FairGBM: Gradient Boosting with Fairness Constraints
Authors: André Cruz, Catarina G Belém, João Bravo, Pedro Saleiro, Pedro Bizarro
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate our method on five large-scale public benchmark datasets, popularly known as folktables datasets, as well as on a real-world financial services case-study. We compare Fair GBM with a set of constrained optimization baselines from the Fair ML literature. |
| Researcher Affiliation | Collaboration | 1Feedzai 2MPI for Intelligent Systems, Tübingen 3UC Irvine |
| Pseudocode | Yes | Algorithm 1 Fair GBM training pseudocode |
| Open Source Code | Yes | Our implementation1 shows an order of magnitude speedup in training time relative to related work, a pivotal aspect to foster the widespread adoption of Fair GBM by real-world practitioners. (footnote 1: https://github.com/feedzai/fairgbm) |
| Open Datasets | Yes | We validate our method on five large-scale public benchmark datasets, popularly known as folktables datasets, as well as on a real-world financial services case-study. The folktables datasets were put forth by Ding et al. (2021) and are derived from the American Community Survey (ACS) public use microdata sample from 2018. |
| Dataset Splits | Yes | Each task is randomly split in training (60%), validation (20%), and test (20%) data. |
| Hardware Specification | Yes | ACSIncome and AOF experiments: Intel i7-8650U CPU, 32GB RAM. ACSEmployment, ACSMobility, ACSTravel Time, ACSPublic Coverage experiments: each model trained in parallel on a cluster. Resources per training job: 1 v CPU core (Intel Xeon E5-2695), 8GB RAM3. |
| Software Dependencies | No | The paper mentions "Light GBM implementation" and its language "C++" and "Python interface" but does not provide specific version numbers for Light GBM, Python, or any other critical libraries/dependencies. The reproducibility checklist points to supplementary materials but does not explicitly list software versions in the text. |
| Experiment Setup | Yes | To control for the variability of results when selecting different hyperparameters, we randomly sample 100 hyperparameter configurations of each algorithm. In the case of EG and GS, both algorithms already fit n base estimators as part of a single training procedure. Hence, we run 10 trials of EG and GS, each with a budget of n = 10 iterations, for a total budget of 100 models trained (leading to an equal budget for all algorithms). |