Enhancing fairness in transfer learning for machine learning models with missing protected attributes in source or target domains

Patent metadata

Field	Value
Publication number	US-11443236-B2
Application number	US-201916692974-A
Country	US
Kind code	B2
Filing date	Nov 22, 2019
Priority date	Nov 22, 2019
Publication date	Sep 13, 2022
Grant date	Sep 13, 2022

Abstract

A method of utilizing a computing device to correct source data used in machine learning includes receiving, by the computing device, first data. The computing device corrects the source data via an application of a covariate shift to the source data based upon the first data where the covariate shift re-weighs the source data.

First claim

What is claimed is:

1. A method of utilizing a computing device to correct source data used in machine learning, the method comprising:
receiving, by a computing device, first data;
providing, by the computing device, transfer learning processing in absence of one or more protected attributes using a covariate shift combined with re-weighing to reduce a difference in group-specific prevalences for the source data;
wherein the source data is partially labeled data, the first data is target data that includes the one or more protected attributes, and the transfer learning processing uses a target-fair covariate shift:
wherein the target-fair covariate shift uses weights that are chosen to minimize a linear combination of a fairness loss with a classification loss; and
wherein the fairness loss is evaluated on the target data where the one or more protected attributes are available.

2. The method of claim 1 , further comprising:
training, by the computing device one or more machine learning models using the re-weighed source data.

3. The method of claim 1 , wherein the source data is fully labeled data having the one or more protected attributes, the first data is target data, and the transfer learning processing uses a prevalence-constrained covariate shift.

4. The method of claim 3 , wherein the prevalence-constrained covariate shift uses learned weights based on a difference as compared to covariate shift weights subject to constraints on weighted prevalences.

5. A computer program product for correcting source data used in machine learning, the computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions executable by a processor to cause the processor to:
receive, by the processor, first data;
provide, by the processor, transfer learning processing in absence of one or more protected attributes using a covariate shift combined with re-weighing to reduce a difference in group-specific prevalences for the source data;
wherein the source data is partially labeled data, the first data is target data that includes the one or more protected attributes, and the transfer learning processing uses a target-fair covariate shift;
wherein the target-fair covariate shift uses weights that are chosen to minimize a linear combination of a fairness loss with a classification loss; and
wherein the fairness loss is evaluated on the target data where the one or more protected attributes are available.

6. The computer program product of claim 5 , wherein the program instructions executable by the processor further causes the processor to:
train, by the processor, one or more machine learning models using the re-weighed source data.

7. The computer program product of claim 5 , wherein the source data is fully labeled data having the one or more protected attributes, the first data is target data, and the transfer learning processing uses a prevalence-constrained covariate shift.

8. The computer program product of claim 7 , wherein the prevalence-constrained covariate shift uses learned weights that are chosen based on a difference as compared to covariate shift weights subject to constraints on weighted prevalences.

9. An apparatus comprising:
a memory configured to store instructions; and
a processor configured to execute the instructions to:
receive first data;
provide transfer learning processing in absence of one or more protected attributes using a covariate shift combined with re-weighing to reduce a difference in group-specific prevalences for a source data;
wherein the source data is partially labeled data, the first data is target data that includes the one or more protected attributes, and the transfer learning processing uses a target-fair covariate shift; and
wherein the target-fair covariate shift uses weights that are chosen to minimize a linear combination of a fairness loss with a classification loss, and the fairness loss is evaluated on the target data where the one or more protected attributes are available.

10. The apparatus of claim 9 , wherein the processor is configured to further execute the instructions to:
train one or more machine learning models using the re-weighed source data.

11. The apparatus of claim 9 , wherein the source data is fully labeled data having the one or more protected attributes, the first data is target data, and the transfer learning processing uses a prevalence-constrained covariate shift.

12. The apparatus of claim 11 , wherein the prevalence-constrained covariate shift uses learned weights that are chosen based on a difference as compared to covariate shift weights subject to constraints on weighted prevalences.

Assignees

• IBM

Inventors

• Natesan Ramamurthy Karthikeyan
• Coston Amanda
• Wei Dennis
• Varshney Kush Raj
• Speakman Skyler
• Mustahsan Zairah
• Chakraborty Supriyo

Classifications

• G06N20/00 Primary

  Machine learning · CPC title