Bias source identification and de-biasing of a dataset

1. A method, in a data processing system comprising a processor and memory, the memory comprising instructions executed by the processor to cause the processor to implement a source of bias identification (SoBI) computing tool that identifies sources of bias in a dataset, the method comprising:

executing, by a bias detection tool executing in the data processing system, a bias detection operation on results of an operation of a computer model, based on an input dataset, to generate a plurality of separate groupings of values for a protected attribute corresponding to a detected bias in the operation of the computer model;

generating, by the SoBI computing tool, a plurality of sub-groups for each corresponding grouping of values, wherein each sub-group comprises an individual value, or a sub-range of values less than a total number of values in the corresponding grouping of values, for the protected attribute;

analyzing, by the SoBI computing tool, each of the sub-groups in the plurality of sub-groups, based on at least one source of bias identification criterion, to identify one or more sources of bias in the input dataset, wherein each source of bias in the one or more sources of bias corresponds to a corresponding sub-group in the plurality of sub-groups; and

outputting, by the SoBI computing tool, a bias notification to an authorized computing device specifying the one or more sources of bias in the input dataset.

2. The method of claim 1 , wherein executing the bias detection operation to generate the plurality of separate groupings of values comprises:

generating a majority grouping having a majority grouping range of values for the protected attribute, wherein the majority grouping corresponds to entries in the input dataset associated with a privileged set of entities identified via the bias detection operation; and

generating a minority grouping having a minority grouping range of values for the protected attribute, wherein the minority grouping corresponds to entries in the input data set associated with a non-privileged set of entities identified via the bias detection operation, and wherein the plurality of sub-groups comprises a first set of sub-groups corresponding to the majority grouping and a second set of sub-groups corresponding to the minority grouping.

3. The method of claim 2 , wherein values for the protected attribute are numerical values, and wherein at least one of the first set of sub-groups or second set of sub-groups is generated by dividing the corresponding majority grouping range of values or corresponding minority grouping range of values into a plurality of sub-ranges of a same size.

4. The method of claim 2 , wherein values for the protected attribute are categorical values, and wherein at least one of the first set of sub-groups or second set of sub-groups is generated by dividing the corresponding majority grouping range of values or corresponding minority grouping range of values into individual category values.

5. The method of claim 1 , wherein the bias identification criterion comprises at least one predetermined relationship between a first fairness metric calculated for a grouping of values of the protected attribute, and a second fairness metric calculated for a sub-group of the grouping of values of the protected attribute, and wherein the predetermined relationship is evaluated for each combination of grouping of values and corresponding sub-groups of the grouping of values.

6. The method of claim 5 , wherein the at least one predetermined relationship comprises the second fairness metric being at least one of greater than the first fairness metric, or less than the first fairness metric, by a predetermined amount.

7. The method of claim 1 , wherein the bias identification criterion comprises a correspondence between a predetermined number of features identified in an explanation generated by a model interpretability tool executed on the computer model, and the protected attribute.

8. The method of claim 1 , further comprising:

performing, by the SoBI computing tool, a de-biasing operation on the input dataset to thereby generate a de-biased dataset, wherein the de-biasing operation comprises removing entries in the input data set that correspond to the one or more sources of bias to thereby generate the de-biased dataset comprising entries in the input dataset that do not correspond to the one or more sources of bias.

9. The method of claim 8 , further comprising:

executing, by the SoBI computing tool, a machine learning operation on the computer model to retrain the computer model using the de-biased dataset as a training dataset for the machine learning operation.

10. The method of claim 1 , wherein the computer model is a machine learning computer model trained by a machine learning process based on the input dataset being a training dataset for the machine learning computer model.

