On Human-Aligned Risk Minimization

Authors: Liu Leqi, Adarsh Prasad, Pradeep K. Ravikumar

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically study these risk measures, and demonstrate their improved performance on desiderata such as fairness, in contrast to the traditional workhorse of expected loss minimization.
Researcher Affiliation Academia Liu Leqi Carnegie Mellon University Pittsburgh, PA 15213 leqil@cs.cmu.edu Adarsh Prasad Carnegie Mellon University Pittsburgh, PA 15213 adarshp@cs.cmu.edu Pradeep Ravikumar Carnegie Mellon Universit Pittsburgh, PA 15213 pradeepr@cs.cmu.edu
Pseudocode No The paper describes an 'iterative update rule' for optimization using mathematical equations but does not present it as formally labeled pseudocode or an algorithm block.
Open Source Code No The paper does not provide any explicit statements about releasing source code or links to a code repository.
Open Datasets Yes Using the fair ML toolkit version of the COMPAS recidivism dataset [1]... we use the UTKFace dataset [31] to train a neural network
Dataset Splits No The paper mentions '2000 training and 20000 testing data points' and 'With a 90% and 10% train-test split', but does not explicitly specify a separate validation split or its proportions.
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments (e.g., GPU models, CPU types, or memory specifications).
Software Dependencies No The paper mentions software like 'logistic regression model', 'neural network', 'mini-batch stochastic gradient descent', and 'AI Fairness 360 toolkit [2]', but does not provide specific version numbers for any of these or other software dependencies.
Experiment Setup Yes We fix the squared error (θ; (x, y)) = 1/2(y - x^Tθ)^2 as our loss function. ... logistic regression model with L2 regularization. ... For EHRM, we have chosen b to be .3... To minimize empirical human risk, we have used a variant of mini-batch stochastic gradient descent. At each step t, θt+1 = θt - ηt/B Σi=1 to B wt i θ (θ; Zi), where wt i = w POLY(Fn( (θt; Zi))), Fn( ) is the empirical CDF of the mini-batch losses, B is the mini-batch size and ηt is the learning rate.