Fair Generalized Linear Models with a Convex Penalty
Authors: Hyungrok Do, Preston Putzel, Axel S Martin, Padhraic Smyth, Judy Zhong
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To empirically demonstrate the efficacy of the proposed fair GLM, we compare it with other wellknown fair prediction methods on an extensive set of benchmark datasets for binary classification and regression. In addition, we demonstrate that the fair GLM can generate fair predictions for a range of response variables, other than binary and continuous outcomes. ... 6. Experiments and Results We performed experiments for a comprehensive list of benchmark datasets to evaluate the proposed F-GLM, comparing it with the naive GLM and with multiple in-process linear model-based fairness-aware methods. |
| Researcher Affiliation | Academia | 1Department of Population Health, NYU Grossman School of Medicine, New York, NY, USA 2Department of Computer Science, University of California, Irvine, CA, USA. |
| Pseudocode | Yes | Algorithm 1 describes this discretization procedure for continuous outcomes. |
| Open Source Code | Yes | Full code and datasets for our experiments are available on https://github.com/hyungrok-do/fair-glm-cvx. |
| Open Datasets | Yes | We consider four different tasks/outcome types: binary classification (5 datasets), multiclass classification (2 datasets), continuous outcomes (4 datasets), and count outcomes (1 dataset). General characteristics of the datasets are summarized in Table 3. |
| Dataset Splits | No | We randomly divided each dataset into training (70%) and testing (30%) sets, except for the Adult dataset which has predefined train/test splits. ... For each value of the hyperparameter, the performance and disparity measures were then estimated by averaging over 20 replicates of random splits of the training and testing sets (except for the Adult dataset). |
| Hardware Specification | No | The paper mentions computation times ('fitting an F-GLM took 7 seconds to compute D and an additional 9 seconds until convergence') but does not provide specific hardware details such as CPU/GPU models or memory specifications used for running experiments. |
| Software Dependencies | No | The paper mentions software like 'Python code', 'fairlearn Python package', 'CVXPY', and 'CPLEX', but it does not specify any version numbers for these software dependencies, which are required for reproducible description. |
| Experiment Setup | Yes | Each model was trained on the training set by varying its fairness-related hyperparameter (if it exists) over a suitable range. ... We presented the range of the hyperparameters used for our experiments in Table 4. For some datasets, we used slightly different range of hyperparameters for some methods. Details can be found in our code. |