US 11,531,845 B1Granted

Bias mitigating machine learning training system

AssigneeSAS Institute Inc.

InventorsXin Jiang Hunt, Xinmin Wu, Ralph Walter Abbey

Life of this patent

Estimating the term…

Loading patent…

Abstract

A computing device trains a fair machine learning model. A prediction model is trained to predict a target value. For a number of iterations, a weight vector is computed using the bound value based on fairness constraints defined for a fairness measure type; a weight value is assigned to each observation vector based on the target value and a sensitive attribute value; the prediction model is trained with each weighted observation vector to predict the target value; and a conditional moments vector is computed based on the fairness constraints and the target and sensitive attribute values. Conditional moments difference values are computed. When the conditional moments difference values indicate to adjust the bound value, the bound value is updated and the process is repeated with the bound value replaced with the updated bound value until the conditional moments difference values indicate no further adjustment of the bound value is needed.

Claims30

Dependent claims

A non-transitory computer-readable medium having stored thereon computer-readable instructions that when executed by a computing device cause the computing device to:

(A) train a prediction model with a plurality of observation vectors, wherein each observation vector of the plurality of observation vectors includes a target variable value, a sensitive attribute variable value, and a plurality of attribute variable values;
(B) predict a second target variable value for each observation vector of the plurality of observation vectors using the trained prediction model;
(C) initialize a bound value using a predefined bound value;
(D) initialize a number of iterations;
(E) compute a weight vector that includes a plurality of weight values using the bound value, wherein each weight value is computed based on fairness constraints defined for a predefined fairness measure type;
(F) assign a weight value from the computed weight vector to each observation vector of the plurality of observation vectors based on the predicted second target variable value and the sensitive attribute variable value of each respective observation vector of the plurality of observation vectors;
(G) train the prediction model with each observation vector of the plurality of observation vectors weighted by the respective assigned weight value;
(H) predict the second target variable value for each observation vector of the plurality of observation vectors using the prediction model trained in (G);
(I) compute a conditional moments vector based on the fairness constraints and the second target variable value predicted in (H) and the sensitive attribute variable value of each respective observation vector of the plurality of observation vectors;
(J) increment the initialized number of iterations;
(K) repeat (E) through (J) until a predefined number of bound test update iterations is performed based on the incremented number of iterations;
(L) compute conditional moments difference values based on the fairness constraints;
(M) when the computed conditional moments difference values indicate to adjust the bound value, update the bound value and repeat (D) through (M) with the bound value replaced with the updated bound value until the computed conditional moments difference values indicate that no further adjustment of the bound value is needed;
(N) store the updated bound value as a first bound value and repeat (A) through (M);
(O) after (N), compute the updated bound value from the stored first bound value and the updated bound value determined from the repeat of (A) through (M);
(P) train a fair prediction model with the updated bound value computed in (O);
and
(Q) output the fair prediction model trained in (P).

15 more elements

A computing device comprising:

a processor;
and
a non-transitory computer-readable medium operably coupled to the processor, the computer-readable medium having computer-readable instructions stored thereon that, when executed by the processor, cause the computing device to
(A) train a prediction model with a plurality of observation vectors, wherein each observation vector of the plurality of observation vectors includes a target variable value, a sensitive attribute variable value, and a plurality of attribute variable values;
(B) predict a second target variable value for each observation vector of the plurality of observation vectors using the trained prediction model;
(C) initialize a bound value using a predefined bound value;
(D) initialize a number of iterations;
(E) compute a weight vector that includes a plurality of weight values using the bound value, wherein each weight value is computed based on fairness constraints defined for a predefined fairness measure type;
(F) assign a weight value from the computed weight vector to each observation vector of the plurality of observation vectors based on the predicted second target variable value and the sensitive attribute variable value of each respective observation vector of the plurality of observation vectors;
(G) train the prediction model with each observation vector of the plurality of observation vectors weighted by the respective assigned weight value;
(H) predict the second target variable value for each observation vector of the plurality of observation vectors using the prediction model trained in (G);
(I) compute a conditional moments vector based on the fairness constraints and the second target variable value predicted in (H) and the sensitive attribute variable value of each respective observation vector of the plurality of observation vectors;
(J) increment the initialized number of iterations;
(K) repeat (E) through (J) until a predefined number of bound test update iterations is performed based on the incremented number of iterations;
(L) compute conditional moments difference values based on the fairness constraints;
(M) when the computed conditional moments difference values indicate to adjust the bound value, update the bound value and repeat (D) through (M) with the bound value replaced with the updated bound value until the computed conditional moments difference values indicate that no further adjustment of the bound value is needed;
(N) store the updated bound value as a first bound value and repeat (A) through (M);
(O) after (N), compute the updated bound value from the stored first bound value and the updated bound value determined from the repeat of (A) through (M);
(P) train a fair prediction model with the updated bound value computed in (O);
and
(Q) output the fair prediction model trained in (P).

