Efficient Discrete Multi Marginal Optimal Transport Regularization
Efficient Discrete Multi Marginal Optimal Transport Regularization
复制标题
DOI：
--
复制
发表时间：
2023
期刊：
影响因子：
--
通讯作者：
Ronak R. Mehta;Vishnu Suresh Lokhande
中科院分区：
文献类型：
--
作者：
Ronak R. Mehta;Vishnu Suresh Lokhande
我们建立了一个简单的三层神经网络用于分类任务，并增加了一个公平型正则化函数。我们将我们的构造与4个现成的插件正则化程序进行了比较：(1)无正则化，(2)人口均等(DP)，(3)均衡赔率(EO)，和(4)直方图重心构造。DP和EO正则化是使用FairLain的PyTorch版本(Bird等人，2020年)计算的，而重心版本是使用POT库(具有GPU后端)实现的。因为正则化项的尺度不是直接可比较的，所以我们扫描正则化权重，并为每个数据集/正则化对的所有度量选择最好的。我们使用10个箱子，并在三个随机种子上重复所有实验。模特们。我们用一个标准的Logistic回归模型和一个两层神经网络建立了两个模型。我们比较了三种类型的插件正则化：(1)人口统计平等(DP)，(2)均衡赔率(EO)，和(3)广义EMD。DP和EO正则化因子是使用FairLain的PyTorch实现来计算的(Bird等人，2020年)。
We set up a simple three-layer neural network for classification tasks with the addition of a fairness-type regularizer. We compare our construction with 4 off-the-shelf plug-in regularizers: (1) No regularization, (2) Demographic Parity (DP), (3) Equalized Odds (EO), and (4) a histogrammed barycenter construction. DP and EO regularizers were computed using a PyTorch version of FairLearn (Bird et al., 2020), and the barycenter version was implemented using POT library (with GPU backend). Because the scale of the regularization term is not directly comparable, we sweep regularization weights and select the best over all measures for a each dataset/regularizer pair. We use 10 bins and replicate all experiments over three random seeds. Models. We set up two model settings with a standard logistic regressor and a 2-layer neural network. We compare three types of plug-in regularizers: (1) Demographic Parity (DP), (2) Equalized Odds (EO), and (3) the Generalized EMD. DP and EO regularizers were computed using a PyTorch implementation of FairLearn (Bird et al., 2020).