training an ML module that depends at least on the aforementioned additional dataset, Define at least a first subset of the values ​​of the further features and a second subset of the values ​​of the further features that are different from the first subset, according to the values ​​of the further features. If the first correlation metric is greater than the first threshold, a bias metric is determined that indicates the strength of the bias of the ML-module on the first subset of the values ​​of the further features or the second subset of the values ​​of the further features. To verify whether the bias metric satisfies the bias constraint, If the bias metric satisfies the bias constraint, the ML module is released for use. Methods that include...

2. The method according to claim 1, further comprising generating the further datasets that depend on the selected datasets.

3. The method according to claim 1, further comprising generating the further datasets that depend on the first dataset and the second dataset, which are not the selected datasets.

4. The method according to claim 1, further comprising generating the further datasets that depend on the selected dataset and the datasets of the first dataset and the second dataset that are not the selected dataset.

5. The selected feature quantities are selected by the user, and the method is, A second correlation metric is calculated as a measure of the correlation between the first and second features that are not selected (non-selected features) and the selected features that depend on the first and second datasets that are not selected datasets and the selected datasets. To confirm whether the second correlation metric is greater than the second threshold, The method according to claim 1, further comprising:

6. The method according to claim 1, further comprising, depending on user correlation input data, performing the calculation of the first correlation metric, or the confirmation of whether the first correlation metric is greater than the first threshold, or both.

7. The method according to claim 1, further comprising defining at least a first subset of the values ​​of the further features and a second subset of the values ​​of the further features, depending on user bias input data.

8. The method according to claim 1, further comprising using the trained ML module to compute an output value of the ML module that depends on an application input dataset each containing values ​​corresponding to at least the further features, and storing the bias metric for monitoring the bias of the trained ML module with respect to the further features in the form of metadata for the trained ML module.

9. The method according to claim 1, further comprising using the trained ML module to compute an output value of the ML module that depends on an application input dataset each containing values ​​corresponding to at least the further features, and storing the first correlation metric for monitoring the bias of the trained ML module with respect to the further features in the form of metadata for the trained ML module.

10. The method according to claim 5, further comprising storing in the form of metadata for the trained ML module a further bias metric for monitoring the bias of the trained ML module with respect to the unselected features, when using the trained ML module to compute an output value of the ML module that depends on an application input dataset each containing values ​​corresponding to at least the unselected features.

11. The method according to claim 5, further comprising using the trained ML module to compute an output value of the ML module that depends on an application input dataset each containing values ​​corresponding to at least the unselected features, and storing the second correlation metric for monitoring the bias of the trained ML module with respect to the unselected features in the form of metadata for the trained ML module.

12. The aforementioned method, Modify the structure of the ML module, Repeat the execution of the training of the ML module, Repeating the calculation of the bias metric, To confirm whether the aforementioned bias metric has decreased, If the bias metric has decreased, the ML module having the modified structure will be made available for use, and the previous version of the ML module will be rejected. The method according to claim 1, including the method described in claim 1.

13. The aforementioned method, Modify the structure of the ML module, The generation of the further dataset, which depends on at least one of the first dataset and the second dataset, is repeated using different mathematical functions, Repeat the execution of the training of the ML module, Repeating the calculation of the bias metric, To confirm whether the aforementioned bias metric has decreased, If the bias metric has decreased, the ML module with the modified structure will be made available for use, and the previous version of the ML module will be rejected. The method according to claim 1, including the method described in claim 1.

14. The method further includes repeating the modification of the structure of the ML module, the execution of the training of the ML module, and the calculation of the bias metric, and, in order to further reduce the bias of the ML module, specifying the structure of the ML module for each iteration and storing each set of structural parameter values ​​together with the respective bias metric, in order to provide a database for further optimization of the structure of the ML module. The method according to claim 12.

15. The method further includes storing the set of structural parameter values, along with their respective bias metrics, in the form of metadata for the ML-module. The method according to claim 14.

16. A computer program, wherein the computer program includes program instructions, and the program instructions are To provide training data comprising a first dataset corresponding to at least a first feature and a second dataset corresponding to a second feature, wherein one of the first and second features is a selected feature, and the dataset corresponding to the selected feature is a selected dataset. The method involves automatically generating a further dataset that depends on at least one of the first and second datasets, wherein the further dataset corresponds to additional features. A first correlation metric is calculated as a measure of the correlation between the selected feature and the further feature, depending at least on the further dataset and the selected dataset. To confirm whether the first correlation metric is greater than the first threshold, This involves training an ML module that depends at least on the aforementioned additional dataset, Define at least a first subset of the values ​​of the further features and a second subset of the values ​​of the further features that are different from the first subset, according to the values ​​of the further features. If the first correlation metric is greater than the first threshold, a bias metric is determined that indicates the strength of the bias of the ML-module on the first subset of the values ​​of the further features or the second subset of the values ​​of the further features. To verify whether the bias metric satisfies the bias constraint, If the bias metric satisfies the bias constraint, the ML module is released for use. A computer program that includes [this].

17. A computer system, wherein the computer system is To provide training data comprising a first dataset corresponding to at least a first feature and a second dataset corresponding to a second feature, wherein one of the first and second features is a selected feature, and the dataset corresponding to the selected feature is a selected dataset. The method involves automatically generating a further dataset that depends on at least one of the first and second datasets, wherein the further dataset corresponds to additional features. A first correlation metric is calculated as a measure of the correlation between the selected feature and the further feature, depending at least on the further dataset and the selected dataset. To confirm whether the first correlation metric is greater than the first threshold, This involves training an ML module that depends at least on the aforementioned additional dataset, Define at least a first subset of the values ​​of the further features and a second subset of the values ​​of the further features that are different from the first subset, according to the values ​​of the further features. If the first correlation metric is greater than the first threshold, a bias metric is determined that indicates the strength of the bias of the ML-module on the first subset of the values ​​of the further features or the second subset of the values ​​of the further features. To verify whether the bias metric satisfies the bias constraint, If the bias metric satisfies the bias constraint, the ML module is released for use. A computer system, including a computer system.

18. The computer system according to claim 17, further comprising generating a further dataset that depends on the selected dataset.

19. The computer system according to claim 17, further comprising generating the further datasets which depend on the first dataset and the second dataset, which are not the selected datasets.

20. The computer system according to claim 17, further comprising generating the further datasets that depend on the selected dataset and the datasets of the first dataset and the second dataset that are not the selected dataset.

Images