Fair Bayes-Optimal Classifiers Under Predictive Parity
Authors: Xianli Zeng, Edgar Dobriban, Guang Cheng
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We provide supporting experiments conducted on synthetic and empirical data. |
| Researcher Affiliation | Academia | Xianli Zeng NUS (Chongqing) Research Institute Chongqing, China zengxl19911214@gmail.com Edgar Dobriban University of Pennsylvania Philadelphia, PA 19104 dobriban@wharton.upenn.edu Guang Cheng University of California, Los Angeles Los Angeles, CA 90095 guangcheng@ucla.edu |
| Pseudocode | Yes | Algorithm 1 Fair Bayes-DPP |
| Open Source Code | No | The paper does not provide a specific link or explicit statement about the availability of its source code. |
| Open Datasets | Yes | We test Fair Bayes-DPP on two benchmark datasets for fair classification: Adult [14] and COMPAS [27]. To further test the performance of our algorithm on a large-scale dataset, we conduct experiments on the Celeb Faces Attributes (Celeb A) Dataset [32]. |
| Dataset Splits | Yes | For each dataset, we randomly sample (with replacement) 70%, 50% and 30% as the training, validation and test set, respectively. |
| Hardware Specification | No | The paper mentions training models and using PyTorch, and a ResNet50 model, but does not specify any particular GPU or CPU models, memory sizes, or detailed hardware specifications used for experiments. |
| Software Dependencies | No | The paper mentions "Py Torch" as a software dependency but does not specify its version number or any other software dependencies with their versions. |
| Experiment Setup | Yes | We set the cost parameter c = 0.5, while P(A = 1) = 0.3 and P(Y = 1|A = 0) = 0.2. In the Gaussian case, the Bayes-optimal classifier is linear in x and thus we employ logistic regression to learn η1( ) and η0( ). We then search over a grid with spacings equal to 0.001 over the range we identified in Section 5 for the empirically optimal thresholds under fairness. The conditional probabilities are learned via a three-layer fully connected neural network architecture with 32 hidden neurons per layer. We refer readers to the Appendix for more training details, including optimizer, learning rates, batch sizes and training epochs. |