Code for "Classification Under Misspecification: Halfspaces, Generalized Linear Models, and Connections to Evolvability"

These are scripts for reproducing our paper "Classification Under Misspecification: Halfspaces, Generalized Linear Models, and Connections to Evolvability" (preprint).

Contents

• adult.csv contains a copy of the UCI Adult dataset.
• gauss/ contains the results of our synthetic experiment
• real/ contains the results of our UCI Adult experiment
• noncensor/ contains the experiment where demographic fields were not hidden
• censor/ contains the experiment where demographic fields were hidden
• To reproduce the plots in this work, run python plots.py
• experiment.py is the wrapper for running both our real and synthetic experiments. Setting the flag EXPERIMENT_TYPE therein to one of ['real', 'synthetic-rcn', 'synthetic-massart'] and running python experiment.py will populate gauss/ and real/ with the relevant experimental data (these folders have been pre-populated for the user's convenience).