AI 용어집
인공지능 완전 사전
Fair resampling
Preprocessing technique that modifies the training data distribution by over-representing minority or underrepresented groups to reduce algorithmic disparities in model predictions.
Inverse probability weighting
Bias correction method assigning weights to training examples inversely proportional to their frequency in the population, thus compensating for demographic group imbalance.
Adversarial learning for fairness
Approach of simultaneously training a main predictor and an adversary trying to predict sensitive attributes, forcing the main model to generate representations invariant to protected characteristics.
Fair prediction calibration
Post-processing technique adjusting prediction scores to ensure predicted probabilities correspond consistently to observed frequencies across different demographic groups.
Constrained fairness optimization
Training method incorporating mathematical constraints on fairness metrics directly into the objective function, ensuring fairness criteria are met during model optimization.
Optimized equalized odds
Processing technique ensuring equal true positive rates between groups while maximizing overall performance, often implemented through specific loss functions or threshold adjustments.
Adjusted demographic parity
Correction method ensuring positive predictions are distributed proportionally across different demographic groups, regardless of their intrinsic characteristics.
Causal debiasing
Approach using causal graphs to identify and neutralize causal paths introducing bias, preserving only relevant relationships for the prediction task.
Group invariance learning
Training technique forcing the model to learn representations invariant to variations between demographic groups while preserving information relevant to the main task.
Post-hoc correction through adaptive thresholds
Method applied after training that dynamically adjusts decision thresholds by group to balance performance metrics and ensure fairness in final predictions.
Disparity reduction through reweighting
Preprocessing technique that recalculates the weights of training instances to minimize statistical divergence between the observed distribution and a fair target distribution.
Fair feature masking
Processing strategy that selectively masks or transforms potentially biased features during training to force the model to rely on non-discriminatory attributes.
Selection bias correction
Set of techniques identifying and compensating for distortions introduced by non-random sampling processes that systematically favor certain subgroups of the population.
Robust learning against fairness attacks
Training methodology incorporating adversarial examples designed to amplify biases, thereby strengthening the model's resistance against manipulations aimed at degrading its fairness.
Debiasing through counterfactuals
Technique generating counterfactual examples by modifying sensitive attributes to train the model to produce predictions invariant to changes in these protected characteristics.
Distribution balancing through optimal transport
Advanced method using optimal transport theory to transform the data distribution of a minority group to bring it closer to that of the majority group, thereby reducing systemic biases.
Fair regularization by divergence
Training technique adding a penalty term based on divergence measures (KL, JS, Wasserstein) between the prediction distributions of different groups to ensure statistical fairness.