AI 용어집
인공지능 완전 사전
Measurement bias
Systematic inaccuracy in the collection or recording of variables that differs between demographic groups, affecting the quality and fairness of predictions.
Disparate impact analysis
Quantitative methodology for evaluating the disproportionate effects of algorithmic decisions on different demographic groups to identify and correct discriminations.
Causal attribution
Set of statistical techniques aiming to distinguish correlations from causal relationships in training data to avoid reproducing systemic injustices.
Individual fairness metrics
Evaluation criteria ensuring that similar individuals receive similar treatments, opposing purely group-based approaches to algorithmic fairness.
Representation bias
Systematic distortion where certain groups are underrepresented or misrepresented in training data, leading to biased predictions for these populations.
Inverse weighting
Bias correction technique assigning higher weights to examples from underrepresented groups to balance the influence of all demographic groups in training.
Balanced resampling
Data preprocessing method modifying the distribution of the training set to ensure fair representation of all demographic groups.
Adversarial debiasing
Machine learning approach using adversarial networks to learn data representations invariant to sensitive characteristics while preserving predictive performance.
Algorithm Audit
Systematic and independent examination of an AI system to identify, document, and quantify potential biases and discriminations in its decisions.