bias and fairness编辑历史

# 1 Overview of Bias and Fairness

## 1.1 Definitions and Scope

### 1.1.1 Bias in Data vs. Bias in Algorithms

Bias in data refers to systematic errors or skews present in the datasets used to train machine learning models. These can arise from historical prejudices, unrepresentative sampling, or flawed measurement processes. Bias in algorithms, by contrast, emerges from the model’s design, optimization, or decision logic—for example, when a model learns spurious correlations or when its objective function inadvertently rewards discriminatory outcomes. Distinguishing between these two sources is critical for selecting appropriate mitigation strategies.

### 1.1.2 Fairness as a Sociotechnical Concept

Fairness in information technology is not a purely mathematical property but a sociotechnical concept that must account for legal, ethical, and cultural norms. Technical definitions of fairness (such as demographic parity or equal opportunity) operationalize different intuitions about what it means to treat individuals or groups equitably, yet no single definition is universally applicable. The context of deployment—including the domain, stakeholders, and potential harms—shapes which fairness criteria are most appropriate.

## 1.2 Historical Context

### 1.2.1 Early Cases of Algorithmic Discrimination

Concerns about bias in automated systems date back to the 1970s and 1980s, notably with the use of statistical models in credit scoring and hiring that inadvertently penalized minority groups. Early examples include the denial of mortgage loans based on zip codes that correlated with race and the use of personality tests that disadvantaged non‑native English speakers. These cases highlighted the need for auditing mechanisms and oversight.

### 1.2.2 Rise of Fair Machine Learning (2010s–Present)

The modern field of fair machine learning began to coalesce around 2010, driven by high‑profile incidents such as gender‑biased word embeddings, racially biased p