Bias in Machine Learning

These flashcards cover the key concepts and terms related to bias in machine learning as discussed in the lecture.

Bias in Machine Learning

The tendency of machine learning algorithms to produce skewed results due to prejudiced data.

Compounded Bias

Initial bias that grows over time, leading to greater disparities; for example, in crime prediction leading to increased policing in biased locations.

Example of Bias in Algorithms

Bank loan algorithms influenced by previous human decisions, which may carry inherent biases.

Sample Size Disparity

The difference in data availability between majority and minority groups, affecting the predictive accuracy of ML models for underrepresented groups.

Proxies

Features in a dataset that may indirectly represent sensitive attributes (e.g., using zip code as a proxy for race or income).

Conditions of Fairness

Frameworks that define fairness in machine learning, including awareness, statistical parity, demographic parity, and individual fairness.

Statistical Parity

A measure of fairness that ensures equal positive and negative probabilities for both majority and minority groups.

Demographic Parity

Guarantees equal positive probability of outcomes between majority and minority groups in a model.

Individual Fairness

The principle that similar individuals should receive similar outcomes regardless of group affiliation.

Counterfactual Fairness

A property where the model's output does not change when the protected attribute is altered.

Predictive Parity

Ensures that models have equal precision across different groups.

Equalized Odds

The requirement that true positive rates (TPR) and false positive rates (FPR) should be equal across groups.

Disparate Treatment

Intentional discrimination against a group, which is less common in machine learning.

Disparate Impact

Unjustifiably adverse effects on a group, which may occur unintentionally.

Algorithmic Accountability

The concept of who is responsible for the outcomes and societal effects of algorithms.

Reprocessing Techniques in ML

Methods like massaging, reweighting, and sampling to mitigate bias in machine learning processes.

Adversarial Training

A training method in machine learning that involves training against adversarial examples to improve model robustness.