NSF-CSIRO: Towards Interpretable and Responsible Graph Modeling for Dynamic Systems

现实世界的自然和工程系统（例如，食物网、电网、河流网络和洋流网络）本质上是复杂的，并由许多具有依赖关系的因素驱动。图通常用于表示这些系统的结构和内容，以进行事件预测和风险估计。迄今为止，已经提出了许多图学习方法，如图神经网络，但主要是针对静态图。在动态系统中，结构和内容同时随着新出现的趋势和事件而发展，这使得很难理解和解释图形的每个部分是如何形成可靠的预测模型的。该项目致力于通过结合传感器模式发现、节点交互和网络功能分析以及物理和知识学习，为动态系统构建一个图形学习和解释框架。该项目将提出使用图形建模和理解大规模动态系统的新算法，并为领域专家开发一个原型，以分析他们的数据，解释系统中当前发生的事情，了解由此产生的后果，并提供可能的缓解策略。美国和澳大利亚团队的共同努力将有助于了解/揭示不同地形类型、内陆和沿海水交换、有毒藻华以及农村和地区社区恢复力的水监测系统的动态。该项目包括三个主要重点：(1)传感器信号的特征提取和理解；(2)动态网络节点建模与解释；(3)动态网络的功能性和可信度。该研究将研究信号片段模式（SSP）提取和交互分析，以了解在重大事件出现期间特征如何相互作用。在节点级，新的时间编码和时空图神经网络将用于学习节点事件预测和异常检测的早期预警模型。对节点交互的研究将回答两个节点为何、何时以及如何相互交互。在节点级解释之外，该项目将针对图的功能单元，估计每个快照图的贡献，并定位与输出系统相关的最重要的子图。基于扰动的事后解释器将提供反事实的解释，以增强动态图神经网络系统的可解释性和可信度。该研究还将研究将物理定律和领域知识结合到动态图神经网络中，以开发数据高效、鲁棒和负责任的图建模框架。这是美国和澳大利亚研究人员之间的一个联合项目，由美国国家科学基金会和澳大利亚联邦科学与工业研究组织（CSIRO）的负责任和公平人工智能合作机会资助。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。

Real-world natural and engineered systems (e.g., food web, power grids, river networks, and ocean current networks) are inherently complicated and are driven by many factors with dependency relationships. Graphs have been commonly used to represent the structure and content of these systems for event prediction and risk estimation. To date, many graph learning methods, such as graph neural networks, have been proposed, but primarily for static graphs. In dynamic systems, the structure and content are simultaneously evolving in response to emerging trends and events, making it difficult to understand and interpret how each part of the graph functions in forming reliable models for predictions. This project strives to build a graph learning and interpretation framework for dynamic systems by combining sensor pattern discovery, node interaction and network functionality analysis, and physics- and knowledge-informed learning. The project will propose new algorithms for modeling and understanding large-scale dynamic systems using graphs, as well as develop a prototype for domain experts to analyze their data, explain what is currently happening in the system, understand the resulting consequences, and provide possible mitigation strategies. The joint effort between the US and Australian teams will help understand/uncover the dynamics of water monitoring systems for different terrain types, inland and coastal water exchange, toxic algal blooms, and resilience of rural and regional communities.The project includes three main thrusts: (1) sensor signal to feature extraction and understanding; (2) dynamic network node modeling and interpretation; and (3) dynamic network functionality and trustworthiness. The research will study signal snippet pattern (SSP) extraction and interaction analysis to understand how features interact with each other during the emergence of significant events. At the node level, new temporal encoding and spatial-temporal graph neural networks will be used to learn models for node event prediction and anomaly detection for early warning. The study of node interaction will answer why, when, and how two nodes may be interacting with each other. Beyond node level interpretation, the project will target graph functional units, estimate each snapshot graph’s contribution, and locate subgraphs with the highest significance concerning output systems. A perturbation-based post-hoc explainer will provide counterfactual explanations to enhance the explainability and trustworthiness of dynamic graph neural network systems. The research will also investigate combining physics laws and domain knowledge into dynamic graph neural networks to develop a data-efficient, robust, and responsible graph modeling framework. This is a joint project between U.S. and Australian researchers funded by the Collaboration Opportunities in Responsible and Equitable AI under the U.S. NSF and the Australian Commonwealth Scientific and Industrial Research Organisation (CSIRO).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.