Blind Pareto Fairness and Subgroup Robustness
Authors: Natalia L Martinez, Martin A Bertran, Afroditi Papadaki, Miguel Rodrigues, Guillermo Sapiro
ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experimental results show that the proposed framework improves worst-case risk in multiple standard datasets, while simultaneously providing better levels of service for the remaining population. |
| Researcher Affiliation | Academia | 1Duke University, Durham, NC, USA 2University College London, London, UK. |
| Pseudocode | Yes | Algorithm 1 Blind Pareto Fairness Require: Inputs: Dataset {(xi, yi)}n i=1, partition size ρ Require: Hyper-parameters: rounds T, parameter η, adversary boundary coefficient ϵ > 0, γ-approximate Bayesian solver M( ) arg minh H L(h, ) Initialize α0 = ˆα = {ρ}n i=1 Initialize classifier and loss h0 = M( ˆα), L0 = L(h0, ˆα) for round t = 1, . . . , T do Adversary updates partition function by projected gradient ascent: αt αt 1 + η αL(ht, ˆα) = αt 1 + η ℓ(ht,y) nρ ˆα Q(αt) Uϵ,ρ , Uϵ,ρ = {α : αi [ϵ, 1], P Solver approximately solves for the current partition: ht M( ˆα) end for Return: Classifier h T |
| Open Source Code | Yes | The code is available at github.com/natalialmg/Blind Pareto Fairness. |
| Open Datasets | Yes | Datasets. We used four standard fairness datasets for comparison. The UCI Adult dataset (Dua & Graff, 2017) which contains... The Law School dataset (Wightman, 1998) contains... The COMPAS dataset (Barenstein, 2019) which contains... Lastly we used the MIMIC-III dataset, which consists of clinical records... (Johnson et al., 2016). |
| Dataset Splits | Yes | Results correspond to the best hyper-parameter for each group size; mean and standard deviations are computed using 5-fold cross-validation, all figures are reported on held out (test) data. |
| Hardware Specification | No | The paper does not provide any specific hardware details such as GPU/CPU models, memory, or cloud instances used for running the experiments. |
| Software Dependencies | No | The paper does not provide specific software dependency details, such as programming language versions or library versions (e.g., Python, PyTorch, TensorFlow versions), that would be necessary to precisely replicate the experiments. |
| Experiment Setup | Yes | We train BPF for 18 minimum group sizes ρ = {0.05, . . . , 0.9} and ϵ = 0.01... DRO models were trained on 18 equispaced values of their threshold parameter (η ∈ [0, 1]). For ARL, we tried four configurations for their adversarial network (adversary with 1 or 2 hidden layers with 256 or 512 units each)... The classifier architecture for BPF, ARL, and DRO was standardized to a single-layer MLP with 512 hidden units. In all cases we use cross-entropy loss and same input data. Results correspond to the best hyper-parameter for each group size; mean and standard deviations are computed using 5-fold cross-validation... |