Robust Fairness Under Covariate Shift
Authors: Ashkan Rezaei, Anqi Liu, Omid Memarrast, Brian D. Ziebart9419-9427
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the benefits of our approach on benchmark prediction tasks. We compare our proposed method with baselines that only account for covariate shift or fairness or both. We demonstrate that our method outperforms the baselines on both predictive performance and the fairness constraints satisfaction under covariate shift settings. Experiments We demonstrate the effectiveness of our method on biased samplings from four benchmark datasets: The COMPAS criminal recidivism risk assessment dataset... UCI German dataset... UCI Drug dataset... UCI Arrhythmia dataset... Results Figure 2 shows our experimental results on three samplings from close to IID (left) to mild (middle) and strong covariate shift (right). |
| Researcher Affiliation | Academia | Ashkan Rezaei1, Anqi Liu2, Omid Memarrast1, Brian D. Ziebart1 1 University of Illinois at Chicago 2 California Institute of Technology |
| Pseudocode | Yes | Algorithm 1: Batch Fair Robust Log-Loss learning under Covariate Shift |
| Open Source Code | No | The paper states 'We use the implementation from https://fairlearn.github.io.' for a baseline method (HARDT), but does not provide a link or statement about open-sourcing the code for their proposed method. |
| Open Datasets | Yes | The COMPAS criminal recidivism risk assessment dataset (Larson et al. 2016). UCI German dataset (Dheeru and Karra Taniskidou 2017). UCI Drug dataset (Fehrman et al. 2017). UCI Arrhythmia dataset (Dheeru and Karra Taniskidou 2017). UCI Machine Learning Repository. URL http://archive.ics.uci.edu/ml. Accessed on June 2020. |
| Dataset Splits | No | The paper describes splitting data into source and target, but does not provide explicit training, validation, and test splits. It states: 'Unfortunately since the target distribution is assumed to be unavailable for this problem, properly obtaining optimal regularization via cross validation is not possible.' |
| Hardware Specification | No | The paper does not provide any specific hardware specifications (e.g., GPU/CPU models, memory details) used for running the experiments. |
| Software Dependencies | No | The paper mentions 'We use https://pypi.org/project/KDEpy package' but does not specify a version number for it or for any other core software libraries like Python, PyTorch, TensorFlow, etc. |
| Experiment Setup | Yes | We select the L2 regularization parameter by choosing the best C from {10-5, 10-4, 10-3, 10-2, 10-1, 1, 10} under the IID setting. We use first-order features for our implementation, i.e., φ(x, y) = [x1y, x2y, . . . xmy] , where m is the size of features. we find the exact zero point of the approximated violation efficiently by binary search for µ [-1.5, 1.5] in our experiments. We employ a batch gradient-only optimization to learn our model parameters. We perform a joint gradient optimization that updates the gradient of θ (which requires true label) from the source data and λ (which does not require true label) from the target batch at each step. decaying learning rate ηt |