Core Concepts
The authors propose a framework to model the trade-off between accuracy and fairness in data-restricted binary classification scenarios, analyzing various constraints and their impact on the Bayesian classifier's decisions.
Abstract
The paper introduces a framework to balance accuracy and fairness in machine learning models under data restrictions. It explores different scenarios, constraints, and their effects on decision-making. The study emphasizes the importance of considering both accuracy and fairness in classification tasks.
The content discusses the challenges of bias in ML models, presents mathematical definitions for fairness, analyzes the trade-off between accuracy and fairness, proposes optimization problems to model decisions by the Bayesian classifier, and conducts experiments on three datasets to quantify trade-offs among different fairness notions.
Key points include proposing a framework for modeling trade-offs between accuracy and fairness, analyzing group and individual fairness definitions, formulating optimization problems for fair classifiers, conducting experiments on real datasets to validate the framework's utility.
The study highlights the complexity of balancing fairness and accuracy in machine learning models when faced with data restrictions. It provides insights into how different constraints impact decision-making processes and offers a practical tool for quantifying trade-offs among various fairness notions.
Stats
Applications dealing with sensitive information may restrict data available to ML classifiers.
The proposed framework aims to model the trade-off between accuracy and fairness under different scenarios.
Experiments conducted on three datasets demonstrate the utility of the framework for quantifying trade-offs among different fairness notions.