Fair Classifiers that Abstain without Harm

Authors: Tongxin Yin, Jean-Francois Ton, Ruocheng Guo, Yuanshun Yao, Mingyan Liu, Yang Liu

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We have carried out extensive experiments to demonstrate the benefits of our solution compared to strong existing baselines.
Researcher Affiliation Collaboration 1 University of Michigan 2 Byte Dance Research tyin@umich.edu, jeanfrancois@bytedance.com, rguo.asu@gmail.com, kevin.yao@bytedance.com, mingyan@umich.edu, yang.liu01@bytedance.com
Pseudocode Yes Algorithm 1 Prediction Adjustment
Open Source Code Yes Code: https://github.com/tsy19/FAN
Open Datasets Yes We adopt three real-world datasets: Adult (Dua & Graff, 2017), Compas (Bellamy et al., 2018), and Law (Bellamy et al., 2018).
Dataset Splits Yes Table 5: Size of train, val, test data of each dataset.
Hardware Specification Yes We run the experiments on a single T100 GPU.
Software Dependencies No The paper mentions 'Py Torch' but does not specify its version or any other software dependencies with version numbers.
Experiment Setup Yes The architecture configuration for the Adult dataset consists of two layers, each with a dimension of 300. For the Compas and Law datasets, we employed two layers, each with a dimension of 100. A dropout layer with a dropout probability of 0.5 was applied between the two hidden layers. The Rectified Linear Unit (Re LU) function was used as the activation function.