Fairness Through Matching
Authors: Kunwoong Kim, Insung Kong, Jongjin Lee, Minwoo Chae, Sangchul Park, Yongdai Kim
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5 Experiments This section presents our experimental results, showing that FTM with the proposed transport maps in Section 4 empirically works well to learn group-fair models. The key findings throughout this section are summarized as follows. FTM with the marginal OT map successfully learns group-fair models that exhibit (i) competitive prediction performance (Section 5.2.1) and (ii) higher levels of subset fairness (Section 5.2.2), when compared to other group-fair models learned by existing baseline algorithms. Beyond subset fairness, we further evaluate the self-fulfilling prophecy (Dwork et al., 2012) as an additional benefit of low transport cost (see Table 10 and 11 in Section E of Appendix). FTM with the joint OT map has the ability to learn group-fair models with improved prediction performance as well as improved levels of equalized odds, when compared to FTM with the marginal OT map (Section 5.3). |
| Researcher Affiliation | Collaboration | Kunwoong Kim EMAIL Department of Statistics Seoul National University Insung Kong EMAIL Department of Applied Mathematics University of Twente Jongjin Lee EMAIL Samsung Research Minwoo Chae EMAIL Department of Industrial and Management Engineering Pohang University of Science and Technology (POSTECH) Sangchul Park EMAIL School of Law Seoul National University Yongdai Kim EMAIL Department of Statistics Interdisciplinary Program in Artificial Intelligence Seoul National University |
| Pseudocode | Yes | D.3 Pseudo-code Here, we provide a Pytorch-style psuedo code of calculating the MDP constraint in FTM. Algorithm 1: Py Torch-style pseudo-code of calculating the MDP constraint in FTM. |
| Open Source Code | Yes | 1The source code of FTM is publicly available at https: // github. com/ kwkimonline/ FTM . |
| Open Datasets | Yes | Datasets We use four real-world benchmark tabular datasets in our experiments: Adult (Dua & Graff, 2017), German (Dua & Graff, 2017), Dutch (Van der Laan, 2001), and Bank (Moro et al., 2014). The basic information about these datasets is provided in Table 6 in Section D.1 of Appendix. ... Adult: the Adult income dataset (Dua & Graff, 2017) can be downloaded from the UCI repository3. German: the German credit dataset (Dua & Graff, 2017) can be downloaded from the UCI repository4. Dutch: the Dutch census dataset can be downloaded from the public Github of Quy et al. (2022) 5. Bank: the Bank marketing dataset can be downloaded from the UCI repository6. |
| Dataset Splits | Yes | We randomly partition each dataset into training and test datasets with the 8:2 ratio. For each split, we learn models using the training dataset and evaluate the models on the test dataset. This process is repeated 5 times, and the average performance on the test datasets is reported. |
| Hardware Specification | Yes | We utilize several Intel Xeon Silver 4410Y CPU cores and RTX 3090 GPU processors. |
| Software Dependencies | No | The Adam optimizer (Kingma & Ba, 2014) with the initial learning rate of 0.001 is used. To obtain the OT map for each mini-batch, we solve the linear program by using the POT library (Flamary et al., 2021). We utilize several Intel Xeon Silver 4410Y CPU cores and RTX 3090 GPU processors. More implementation details with Pytorch-style psuedo-code are provided in Section D.2 and D.3 of Appendix. |
| Experiment Setup | Yes | For all algorithms, we employ MLP networks with Re LU activation and two hidden layers, where the hidden size is equal to the input dimension. We run all algorithms for 200 epochs and report their final performances on the test dataset. The Adam optimizer (Kingma & Ba, 2014) with the initial learning rate of 0.001 is used, and the learning rate is scheduled by multiplying 0.95 at each epoch. To obtain the OT map for each mini-batch, we solve the linear program by using the POT library (Flamary et al., 2021). We utilize several Intel Xeon Silver 4410Y CPU cores and RTX 3090 GPU processors. More implementation details with Pytorch-style psuedo-code are provided in Section D.2 and D.3 of Appendix. ... Table 7: Hyper-parameters used for controlling fairness levels for each algorithm. |