1/18
Flashcards about Imbalanced Data and Algorithmic Fairness
Name | Mastery | Learn | Test | Matching | Spaced |
---|
No study sessions yet.
overall error
Classifiers can be biased towards the majority class because they try to reduce the .
Oversampling, Undersampling, Create synthetic data
Three methods of Data Balancing are: __, __, and __.
oversampling
Random __: randomly duplicate data points from the minority class.
undersampling
Random __: randomly delete data points from the majority class.
SMOTE
__ operates in the feature space and creates new data points from the minority class.
The generator
__ tries to mimic examples from a training dataset by transforming a random source of noise received as input into a synthetic sample.
cost matrix
In cost-sensitive classification, the probability threshold can be changed using a __.
fit()
In sklearn models, the __ function has a class_weight parameter.
Precision
__ = TP / (TP + FP)
Recall
__ = TP / (TP + FN)
Specificity
__ = TN / (TN + FP)
prejudice, favoritism
Fairness is the absence of any or towards an individual or a group based on their intrinsic or acquired traits in the decision-making context.
Demographic
__ groups are determined based on sensitive attributes.
positive
The instances in both unprivileged and privileged groups should have equal probability to receive __ outcomes.
True Positive, False
EO states that instances from privileged and unprivileged groups should have equal Rate (TPR) and Positive Rate (FPR).
Pre-Processing, In-Processing, Post-Processing
Three types of Mitigation Algorithms are: __, __, and __.
Pre-Processing
__ techniques pre-process the dataset only by using strategies such as fairness through "unawareness", preferential sampling, massaging, and reweighing.
In-Processing
__ techniques adjust/tune the classification algorithm during the model training, and are dependent on the implemented classifier.
Post-Processing
__ techniques eliminate the discrimination from the final predictions and change the predicted outcomes of classifiers based on certain rules or constraints.