Learning Mahalanobis distance metrics from data

Patent metadata

Field	Value
Publication number	US-12287848-B2
Application number	US-202117345730-A
Country	US
Kind code	B2
Filing date	Jun 11, 2021
Priority date	Jun 11, 2021
Publication date	Apr 29, 2025
Grant date	Apr 29, 2025

Abstract

The present invention provides techniques for learning Mahalanobis distance similarity metrics from data for individually fair machine learning models. In one aspect, a method for learning a fair Mahalanobis distance similarity metric includes: obtaining data with similarity annotations; selecting, based on the data obtained, a model for learning a Mahalanobis covariance matrix Σ; and learning the Mahalanobis covariance matrix Σ from the data using the model selected, wherein the Mahalanobis covariance matrix Σ fully defines the fair Mahalanobis distance similarity metric.

First claim

What is claimed is:

1. A method for improving algorithmic fairness of machine learning models using learned fair Mahalanobis distance similarity metrics, the method comprising:
obtaining training data comprising similarity annotations;
determining one model out of a plurality of models to use in learning a Mahalanobis covariance matrix Σ based on the obtained training data;
learning the Mahalanobis covariance matrix Σ from the obtained training data using the determined one model, wherein the Mahalanobis covariance matrix Σ represents a fair Mahalanobis distance similarity metric; and
training one or more machine learning models using, at least in part, the fair Mahalanobis distance similarity metric, for one or more machine learning model tasks.

2. The method of claim 1 , wherein the fair Mahalanobis distance similarity metric is of a form:
d x ( x 1 ,x 2 )@ φ( x 1 )−φ( x 2 )Σφ( x 1 )−φ( x 2 )) ,

wherein φ(x):X→R d is an embedding map and Σ∈S + d .

3. The method of claim 1 , wherein the obtained training data comprises groups of comparable samples.

4. The method of claim 3 , wherein the determined one model comprises a factor model.

5. The method of claim 4 , wherein the factor model comprises:
φ i =A * u i +B * υ i +ò i ,

wherein φ i ∈R d is a learned representation of x i , u i ∈R K is a sensitive attribute of x i for a task at hand, υ i ∈R L is a relevant attribute of x i for the task at hand, and ò i is an error term.

6. The method of claim 5 , further comprising:
choosing an orthogonal complement of ran(A * ) for the Mahalanobis covariance matrix Σ, wherein ran (A * ) is a column space of A * ; and
solving for ran (A * ).

7. The method of claim 1 , wherein the obtained training data comprises pairs of samples that are comparable, incomparable, or combinations thereof.

8. The method of claim 7 , wherein the determined one model comprises a binary response model.

9. The method of claim 8 , wherein the data comprises human user feedback in a form of triplets {(x i1 , x i2 , y i )} i=1 n , where y i ∈{0,1} indicates whether a human user considers x i1 and x i2 comparable, wherein (x i1 , x i2 , y i ) satisfies the binary response model:

y
i

❘
"\[RightBracketingBar]"

⁢

x

i
⁢
1

,

x

i
⁢
2

:

Ber
⁡
(

2
⁢

σ
⁡
(

-

d
i

)

)

,

(

d
i

@



φ

i
⁢
1

-

φ

i
⁢
2



)

∑
0

2

=

(

φ

i
⁢
1

-

φ

i
⁢
2

)

T

⁢

∑

0

⁢

(

φ

i
⁢
1

-

φ

i
⁢
2

)

=

〈

(

φ

i
⁢
1

-

φ

i
⁢
2

)

⁢

(

φ

i
⁢

Assignees

• IBM
• Univ Michigan

Inventors

• Yurochkin Mikhail
• Mukherjee Debarghya
• Banerjee Moulinath
• Sun Yuekai
• Upadhyay Sohini

Classifications

• G06N3/09

  Supervised learning · CPC title

• G06F18/285 Primary

  Selection of pattern recognition techniques, e.g. of classifiers in a multi-classifier system · CPC title

• G06F18/214

  Generating training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G06F18/40

  Software arrangements specially adapted for pattern recognition, e.g. user interfaces or toolboxes therefor · CPC title

• G06N3/08

  Learning methods · CPC title