FAIRER: Fairness as Decision Rationale Alignment
Authors: Tianlin Li, Qing Guo, Aishan Liu, Mengnan Du, Zhiming Li, Yang Liu
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on a variety of datasets show that our method can significantly enhance fairness while sustaining a high level of accuracy and outperforming other approaches by a wide margin. Extensive experiments on three public datasets, i.e., Adult, Celeb A, and Credit, demonstrate that our method can enhance the fairness of DNNs effectively and outperform others largely. |
| Researcher Affiliation | Collaboration | 1Nanyang Technological University, Singapore 2Institute of High Performance Computing (IHPC), A*STAR, Singapore 3Centre for Frontier AI Research (CFAR), A*STAR, Singapore 4Beihang University, China 5New Jersey Institute of Technology, USA 6Zhejiang Sci-Tech University, China. Corresponding author: Qing Guo <tsingqguo@ieee.org>. |
| Pseudocode | Yes | Algorithm 1 Gradient-guided Parity Alignment |
| Open Source Code | No | The paper does not provide a direct link or explicit statement for open-sourcing the code. |
| Open Datasets | Yes | In our experiments, we use two tabular benchmarks (Adult and Credit) and one image dataset (Celeb A) that are all for binary classification tasks: ❶Adult (Dua & Graff, 2017a). ❷Celeb A (Liu et al., 2015). ❸Credit (Dua & Graff, 2017b). |
| Dataset Splits | Yes | We employ average precision for the accuracy score and DP for the fairness score since a smaller DP means better fairness. For each method, we can draw a plot w.r.t. different λ. Besides, we also train a network without the fairness regularization term and denote it as w.o.Fair Reg. Based on w.o.Fair Reg, we can conduct oversampling on the training samples to balance the samples across different subgroups (Wang et al., 2020) and denote it as w.o.Fair Reg-Oversample. The model is chosen according to the performance on the validation dataset. |
| Hardware Specification | No | The paper does not specify the exact hardware (e.g., GPU/CPU models) used for running the experiments. |
| Software Dependencies | No | The paper mentions using Adam as the learning optimizer but does not specify version numbers for any software dependencies like Python, PyTorch, TensorFlow, or specific libraries. |
| Experiment Setup | Yes | The parameter setting of the Adult Dataset is shown in Table 2. We follow the settings in Chuang & Mroueh (2021) for data preprocessing. The hidden size of MLP is 200. We use Adam as the learning optimizer and the batch size is set as 1000 for the DP metric and 2000 for the EO metric following the setting in Chuang & Mroueh (2021). The parameter setting of the Celeb A Dataset is shown in Table 3. The parameter setting of the Credit Dataset is shown in Table 4. |