EAGER: Constraint Aware Generative Adversarial Networks

生成性对抗性网络(GAN)是通过对抗性游戏估计真实数据分布的网络，在图像生成和文本生成方面取得了巨大的成功。然而，现实世界应用程序中的许多决策模型都是在包含数字和分类属性的关系数据上进行训练的。此外，现实世界中的应用程序通常对生成的数据具有法律或伦理要求，例如，无歧视或隐私保护，或者期望生成与其现有数据互补的样本，而现有数据可能只包含正常样本。该研究显著提高了从图像/文本数据到具有应用需求的关系数据的生成性对抗网络的适用性，并有助于解决使用生成性对抗网络生成约束感知关系数据的知识基础有限的问题。调查结果、工具、软件代码和课程材料文档将分发给研究社区、IT行业和用户，以帮助领域用户生成可识别约束的数据以满足其业务需求。这项急切的研究开发了新的技术，使当前的生成性对抗网络能够生成具有约束的现实关系数据。该框架在生成器中增加一个解码器来生成数值型和类别型数据，并在生成器和鉴别器的目标函数中加入约束项或引入多个鉴别器来执行需求约束。该研究采用f-发散度分析了在引入复杂约束时约束感知GaN框架的收敛问题。然后，研究重点是在统一的框架下开发两个具体的模型，公平的GAN用于生成无歧视数据，以及互补的GAN用于在训练数据中只有正样本时生成负样本。本研究从精确度和收敛两个方面对该框架和两个具体模型进行了实证评估，并将所开发的算法实现并集成到开源深度学习软件系统TensorFlow中。所开发的框架有望促进对生成性对抗网络的理论理解，两个具体的GAN模型有望改善当前公平意识学习和欺诈检测的研究。特别是，公平GAN引入了基于GAN的公平数据生成的新方法，因为目前的公平感知学习研究主要采用数据修改。免费赠送的GAN通过生成补充样本并使训练有素的鉴别器能够准确区分异常样本和正常样本，在欺诈检测方面优于现有的一级分类模型。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Generative adversarial networks (GAN), which estimate the real data distribution through an adversarial game, have achieved great success in image generation and text generation. However, many decision models from real world applications are trained on relational data that contains both numerical and categorical attributes. Furthermore, real world applications often have legal or ethical requirements on the generated data, e.g., discrimination free or privacy preservation, or expect to generate complementary samples to their existing data that may only contain normal samples. The research significantly improves the applicability of generative adversarial networks from image/text data to relational data with application requirements and contributes to the limited base of knowledge in the area of using generative adversarial networks for constraint aware relational data generation. The findings, tools, software code, and curricular materials documents are disseminated to the research community, IT industry, and users, which expects to help domain users generate constraint aware data to meet their business needs. This EAGER research develops novel techniques that enable the current generative adversarial networks to generate realistic relational data with constraints. The developed framework adds a decoder to the generator to generate both numerical and categorical data and incorporates constraint terms into the objective functions of generator and discriminator or introduces multiple discriminators to enforce requirement constraints. The research adopts f-divergences to analyze the convergence of the constraint aware GAN framework when complex constraints are introduced. The research then focuses on developing under the unifying framework two specific models, fair GAN for generating discrimination-free data, and complementary GAN for generating negative samples when only positive samples are available in the training data. The research conducts empirical evaluations of the framework and two specific models in terms of accuracy and convergence, implements and integrates the developed algorithms into TensorFlow, an open source deep learning software system. The developed framework expects to advance theoretical understanding of generative adversarial networks and the two specific GAN models expect to improve the current research on fairness aware learning and fraud detection. In particular, the fair GAN introduces the new approach of fair data generation based on GAN as current fairness aware learning research mainly adopts data modification. The complimentary GAN outperforms existing one-class classification models for fraud detection by generating complementary samples and enabling the trained discriminator to accurately separate abnormal samples from normal ones.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.