d6539d3b57159babf6a72e106beb45bd-AuthorFeedback.pdf

Hyper-Parameter Selection: We fixed the number of iterations of algorithm 1 to be10 for EO and5 for DP.22 By theorem 1, increasing this parameter will only increase the accuracy of the final randomized classifier. All the23 other hyperparameters were chosen by performing grid search using the validation set. In terms of attribute measurement errors, it can handle the situation when errors across all the attributes are30 similar e.g. Limited Loss Function: Fairness definitions in binary classification setting are usually considered with 0-147 predictions.