Learning Fair Naive Bayes Classifiers by Discovering and Eliminating Discrimination Patterns

Authors: YooJung Choi, Golnoosh Farnadi, Behrouz Babaki, Guy Van den Broeck10077-10084

AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental An empirical evaluation on three real-world datasets demonstrates that we can remove exponentially many discrimination patterns by only adding a small fraction of them as constraints.
Researcher Affiliation Academia 1University of California, Los Angeles, 2Mila, 3Universit e de Montr eal, 4Polytechnique Montr eal {yjchoi, guyvdb}@cs.ucla.edu, farnadig@mila.quebec, behrouz.babaki@polymtl.ca
Pseudocode Yes Algorithm 1 DISC-PATTERNS(x, y, E)
Open Source Code Yes The processed data, code, and Appendix are available at https: //github.com/UCLA-Star AI/Learn Fair NB.
Open Datasets Yes We use three datasets: The Adult dataset and German dataset are used for predicting income level and credit risk, respectively, and are obtained from the UCI machine learning repository5; the COMPAS dataset is used for predicting recidivism.
Dataset Splits Yes Table 4 reports the 10-fold CV accuracy of our method (δ-fair) compared to a max-likelihood naive Bayes model (unconstrained) and two other methods of learning fair classifiers
Hardware Specification Yes All experiments were run on an AMD Opteron 275 processor (2.2GHz) and 4GB of RAM running Linux Centos 7.
Software Dependencies No To solve the signomial programs, we use GPkit, which finds local solutions to these problems using a convex optimization solver as its backend.7 Throughout our experiments, Laplace smoothing was used to avoid learning zero probabilities. 7We use Mosek (www.mosek.com) as backend.
Experiment Setup Yes Throughout our experiments, Laplace smoothing was used to avoid learning zero probabilities.