EAGER: Scalable, Content-Based, Domain-Agnostic Search of Scientific Data through Concise Topological Representations

尖端科学依赖于科学家筛选和访问最新研究产生的大量数据的能力。大部分数据存储在在线数据库中，只能通过使用特定的科学术语（如关键字，标签或描述）进行搜索。如果有人不知道正确的术语，他们往往无法访问所有可能对他们的研究有用的数据。通过以新的方式使用数学方法进行信息检索，该项目将研究是否可以为科学数据修改称为基于内容的搜索的强大搜索工具。如果成功，该项目将使数据用户无需确切知道要使用哪些关键字，从而改变科学家访问和共享数据的方式，并为具有截然不同专业知识的科学家创造新的合作机会。描述科学数据内容的一种特别有前途的方法是通过数据集的拓扑结构。因此，该项目将开发计算拓扑相似性的方法，这些方法比以前认为的更小，更快，更具可扩展性，目标是创建一种基于内容的科学数据交叉搜索方法。具体来说，研究人员将开发一个学习散列函数来将数据集的持久性图（其拓扑结构的常见编码）转换为简单的二进制代码。该哈希将被训练，使得代码之间的按位距离将保持数据集之间的拓扑相似性的度量。 这将把拓扑比较从当前昂贵的瓶颈状态转换为具有名义处理成本的状态，可以扩展到大型数据库查询。最初，该项目将专注于维护集群和邻域的二进制代码，最终开发保持秩或半度量的代码。研究人员还将探索在合成数据上训练学习散列函数的策略，目标是开发一种完全不受领域影响的基于内容的搜索方法。该奖项反映了NSF的法定使命，并被认为值得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来支持。

Cutting-edge science relies on scientists’ ability to sift through and access the massive amounts of data that are being produced by the latest research. Much of that data is stored in online databases and is searchable only by using specific, scientific terms, like keywords, tags, or descriptions. If someone doesn’t know exactly the right terms to use, they often can’t access all the data that might be useful for their research. By using mathematical approaches for information retrieval in a new way, this project will study whether a powerful search tool, called content-based search, can be modified for scientific data. If successful, this project will free data users from needing to know exactly which keywords to use, transforming how scientists are able to access and share data and creating new opportunities for scientists with vastly different expertise to work together.One particularly promising way to describe the content of scientific data is through a dataset’s topology. Therefore, this project will develop approaches to compute topological similarity that are smaller, faster, and more scalable than previously thought possible, with the goal of creating a method for cross-cutting, content-based search of scientific data. Specifically, the investigators will develop a learned-hash function to convert a dataset’s persistence diagram - the common encoding of its topology - to a simple binary code. This hash will be trained such that the bitwise distance between codes will maintain a measure of topological similarity between datasets. This will convert topological comparisons from the current state of an expensive bottleneck to one with nominal processing costs that can scale to large database queries. Initially, this project will focus on binary codes that maintain clusters and neighborhoods, ultimately developing codes that are rank or semi-metric preserving. The investigators will also explore strategies for training a learned-hash function on synthetic data, with the goal of developing a fully domain-oblivious approach to content-based search.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.