CNS Core: Small: Causal Reasoning for Data-Driven Networking

通过Internet提供高质量服务的一个关键部分是优化底层网络的性能。随着数据分析和机器学习的兴起，网络提供商已经能够稳步改善用户体验。然而，任何等待网页或视频加载很长时间的人都知道，仍然需要更好的网络服务。该项目解决了这一挑战，并通过将不同形式的机器学习应用于网络来开辟新天地，这可能会带来巨大的好处。这里，我们的目标是回答假设问题——即，给定现有部署系统的记录数据，如果我们改变系统的设计（例如，部署新的视频流算法），性能会有什么影响？回答这种性质的假设问题被称为因果推理，它考虑的是在记录数据时没有发生的事件的影响。广泛使用的机器学习（ML）工具（例如，现成的神经网络）不足以进行因果推理，因为它们只是捕获收集到的数据中的相关性。这将它们限制在预测与现有设计的部署系统在未来的表现有关，并且当面对“假设”问题时，它们会产生偏差。其他方法，如随机对照试验，可能会破坏现场用户的表现。该项目旨在通过允许运营商对被动收集的网络数据提出关于新网络算法的假设问题，显著提高互联网服务的用户体验，同时使这些服务的运营成本更低（可以传递给消费者的成本节约）。这个项目将把网络系统、机器学习和因果推理结合在一起。更具体地说，该项目将开发：(i)现实世界网络案例研究的因果关系图，以及考虑网络测量的便利性、估计量方差和预测准确性的因果调整方法；通过网络模型推断潜在变量的新方法，并通过仪器收集额外的数据，这些数据可以作为这些潜在变量的代理，而不影响现场用户；（iii）连接机器学习技术的技术，用于推断潜在变量和建模所提出的变化，并通过合并网络模型提出独特的挑战。这个项目中的思想可以更广泛地应用于许多网络和机器学习领域。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。

A critical part of providing high quality services over the Internet is optimizing the performance of the underlying networks. With the rise of data analytics and machine learning, network providers have been able to steadily improve the user experience. However, as anyone who has waited a long time for a webpage or video to load knows, there remains a need for ever-better network service. This project tackles this challenge and breaks new ground by applying a different form of machine learning to networking that could potentially offer significant benefits. Here, the goal is answering what-if questions -- i.e., given recorded data of an existing deployed system, what would be the performance impact if we changed the design of the system (e.g., deployed a new video streaming algorithm)? Answering what-if questions of this nature is known as causal reasoning, which considers the effect of events that did not occur while the data was being recorded. Widely used machine learning (ML) tools (e.g., off-the-shelf neural networks) are inadequate for causal reasoning since they merely capture correlations in collected data. This limits them to predictions that pertain to how a deployed system with an existing design performs in the future, and they incur biases when faced with "what-if" questions. Other approaches such as Randomized Control Trials may be disruptive to the performance of live users. This project aims to significantly increase the user experience of Internet services, while making these services cheaper to operate (cost savings that can be passed on to consumers), by allowing operators to ask what-if questions about new networking algorithms on passively collected networked data.This project will bring together network systems, machine learning, and causal inference. More specifically, this project will develop: (i) Causal dependency graphs for real-world networking case studies, and approaches for causal adjustments that consider the ease of measurement in networking, estimator variance, and accuracy of predictions; (ii) Novel approaches to infer latent variables through network models, and through instrumentation to collect additional data that can serve as proxies for these latent variables in a manner that does not impact live users; and (iii) Techniques to bridge ML techniques for inferring latent variables and modeling the proposed changes, with unique challenges posed by incorporating network models. The ideas in this project can be more generally applied in many networking and machine learning domains.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.