CRII: Interpretable Influence Propagating and Blocking on Graphs

随着包括虚拟(例如，社交网络)和物理接地(例如，交通网络)的网络的复杂性增加，需要了解网络影响的传播是至关重要的。网络影响力因其深远的社会和商业影响而日益受到研究者的关注。例如，信息传播或病毒营销的目标是确定能够影响大量其他人的最突出的潮流引领者，而流行病学的主要目标是确定谁最有可能传播疾病，这有助于制定疫苗和检疫法规。该项目将开发新的工具来分析传播网络的结构和初始状态如何最大化影响流，然后研究控制流的政策选择。这个项目的主要创新将是它能够学习流动和图形几何结构之间的复杂关系，并为决策者提取可理解的规则。主要的挑战是跨结构和属性组合的大量变量组合，以改变影响流。这个项目倡导一种独特的范例，用于学习图上影响流的可解释表示，特别强调解开由图的几何结构和种子选择所施加的组合限制。这项研究旨在开发一种新的框架，在分离影响过程的同时提高准确性和可解释性。由于影响建模的特点，包括复杂的错综复杂的拓扑链、数据不足和汇合效应，使得开发可解释模型的难度变得更加复杂。研究人员将使用情景感知约束和补充观察来缩小搜索范围并确定影响来源。调查员将对有效和高效的控制政策进行深入评估，以改善影响传播或阻止。这个项目将解决以下三个基本研究问题：学习影响的解释性拓扑依赖；学习影响级联和来源；以及学习控制影响流。所提出的模型将根据当前受影响的区域来确定使未来影响最小化的最优种子，这需要对大量依赖于图的元素的交互作用进行建模。根据热力学第二定律，流量是由系统内的能级决定的，这促使我们进一步检查图中的熵概念，以提供更可靠的评估。研究人员将使用全局敏感度分析和扰动矩阵理论来选择最小但最关键和最健壮的集合。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。

As networks including virtual (e.g., social networks) and physically grounded (e.g., transportation networks) increase in complexity, the need to understanding the spread of network influence is crucial. Influence in networks has been the subject of increasing attention among researchers due to its far-reaching social and commercial implications. For instance, the objective of information propagation or viral marketing is to identify the most prominent trend setters capable of influencing vast numbers of others, while the primary objective of epidemiology is to ascertain who is most likely to spread a disease, which aids in the development of vaccine and quarantine regulations. This project will develop novel tools to analyze how the spreading network's structure and initial state maximize the influence flow, and then investigate policy options for controlling the flows. The primary innovation of this project will be its ability to learn the complex relationship between flows and the geometric structure of graphs and extract understandable rules for decision-makers. The main challenge is the huge number of combinations of variables combined across structures and attributes to alter influence flows. This project advocates a unique paradigm for learning interpretable representations of influence flow over graphs, with a particular emphasis on disentangling the combinatorial limitations imposed by both the graph's geometric structure and seed selection. This research aims at developing a new framework that boosts accuracy and interpretability while decoupling the influence process. The difficulty of developing interpretable models is compounded by the specific characteristics of influence modeling, which include complex tangled topological links, insufficient data, and confluence effects. The investigator will use context-aware constraints and complementing observations to narrow the search and determine the effect source. The investigator will perform an in-depth assessment of effective and efficient control policies to improve influence propagation or blocking. This project will address the following three fundamental research issues: learning expository topological dependence of influence; Learning the influence cascade and the sources; and learning to control the influence flows. The proposed model will determine the optimal seeds that minimize future influence based on the currently influenced region, which requires modeling the interaction of numerous graph-dependent elements. According to thermodynamics' second law, the flow rate is determined by the energy levels within the system, which motivates us to examine graph entropy notions further to provide a more robust assessment of them. The investigator will employ global sensitivity analysis and perturbation matrix theory to choose the smallest yet most critical and robust set.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.