Bias

Bias in machine learning refers to systematic differences in model outcomes across demographic groups, which can arise from various sources, including training data, labeling processes, and deployment contexts. Mathematically, bias can be quantified using metrics such as demographic parity, equal opportunity, and disparate impact, which assess the fairness of predictions across different groups. Bias can originate from historical data reflecting societal prejudices, leading to models that perpetuate or exacerbate these inequalities. Techniques for mitigating bias include re-sampling, re-weighting, and adversarial debiasing, which aim to create more equitable models. Understanding and addressing bias is critical in the broader context of fairness in AI, as it directly impacts the ethical implications of deploying machine learning systems in real-world applications.