When do Minimax-fair Learning and Empirical Risk Minimization Coincide?
Authors: Harvineet Singh, Matthäus Kleindessner, Volkan Cevher, Rumi Chunara, Chris Russell
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5. Empirical Study We perform a large-scale study to test how well do our theoretical results generalize to practical scenarios where the assumptions may not hold. |
| Researcher Affiliation | Collaboration | 1New York University, New York, USA 2Amazon Web Services, T ubingen, Germany 3 Ecole Polytechnique F ed erale de Lausanne, Lausanne, Switzerland. |
| Pseudocode | No | The paper does not contain any explicit pseudocode or algorithm blocks. Methods are described via mathematical formulations and textual explanations. |
| Open Source Code | Yes | Code to reproduce the experiments is available at https://github.com/amazon-science/rethinking-minimax-fairness |
| Open Datasets | Yes | Table 1: Datasets. ACS datasets are curated for 4 US states (NY, CA, TX, IN). Datasets for different domains combined with different group types result in 36 datasets in total. All labels are binary. Dataset sources are given in Section E.2 in the appendix. ... Section E.2: 1. ACS Income, ACS Employment, ACS Health Insurance (Ding et al., 2021). Accessed using folktables package https://github.com/zykls/folktables from https://www.census.gov/programs-surveys/acs. |
| Dataset Splits | No | Datasets are divided randomly into 70-30 train-test split. |
| Hardware Specification | Yes | Experiments were run on a compute cluster using 36 nodes with an Intel Xeon 2.9 GHz processor, 1 NVIDIA RTX8000 GPU and 24 GB system memory for each node. |
| Software Dependencies | No | We use the default ERM solvers available in the scikit-learn Python package (Pedregosa et al., 2011). ... MLP is implemented using code from rtdl package4. ... Py Torch (Paszke et al., 2019). |
| Experiment Setup | Yes | Table 5: Hyperparameters used for the model types. Unless specified we use the default settings in scikit-learn. ... MLP, MLP in Gorishniy et al. (2021) hidden layer sizes=[1024,1024], dropout=0.1, lr=0.001, batch size=2048, Adam W optimizer in Py Torch (Paszke et al., 2019), ERM iterations=2000 ... For the optimization procedure of the minimax method, we set the convergence threshold as 10 12 and run at most 10000 iterations, except for MLP we use 200 iterations to reduce compute time. |