11. A computer program product comprising a computer readable storage medium having a computer readable program stored therein, wherein the computer readable program, when executed on a computing device, causes the computing device to implement a source of bias identification (SoBI) computing tool that identifies sources of bias in a dataset, and to:

execute, by a bias detection tool executing in the data processing system, a bias detection operation on results of an operation of a computer model, based on an input dataset, to generate a plurality of separate groupings of values for a protected attribute corresponding to a detected bias in the operation of the computer model;

generate, by the SoBI computing tool, a plurality of sub-groups for each corresponding grouping of values, wherein each sub-group comprises an individual value, or a sub-range of values less than a total number of values in the corresponding grouping of values, for the protected attribute;

analyze, by the SoBI computing tool, each of the sub-groups in the plurality of sub-groups, based on at least one source of bias identification criterion, to identify one or more sources of bias in the input dataset, wherein each source of bias in the one or more sources of bias corresponds to a corresponding sub-group in the plurality of sub-groups; and

output, by the SoBI computing tool, a bias notification to an authorized computing device specifying the one or more sources of bias in the input dataset.

12. The computer program product of claim 11 , wherein the computer readable program further causes the computing device to execute the bias detection operation to generate the plurality of separate groupings of values at least by:

generating a majority grouping having a majority grouping range of values for the protected attribute, wherein the majority grouping corresponds to entries in the input dataset associated with a privileged set of entities identified via the bias detection operation; and

generating a minority grouping having a minority grouping range of values for the protected attribute, wherein the minority grouping corresponds to entries in the input data set associated with a non-privileged set of entities identified via the bias detection operation, and wherein the plurality of sub-groups comprises a first set of sub-groups corresponding to the majority grouping and a second set of sub-groups corresponding to the minority grouping.

13. The computer program product of claim 12 , wherein values for the protected attribute are numerical values, and wherein at least one of the first set of sub-groups or second set of sub-groups is generated by dividing the corresponding majority grouping range of values or corresponding minority grouping range of values into a plurality of sub-ranges of a same size.

14. The computer program product of claim 12 , wherein values for the protected attribute are categorical values, and wherein at least one of the first set of sub-groups or second set of sub-groups is generated by dividing the corresponding majority grouping range of values or corresponding minority grouping range of values into individual category values.

15. The computer program product of claim 11 , wherein the bias identification criterion comprises at least one predetermined relationship between a first fairness metric calculated for a grouping of values of the protected attribute, and a second fairness metric calculated for a sub-group of the grouping of values of the protected attribute, and wherein the predetermined relationship is evaluated for each combination of grouping of values and corresponding sub-groups of the grouping of values.

16. The computer program product of claim 15 , wherein the at least one predetermined relationship comprises the second fairness metric being at least one of greater than the first fairness metric, or less than the first fairness metric, by a predetermined amount.

17. The computer program product of claim 11 , wherein the bias identification criterion comprises a correspondence between a predetermined number of features identified in an explanation generated by a model interpretability tool executed on the computer model, and the protected attribute.

18. The computer program product of claim 11 , wherein the computer readable program further causes the computing device to:

perform, by the SoBI computing tool, a de-biasing operation on the input dataset to thereby generate a de-biased dataset, wherein the de-biasing operation comprises removing entries in the input data set that correspond to the one or more sources of bias to thereby generate the de-biased dataset comprising entries in the input dataset that do not correspond to the one or more sources of bias.

19. The computer program product of claim 18 , wherein the computer readable program further causes the computing device to:

execute, by the SoBI computing tool, a machine learning operation on the computer model to retrain the computer model using the de-biased dataset as a training dataset for the machine learning operation.

20. An apparatus comprising:

a processor; and

a memory coupled to the processor, wherein the memory comprises instructions which, when executed by the processor, cause the processor to implement a source of bias identification (SoBI) computing tool that identifies sources of bias in a dataset, and to:

execute, by a bias detection tool executing in the data processing system, a bias detection operation on results of an operation of a computer model, based on an input dataset, to generate a plurality of separate groupings of values for a protected attribute corresponding to a detected bias in the operation of the computer model;

generate, by the SoBI computing tool, a plurality of sub-groups for each corresponding grouping of values, wherein each sub-group comprises an individual value, or a sub-range of values less than a total number of values in the corresponding grouping of values, for the protected attribute;

analyze, by the SoBI computing tool, each of the sub-groups in the plurality of sub-groups, based on at least one source of bias identification criterion, to identify one or more sources of bias in the input dataset, wherein each source of bias in the one or more sources of bias corresponds to a corresponding sub-group in the plurality of sub-groups; and

output, by the SoBI computing tool, a bias notification to an authorized computing device specifying the one or more sources of bias in the input dataset.