32e54441e6382a7fbacbbbaf3c450059-Supplemental.pdf

We only included a candidate variable if the nearest neighbor match was exact, i.e., we could find We compared the "fnlwgt" data to all weight variables "UH_WGTS_A1", which has a similar distribution. Since we did not identify an exact match for "fnlwgt" and the variable is not a property of an individual, we do not utilize it further in We vary the threshold from 6,000 to 72,000. Concretely, for a given threshold, e.g. In our experiments, as the "unconstrained" base classifier, we use the gradient boosted decision tree B.1 ACSIncome Predict whether US working adults' yearly income is above $50,000. T arget: PINCP (Total person's income): an individual's label is 1 if PINCP > 50000, otherwise 0. ACS PUMS data differently, and construct a new prediction task. Features: AGEP (Age): Range of values: - 0 - 99 (integers) - 0 indicates less than 1 year old.