CAREER: Towards Responsible Graph Neural Networks

图神经网络(GNN)被誉为关系数据(即图形)机器学习的突破，它将允许神经网络在计算机视觉或自然语言处理中实现同样的“人工智能复兴”，最近已成为关系数据推理的首选算法之一。然而，尽管GNN在学术数据集上取得了成功，但人们对它们的性质仍然知之甚少。这严重影响了它们在实际环境中的使用，在实际环境中，对严谨、透明和可靠性的要求是至高无上的。为此，这项建议侧重于两个关键研究议程：(1)通过实验和理论确定国民总收入产出的性质、可靠性和敏感性；(2)推进对图形估计器统计性质的理论理解，从而为发展改进的国民总收入奠定更坚实的基础。由此产生的算法将被部署在各种新颖的应用中，旨在从数据中学习可理解和可靠的表示。将努力通过各种手段促进和推广这些方法的使用，包括开发开放源码软件，创建新的研究生课程，以及支持调查员大学与当地社区学院合作的倡议。该项目研究了数据、GNN体系结构参数(如嵌入距离或卷积运算符)和由此产生的嵌入空间几何之间的关系。总体目标是确定不同GNN架构能够实现最佳性能的具体条件。为此，研究人员建议利用丰富的统计文献中关于高维统计和基于图表的正则化的见解。这种方法试图将GNN视为流形上函数的估计器，以便(A)分析GNN可以有效学习的函数类型，以及(B)从包含图形正则化的高维模型中进行比较并获得洞察力。通过追求这些研究方向，该提案旨在将GNN从黑盒模型转变为可解释、值得信任和可靠的可操作分析管道。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。

Heralded as the breakthrough for machine learning on relational data (i.e., graphs) that would allow the same “AI renaissance” that Neural Networks have achieved in Computer Vision or Natural Language Processing, Graph Neural Networks (GNNs) have recently emerged as one of the preferred algorithms for reasoning on relational data. Yet, despite GNNs’ success on academic datasets, their properties remain ill-understood. This severely compromises their use in practical settings, where the requirements for rigor, transparency and reliability are paramount. In response, this proposal focuses on two key research agendas: (1) characterizing the properties, reliability and sensitivity of GNN outputs through experiments and theory; and (2) advancing the theoretical understanding of statistical properties in graph estimators, thereby providing a stronger foundation for the development of improved GNNs. The resulting algorithms will be deployed in diverse novel applications aimed at learning understandable and reliable representations from data. Efforts will be made to promote and spread the utilization of these methods through various means, including developing open-source software, creating new graduate courses, and supporting the investigator's university initiative to collaborate with local community colleges. The project investigates the relationship between data, GNN architecture parameters (such as embedding distance or convolution operator), and the resulting embedding space geometry. The overarching objective is to identify specific conditions under which different GNN architectures can achieve optimal performance. To this end, the investigator suggests leveraging insights from the rich statistics literature on high-dimensional statistics and graph-based regularization. This approach seeks to view GNNs as estimators for functions on a manifold in order to (a) analyze the types of functions that GNNs can effectively learn, and (b) draw comparisons and gain insights from high-dimensional models incorporating graph regularization. By pursuing these research directions, this proposal aims to transform GNNs from black-box models into actionable analysis pipelines that are explainable, trustworthy, and reliable.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.