1. Official Vendor Documentation (AWS): Amazon SageMaker Clarify is a tool designed for this purpose. The documentation states
"Amazon SageMaker Clarify helps improve your machine learning models by detecting potential bias and helping explain how these models make predictions." It supports both pre-training bias analysis on data (aligning with option C) and post-training bias analysis using fairness metrics (aligning with option A).
Source: AWS Documentation
"Amazon SageMaker Clarify
" Section: "Detect Bias in Data and Models."
2. Official Vendor Documentation (AWS): The AWS framework for Responsible AI emphasizes fairness as a key dimension. It states
"Fairness is the dimension of Responsible AI that is concerned with how an AI system's outputs impact different individuals or subgroups of people... Mitigation can occur at different stages of the AI/ML lifecycle
including in the data collection and preparation phase..."
Source: AWS Responsible AI
"Fairness" dimension description.
3. Academic Publication: In the survey paper "A Survey on Bias and Fairness in Machine Learning
" the authors categorize bias mitigation techniques into three families: pre-processing
in-processing
and post-processing. Modifying training data (C) is a primary pre-processing technique
and using fairness metrics (A) is essential for evaluation across all stages.
Source: Mehrabi
N.
Morstatter
F.
Saxena
N.
Lerman
K.
& Galstyan
A. (2021). "A Survey on Bias and Fairness in Machine Learning." ACM Computing Surveys
54(6)
Article 115
Section 4: "Bias Mitigation". DOI: https://doi.org/10.1145/3457607