dc.description.abstract |
In a world where machine learning and AI play an increasing role in decision-making across various sectors, concerns about fairness have emerged. This report delves into the journey of understanding fairness in machine learning, exploring its origins, current challenges, and the critical importance of addressing biases. The report investigates real-world instances of bias, such as racial and gender biases in hiring algorithms, biased facial recognition systems, and discriminatory outcomes in college admissions. It scrutinizes the reasons behind bias, emphasizing the impact of big data’s subjective nature and inherent biases in training data. Various fairness metrics, including Unawareness, Demographic Parity, Equalized Odds, and Predictive Rate Parity, are discussed. The paper then explores fairness issues in the context of college admissions in the United States, applying these metrics to analyze the biases present. Additionally, the Impossibility Theorem of Fairness is introduced, highlighting the inherent challenges of simultaneously achieving demographic parity, predictive parity, and equalized odds. The concept of Rawlsian Equality of Opportunity is presented, examining utility distribution based on circumstances and choices. The document also introduces AIF360, a toolkit for fairness in machine learning, using the COMPAS algorithm as a case study. The architectural overview, bias mitigation approaches, and algorithms employed by AIF360 are discussed. |
en_US |