Confidential Reasoning about Data Using Abstract Types as Meaningful Proxies

拟议研究的目标是制定一种评价和改进的方法和相关系统，以便将大量机密数据记录转换为一组抽象类型，以捕捉数据集中的重要趋势和模式，而不披露任何个别数据记录中的信息。由此产生的抽象类型学预计有两个主要用途。首先，类型学将代表数据集的定性视图，这可能有助于组织(如医院或大型零售商)理解数据的整体结构。其次，从机密数据派生的抽象类型将通知应用程序，这些应用程序根据嵌入在该数据中的知识提出建议。 拟议的方法将使用1)特征提取(将数据变量具体化为代表数据记录中的主要差异来源的表示特征)、2)数据的迭代聚类(使用不向人暴露个别数据记录的方法)和3)由领域专家在每次迭代中创建的类型审查来确定其代表性，从而从机密数据中提取类型。 这项研究将带来两大创新。首先，一种创建数据代理的方法，该方法可以服务于个性化和推荐系统的需求，而不需要检查机密数据。第二，一组派生的抽象类型，可以作为大量个人及其机密数据的有用代理。这项研究应该导致开发确定数据代理的方法，这些方法可以释放机密数据中的知识，并允许将其输出用于研究，而不暴露(或冒着暴露)基础数据的风险。应该从这项研究中受益的应用实例包括临床决策支持和事件推荐。

The goal of the proposed research is to develop a methodology and associated system of evaluation and refinement, for converting a large set of confidential data records into a set of abstract types that capture important trends and patterns in the data set without disclosing the information in any of the individual data records. Two major uses are anticipated for the resulting abstracted typology. First, the typology will represent a qualitative view of the data set that may help an organization such as a hospital or major retailer in understanding the overall structure of data. Second, abstract types derived from confidential data will inform applications that make recommendations based on knowledge embedded in that data. The proposed methodology will abstract typologies from confidential data using 1) feature extraction (reification of data variables into representation features that are representative of the major sources of variance in the data records), 2) iterative clustering of the data (using methods that do not expose individual data records to human view) and 3) review of types, created in each iteration, by domain experts to determine their representativeness. This research will lead to two major innovations. First, a methodology for creating data proxies that can service the needs of personalization and recommendation systems without requiring inspection of confidential data. Second, a set of derived abstract types that can stand as useful proxies for a large set of individuals and their confidential data. This research should lead to development of methods for identifying data proxies that can unleash the knowledge in confidential data and allow it to be exported for research use without exposing (or risking exposure of) the underlying data. Examples of applications that should benefit from this research include clinical decision support and event recommendation.