18 more elements

A method of training a fair machine learning model, the method comprising:

(A) training, by a computing device, a prediction model with a plurality of observation vectors, wherein each observation vector of the plurality of observation vectors includes a target variable value, a sensitive attribute variable value, and a plurality of attribute variable values;
(B) predicting, by the computing device, a second target variable value for each observation vector of the plurality of observation vectors using the trained prediction model;
(C) initializing, by the computing device, a bound value using a predefined bound value;
(D) initializing, by the computing device, a number of iterations;
(E) computing, by the computing device, a weight vector that includes a plurality of weight values using the bound value, wherein each weight value is computed based on fairness constraints defined for a predefined fairness measure type;
(F) assigning, by the computing device, a weight value from the computed weight vector to each observation vector of the plurality of observation vectors based on the predicted second target variable value and the sensitive attribute variable value of each respective observation vector of the plurality of observation vectors;
(G) training, by the computing device, the prediction model with each observation vector of the plurality of observation vectors weighted by the respective assigned weight value;
(H) predicting, by the computing device, the second target variable value for each observation vector of the plurality of observation vectors using the prediction model trained in (G);
(I) computing, by the computing device, a conditional moments vector based on the fairness constraints and the second target variable value predicted in (H) and the sensitive attribute variable value of each respective observation vector of the plurality of observation vectors;
(J) incrementing, by the computing device, the initialized number of iterations;
(K) repeating, by the computing device, (E) through (J) until a predefined number of bound test update iterations is performed based on the incremented number of iterations;
(L) computing, by the computing device, conditional moments difference values based on the fairness constraints;
(M) when the computed conditional moments difference values indicate to adjust the bound value, updating, by the computing device, the bound value and repeating, by the computing device, (D) through (M) with the bound value replaced with the updated bound value until the computed conditional moments difference values indicate that no further adjustment of the bound value is needed;
(N) storing, by the computing device, the updated bound value as a first bound value and repeating, by the computing device, (A) through (M);
(O) after (N), computing, by the computing device, the updated bound value from the stored first bound value and the updated bound value determined from the repeat of (A) through (M);
(P) training, by the computing device, a fair prediction model with the updated bound value computed in (O);
and
(Q) outputting, by the computing device, the fair prediction model trained in (P).

15 more elements

Dependencies are inferred from the claim text.

Facts

Application: 17/837,444Filed 10 Jun 2022 · continuation in part
Priority: 6 Oct 2021
Granted: 20 Dec 2022
Term ends: Not estimatedThe source records don’t support an estimate.

Classification4 codes

G06F 18/2148
Generating training patterns; Bootstrap methods, e.g. bagging or boosting: characterised by the process organisation or structure, e.g. boosting…
G06N 20/00
Machine learning
G06F
Computing and data processing
G06N
AI and machine learning

Family

Application63/252,918Us provisional application · filed 6 Oct 2021
Application63/272,980Us provisional application · filed 28 Oct 2021
Application17/557,298Parent of this continuation-in-part · filed 21 Dec 2021
US 11,436,444 B1Related patent
US 11,531,845 B1This patent

From the US filing history. Foreign family members aren’t listed.