Efficiency and Fairness Dual Focus: The Neyman-Pearson Approach

Project: Research project

Project Details

Description

In the present era, the utilization of AI in decision-making processes has become an increasing practice. However, recent reports have drawn attention to the potential unequal treatment by AI-driven decisions to different social groups. Moreover, recent laws and regulations also enforce fair treatment by automatic decisions. Consequently, the concept of algorithmic fairness has gained significant attention. From a statistical point of view, the goal of algorithmic fairness is to ensure equitable outcomes for all social groups involved in machine learning classification tasks. Starting from this perspective, we aim to tackle two problems in this project: firstly, how can institutions like banks and companies find a balance between prioritizing algorithmic fairness and maintaining the efficiency that they seek by using statistical learning; and secondly, how can fair and efficient classifiers be trained in a scenario where label noise occurs because of biases in the existing datasets as a result of historical inequalities. Addressing these complex challenges is of utmost significance in order to establish decision-making processes in AI that are both fair and efficient.

To address the first question regarding the dual focus on efficiency and fairness, we introduce the Neyman-Pearson (NP) classification paradigm, which measures institutional efficiency in industries like finance and medical diagnosis. Moreover, as there are various fairness criteria, each serving a specific purpose or scenario, we propose a comprehensive framework that combines the NP classification with any algorithmic fairness criterion. Additionally, we present a classifier-training algorithm that can be applied to any classification method, such as logistic regression, ensuring that the resulting classifier meets the requirements of both efficiency and fairness with high probability.

In addressing the second question regarding training under label noise, we propose a robust method that can also be adapted to any classification method. Specifically, we utilize the median-of-means method, a robust statistical estimation technique, to mitigate the impact of label noise during the training process of the fair NP classifier. This method is capable of handling arbitrary types of label noise, as long as the proportion of mislabeled data remains low. By incorporating the median-of-means method, our proposed approach ensures robustness in the presence of label noise, allowing for more accurate and reliable classification outcomes for fairness-efficiency dual focus.
StatusNot started
Effective start/end date1/01/2531/12/27

Fingerprint

Explore the research topics touched on by this project. These labels are generated based on the underlying awards/grants. Together they form a unique fingerprint.