FAI: A novel paradigm for fairness-aware deep learning models on data streams

大量的信息以数据流的形式在不同的领域之间不断传递。社交网络、博客、在线业务和传感器都会产生巨大的数据流。这些数据流以随时间变化的模式接收。虽然这些数据可以分配给特定的类别、对象和事件，但它们的分布不是恒定的。这些类别受分布变化的影响。这些分布变化通常是由于潜在的环境、地理、经济和文化背景的变化。例如，在2019冠状病毒病大流行期间，贷款申请的风险水平受到分布变化的影响。这是因为贷款风险是基于与申请人相关的因素，如就业状况和收入。这些因素通常相对稳定，但由于大流行的经济影响，变化很快。因此，现有的贷款推荐系统需要适应有限的例子。该项目将开发开放软件，以帮助用户评估在线公平性算法，减轻潜在的偏见，并检查效用与公平性之间的权衡。它将实现两个现实世界的应用：从视频数据中识别在线犯罪事件，从点击流数据中预测在线购买行为。为了扩大该项目在研究和教育方面的影响，该项目将为不同背景、性别和种族/民族的学生提供STEM课程。该项目包括研讨会、工作坊、短期课程和学生研究项目等活动。该项目旨在开发一种新的创新范式，用于设计、实施和评估在线公平感知深度学习（DL）模型。这种模型将用于噪声和非平稳数据流中的分类任务。该项目将侧重于五个方面。首先，该项目将探索如何确保在在线和非平稳环境下将各种公平原则纳入DL模型。该项目还将研究如何确定一个神经网络架构，以反映因果结构并适应分布变化。该项目还研究了深度学习模型如何学习原始参数（与模型准确性相关）和双参数（与模型公平性相关）的全局初始化。最后，该项目着眼于如何使在线学习算法对模型公平性估计中的不确定性具有鲁棒性，以及如何最终解释在线DL模型的公平性。通过连接神经架构搜索、在线元学习和公平感知深度学习技术等领域，该项目推进了人工智能公平性方面的最新研究。该项目将提供以下创新：(1)通过因果表示学习从原始特征中分离潜在的敏感和非敏感因果变量；(2)通过差分架构搜索，识别数据流的自适应架构；(3)通过在线中在线的方式学习原始模型参数和对偶模型参数的有效初始化；(4)开发鲁棒版本的算法来处理模型公平性和任务中的不确定性；(5)使用局部和全局解释识别负责模型自适应的训练样例和潜在因果变量。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Massive amounts of information are transferred constantly between different domains in the form of data streams. Social networks, blogs, online businesses, and sensors all generate immense data streams. Such data streams are received in patterns that change over time. While this data can be assigned to specific categories, objects and events, their distribution is not constant. These categories are subject to distribution shifts. These distribution shifts are often due to the changes in the underlying environmental, geographical, economic, and cultural contexts. For example, the risks levels in loan applications have been subject to distribution shifts during the COVID-19 pandemic. This is because loan risks are based on factors associated to the applicants, such as employment status and income. Such factors are usually relatively stable, but have changed rapidly due to the economic impact of the pandemic. As a result, existing loan recommendation systems need to be adapted to limited examples. This project will develop open software to help users evaluate online fairness-in algorithms, mitigate potential biases, and examine utility-fairness trade-offs. It will implement two real-world applications: online crime event recognition from video data and online purchase behavior prediction from click-stream data. To amplify the impact of this project in research and education, this project will leverage STEM programs for students with diverse backgrounds, gender and race/ethnicity. This project includes activities including seminars, workshops, short courses, and research projects for students. This project aims to develop a new and innovative paradigm for designing, implementing, and evaluating online fairness-aware Deep Learning (DL) models. Such models would be used for classification tasks in noisy and non-stationary data streams. This project will focus on five areas. First, the project will explore how to ensure a variety of fairness principles are incorporated in a DL model in online and non-stationary settings. The project will also look at how to identify a neural network architecture that will reflect the causal structure and be adaptive to distribution shifts. The project also looks at how the DL model will learn global initialization of primal parameters (associated with model accuracy) and dual parameters (associated with model fairness). Finally, the project looks at how to make online learning algorithms robust to uncertainties in model estimation of fairness and how to, ultimately, interpret the fairness of an online DL model. By bridging the areas of neural architecture search, online meta-learning, and fairness-aware deep learning techniques, this project advances state-of-the-art research in Fairness in AI. This project will offer the following innovations: (1) disentangle underlying sensitive and non-sensitive causal variables from raw features via causal representation learning; (2) identify adaptive architectures for data streams via differential architecture search; (3) learn effective initializations for both primal and dual model parameters in an online-within-online manner; (4) develop robust versions of the algorithms to deal with uncertainties in model fairness and tasks, and (5) identify the training examples and latent causal variables responsible for model adaption using local and global interpretations.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.