Thuật ngữ AI
Từ điển đầy đủ về Trí tuệ nhân tạo
Equalized Odds
Fairness criterion stating that a model's true positive rates and false positive rates should be equal across different demographic groups, ensuring uniform predictive performance.
Fairness Audit
Systematic process of evaluating an AI model to identify, quantify, and document biases and performance disparities between groups, using fairness metrics and statistical analyses.
Debiasing
Set of techniques aimed at reducing or eliminating biases in training data, algorithms, or model predictions, including pre-processing, in-processing, and post-processing methods.
Sensitive Attributes
Demographic variables or personal characteristics (such as race, ethnicity, gender, age, religion) that are protected by law and whose use in automated decision-making may lead to discrimination.
Fairness Score
Aggregate metric that quantifies the overall level of fairness of a model by combining multiple bias indicators (such as demographic parity or equal opportunity) into a single comparable value.
Fair Reweighing
Pre-processing technique that adjusts the weights of training examples for different demographic groups to correct imbalances and satisfy specific fairness criteria such as demographic parity.
Fairness Robustness Test
Evaluation that checks whether a model's fairness metrics remain stable in the face of slight data perturbations or changes in population distribution, ensuring durable fairness.