Fairness without Harm: An Influence-Guided Active Sampling Approach

Authors: Jinlong Pang, Jialu Wang, Zhaowei Zhu, Yuanshun Yao, Chen Qian, Yang Liu

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on real-world data demonstrate the effectiveness of our proposed algorithm.
Researcher Affiliation Collaboration Jinlong Pang UC Santa Cruz jpang14@ucsc.edu Jialu Wang UC Santa Cruz faldict@ucsc.edu Zhaowei Zhu Docta.ai zzw@docta.ai Yuanshun Yao Meta Gen AI kevinyao@meta.com Chen Qian UC Santa Cruz cqian12@ucsc.edu Yang Liu UC Santa Cruz yangliu@ucsc.edu
Pseudocode Yes Algorithm 1 Fair influential sampling (FIS)
Open Source Code Yes Our code is available at github.com/UCSC-REAL/Fairness Without Harm.
Open Datasets Yes We evaluate the performance of our algorithm on three real-world datasets across three different modalities: Celeb A [49], UCI Adult [8] and Compas [6].
Dataset Splits Yes Then, the test dataset is split into two independent portions: a new test set and a validation set, with 10% of the test data randomly designated as the hold-out validation set.
Hardware Specification No The paper mentions 'running one experiment for the Celeb A dataset on a single GPU roughly requires about 4 hours' but does not specify the model or type of GPU.
Software Dependencies No The paper mentions using 'SGD optimizer' and 'Re LU network' but does not list specific software libraries or their version numbers (e.g., PyTorch, TensorFlow, scikit-learn versions).
Experiment Setup Yes The epochs are split into two phases: warm-up epochs (5 epochs) and training epochs (10 epochs). The default label budget per round, which represents the number of solicited data samples, is set to 256. Additionally, the default values for learning rate, momentum, and weight decay are 0.01, 0.9, and 0.0005, respectively.