Fairness via Representation Neutralization
Authors: Mengnan Du, Subhabrata Mukherjee, Guanchu Wang, Ruixiang Tang, Ahmed Awadallah, Xia Hu
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results over several benchmark datasets demonstrate our RNF framework to effectively reduce discrimination of DNN models with minimal degradation in task-specific performance. |
| Researcher Affiliation | Collaboration | Mengnan Du1 , Subhabrata Mukherjee2, Guanchu Wang3, Ruixiang Tang3, Ahmed Hassan Awadallah2, Xia Hu3 1Texas A&M University 2Microsoft Research 3Rice University |
| Pseudocode | Yes | Algorithm 1: RNF mitigation framework. |
| Open Source Code | No | The paper does not provide any explicit statement or link for open-sourcing the code for the described methodology. |
| Open Datasets | Yes | We use two benchmark tabular datasets and one image dataset to evaluate the effectiveness of RNF. For the Adult income dataset (Adult)... For the Medical Expenditure dataset (MEPS)... The Celeb Faces Attributes (Celeb A) dataset is used... |
| Dataset Splits | Yes | We split all datasets into three subsets with statistics reported in Table 1. Table 1: Dataset statistics. # Training 33120 # Validation 3000 # Test 9102 (for Adult dataset, similar for others). |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU models, CPU types, or memory amounts used for the experiments. |
| Software Dependencies | No | The paper does not specify the versions of software dependencies (e.g., programming languages, libraries, or frameworks) used to run the experiments. |
| Experiment Setup | Yes | For the image classification task, we use Res Net-18... For tabular datasets, we use a three-layer MLP... Dropout is used for the first two layers with the dropout probability fixed at 0.2. We use the same batch size of 64 for tabular datasets and 390 for the image dataset... The optimal temperature T used to calculate the probability is set as 2.0, 5.0, 2.0 for Adult, MEPS and Celeb A datasets respectively. For Eq. (3), we sample λ from the list [0.6, 0.7, 0.8, 0.9]. The hyper-parameter q in Eq. (5) is set as 0.2, 0.6, 0.3 for Adult, MEPS and Celeb A datasets respectively. |