# US 2023/0267362 A1 · Automating bias evaluation for machine learning projects

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2023/0267362 A1
- Document ID: US20230267362A1
- Status: Published application
- Kind: A1
- Application number: 17/652,268
- Filed: 2022-02-23
- Published: 2023-08-24
- Earliest claimed: 2022-02-23
- Assignees: AT&T Intellectual Property I, L.P.
- Applicants: At&T Intellectual Property I, L.P.
- Inventors: Noemi Derzsy; Balachander Krishnamurthy
- Web page: https://patentgrep.com/patent/US20230267362A1

## CPC

- G06N 20/00: Machine learning
- G06N 5/022: Knowledge engineering; Knowledge acquisition
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

A method includes obtaining descriptive information for a first machine learning project, identifying, based on the descriptive information, a plurality of past machine learning projects which are similar to the first machine learning project, retrieving digital documents that describe the bias evaluation pipelines that were used to evaluate the plurality of past machine learning projects, detecting a common bias evaluation pipeline step among at least a subset of the digital documents, extracting, from the subset, a snippet of machine-executable code that corresponds to the common bias evaluation pipeline step, modifying the snippet of machine-executable code with use case data that is specific to the first machine learning project to generate modified machine-executable code, and generating a proposed bias evaluation pipeline for evaluating the first machine learning project, wherein the proposed bias evaluation pipeline includes the modified machine-executable code.

## Claims

### Claim 1

A method comprising:
obtaining, by a processing system including at least one processor, descriptive information for a first machine learning project;
identifying, by the processing system and based on the descriptive information, a plurality of past machine learning projects which are detected to be similar to the first machine learning project;
retrieving, by the processing system, a plurality of digital documents, wherein each digital document of the plurality of digital documents describes a bias evaluation pipeline that was used to evaluate a different past machine learning project of the plurality of past machine learning projects;
detecting, by the processing system, a common bias evaluation pipeline step among at least a subset of the plurality of digital documents;
extracting, by the processing system from the subset of the plurality of digital documents, a snippet of machine-executable code that corresponds to the common bias evaluation pipeline step;
modifying, by the processing system, the snippet of machine-executable code with use case data that is specific to the first machine learning project to generate modified machine-executable code; and
generating, by the processing system, a proposed bias evaluation pipeline for evaluating the first machine learning project, wherein the proposed bias evaluation pipeline includes the modified machine-executable code.

### Claim 2 (depends on 1)

The method of claim 1, wherein the descriptive information describes at least one of: a use case of the first machine learning project, input features of a machine learning model used in the first machine learning project, or a target output of the machine learning model.

### Claim 3 (depends on 1)

The method of claim 1, wherein the plurality of digital documents is stored with other digital documents corresponding to other machine learning projects which have been previously evaluated for bias in a repository that is accessible by the processing system.

### Claim 4 (depends on 1)

The method of claim 1, wherein the detecting comprises using a natural language processing technique to detect the common bias evaluation pipeline step.

### Claim 5 (depends on 1)

The method of claim 1, wherein the common bias evaluation pipeline step is one step in a series of bias evaluation pipeline steps that is common to the subset of the plurality of digital documents.

### Claim 6 (depends on 1)

The method of claim 1, wherein the common bias evaluation pipeline step is one step in a pattern that is common to the subset of the plurality of digital documents.

### Claim 7 (depends on 1)

The method of claim 1, wherein the detecting comprises detecting a plurality of common bias evaluation pipeline steps and ranking the plurality of common bias evaluation steps.

### Claim 8 (depends on 7)

The method of claim 7, wherein the plurality of common bias evaluation pipeline steps is ranked in order from most frequently appearing to least frequently appearing.

### Claim 9 (depends on 8)

The method of claim 8, wherein the common bias evaluation pipeline step is included within a set of n top-ranked common bias evaluation pipeline steps of the plurality of common bias evaluation steps.

### Claim 10 (depends on 9)

The method of claim 9, wherein the extracting further comprises extracting a bias evaluation pipeline step that is not included within the set of n top-ranked common bias evaluation pipeline steps.

### Claim 11 (depends on 1)

The method of claim 1, wherein the use case data comprises a feature intended to be provided as an input to a machine learning model used in the first machine learning project.

### Claim 12 (depends on 1)

The method of claim 1, wherein the proposed bias evaluation pipeline incorporates a suggestion made by a human user.

### Claim 13 (depends on 12)

The method of claim 12, wherein the suggestion is at least one of: a modification to the snippet of machine-executable code or an additional snippet of machine-executable code to include in the proposed bias evaluation pipeline.

### Claim 14 (depends on 1)

The method of claim 1, further comprising:
performing, by the processing system, an evaluation of the first machine learning project for bias, using the proposed bias evaluation pipeline.

### Claim 15 (depends on 14)

The method of claim 14, wherein the performing generates an output that indicates at least one portion of the first machine learning project where a potential bias was detected.

### Claim 16 (depends on 15)

The method of claim 15, wherein the output further includes a suggestion for minimizing the potential bias that was detected.

### Claim 17 (depends on 16)

The method of claim 16, wherein the suggestion is based on an adjustment that was made to at least one machine learning project of the subset of the plurality of past machine learning projects.

### Claim 18 (depends on 15)

The method of claim 15, further comprising:
creating, by the processing system, a digital document for the first machine learning project that includes the proposed bias evaluation pipeline, machine-executable code corresponding to the proposed bias evaluation pipeline, and the output; and
storing, by the processing system, the digital document for the first machine learning model in a repository with the plurality of digital documents.

### Claim 19

A non-transitory computer-readable medium storing instructions which, when executed by a processing system including at least one processor, cause the processing system to perform operations, the operations comprising:
obtaining descriptive information for a first machine learning project;
identifying, based on the descriptive information, a plurality of past machine learning projects which are detected to be similar to the first machine learning project;
retrieving a plurality of digital documents, wherein each digital document of the plurality of digital documents describes a bias evaluation pipeline that was used to evaluate a different past machine learning project of the plurality of past machine learning projects;
detecting a common bias evaluation pipeline step among at least a subset of the plurality of digital documents;
extracting, from the subset of the plurality of digital documents, a snippet of machine-executable code that corresponds to the common bias evaluation pipeline step;
modifying the snippet of machine-executable code with use case data that is specific to the first machine learning project to generate modified machine-executable code; and
generating a proposed bias evaluation pipeline for evaluating the first machine learning project, wherein the proposed bias evaluation pipeline includes the modified machine-executable code.

### Claim 20

A device comprising:
a processing system including at least one processor; and
a non-transitory computer-readable medium storing instructions which, when executed by the processing system, cause the processing system to perform operations, the operations comprising:
obtaining descriptive information for a first machine learning project;
identifying, based on the descriptive information, a plurality of past machine learning projects which are detected to be similar to the first machine learning project;
retrieving a plurality of digital documents, wherein each digital document of the plurality of digital documents describes a bias evaluation pipeline that was used to evaluate a different past machine learning project of the plurality of machine learning projects;
detecting a common bias evaluation pipeline step among at least a subset of the plurality of digital documents;
extracting, from the subset of the plurality of digital documents, a snippet of machine-executable code that corresponds to the common bias evaluation pipeline step;
modifying the snippet of machine-executable code with use case data that is specific to the first machine learning project to generate modified machine-executable code; and
generating a proposed bias evaluation pipeline for evaluating the first machine learning project, wherein the proposed bias evaluation pipeline includes the modified machine-executable code.
