CAREER: Bridging the Data-Model Gap -- Leveraging Surveillance for Propagation Mining over Networks

该研究的长期目标是理解、有效管理和利用动态机制，如在自然、社会和技术系统中发生的大型网络上的传播。了解了这些过程，我们就可以为自己的利益操纵它们。传播和网络在公共卫生和流行病学、系统生物学、网络安全、病毒营销和社交媒体等各个领域都有许多应用，因此，这一领域的进展有望带来科学、商业和社会效益。提出的研究旨在为传播相关问题开发可扩展的、数据驱动的框架，从而获得更多可实现和可推广的工具。PI的调查将导致新的挖掘和学习问题以及可扩展的技术，这些技术可以应用于大规模数据集，有助于为未来做出更明智的选择。教育活动也与这一研究议程紧密结合，包括通过课程、辅导和其他大学项目将研究与教育结合起来。目前大多数传播挖掘工作都假定存在校准良好的模型。执行模型校准通常非常昂贵，而且不健壮。实际上，在许多情况下，不清楚应该校准哪个参数化模型。然而，越来越多的监测数据，如在线媒体和医疗健康记录的可用性。PI方法的独特之处在于，其目的是直接使用监测数据，并根据数据和网络共同制定优化问题。提出的问题包括为流感等疾病发明数据驱动的免疫策略，在级联数据集中自动发现缺失感染/激活，以及基于传播数据和网络的分布式特征表示自动学习图摘要。PI建议为所有这些问题开发一个灵活且富有表现力的框架。此外，开发的算法将应用于各种领域，利用多种协作。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。

The long-term goal of the proposed research is to understand, manage efficiently, and utilize dynamical mechanisms like propagation on large networks, occurring across natural, social, and technological systems. Understanding such processes enables us to manipulate them for our benefit. Propagation and networks have numerous applications in areas as diverse as public health and epidemiology, systems biology, cyber security, viral marketing and social media---hence progress in this domain promises scientific, commercial and social benefits. The proposed research aims to develop extensible, data-driven frameworks for propagation-related problems getting more implementable and generalizable tools. The PI's investigations will lead to novel mining and learning problems and scalable techniques which can be applied to massive datasets, helping make more informed choices for future. Educational activities are also closely integrated with this research agenda, including integrating research with education through courses, tutorials, and other university programs. Most current work in propagation mining assume the existence of well-calibrated models. Performing model calibration is typically very expensive, and not robust. Indeed, in many situations it is not clear which parameterized model should be calibrated. However there is an increasing availability of surveillance data like online media and medical health records. The PI's approach is unique in the sense that the aim is to directly use surveillance data and formulate optimization problems based on the data and network together. The proposed problems include inventing data-driven immunization policies for diseases like influenza, automatically finding missing infections/activations in cascade datasets, and automatically learning graph summaries based on distributed feature representations of propagation data as well as the network. The PI proposes to develop a flexible and expressive framework for all these problems. In addition, the developed algorithms will be applied to various domains, leveraging multiple collaborations.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.