III: Small: Learning to Hash Information Networks

当今许多现实世界的系统都形成了信息网络，其中节点代表多类型实体，链接反映实体之间的关系。例子包括社会网络、用户-产品网络和知识图，它们分别存在于社会网络站点、电子商务系统和数字百科全书系统中。数据科学的一个主要目标是获得信息网络的有效表示，以支持各种分析任务，如相似性搜索、链接预测和节点排序，从而实现搜索和推荐等现实世界应用程序。先前的研究已经开发了用于推断网络数据表示的机器学习和优化解决方案。然而，这些方法仍然受到计算和存储方面的挑战，特别是在响应小型设备上的实时请求和计算时。该项目将开发学习网络数据表示的新方法，可以节省大量的计算成本和存储空间。最终，该项目将使由网络数据驱动的应用程序更有效地运行。该项目将开发新的机器学习方法，用于散列同质和异质信息网络。除了结构信息外，所开发的方法还将考虑许多信息网络可能可用的属性信息。由于形式化的学习目标是二元决策变量的np困难问题，本项目将通过探索解决学习问题的两个不同方向，开发新的优化方法来解决学习问题。第一个是创造性地将形式化的学习问题转化为经过充分研究的最大切割问题，然后利用现有的方法来解决转化后的问题。二是研究已形式化学习问题的连续优化重新表述，然后发展凸逼近来解决重新表述的连续优化问题。学习到的信息网络的二进制表示将应用于几个重要的应用任务，如相似节点搜索、节点分类、链接预测和推荐。将使用多个信息网络数据集来评估基于不同定量评估指标的开发方法的性能。本项目开发的信息网络哈希解决方案将对信息网络的表示学习、网络数据分析的数据挖掘和机器学习、机器学习问题的优化等研究领域产生重大推动作用。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。

Many of today's real-world systems form information networks, where nodes represent multi-type entities and links reflect relation between entities. Examples include social networks, user-product networks, and knowledge graphs, that exist in social networking sites, e-commerce systems, and digital encyclopedia systems, respectively. One major goal of data science is to obtain effective representations of information networks to support various analytical tasks such as similarity search, link prediction, and node ranking that enable real-world applications such as search and recommendation. Prior research has developed machine learning and optimization solutions for inferring network data representations. However, these approaches still suffer from computation and storage challenges, especially while responding to real-time requests and computing on small devices. This project will develop novel methods for learning network data representations that could save much computational cost and storage space. Ultimately, the project will make the applications driven by network data run more effectively and efficiently. This project will develop novel machine learning methods for hashing both homogeneous and heterogenous information networks. In addition to structure information, the developed approach will also consider attribute information that may be available for many information networks. As the formalized learning objectives are NP-hard problems with binary decision variables, this project will develop new optimization methods for solving the learning problems by exploring two different directions for solving the learning problems. The first one is to creatively transform the formalized learning problems to well-studied Max-Cut problems and then leverage existing approaches to solve the transformed problems. The other one is to study continuous optimization reformulations for the formalized learning problems and then develop convex approximations to solve the reformulated continuous optimization problems. The learned binary representations of information networks will be applied to several important applications tasks such as similar node search, node classification, link prediction, and recommendations. Multiple information network datasets will be used to evaluate the performance of developed methods based on different quantitative evaluation metrics. The developed information network hashing solutions in this project will significantly advance the research fields of representation learning of information networks, data mining and machine learning for network data analysis, and optimization for machine learning problems.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.