CAREER: Meta Clustering: Improving User Efficiency on Real-World Clustering Applications

本研究的目标是通过改变用户与聚类系统的交互方式，减少将聚类应用于现实问题所需的知识工程。Meta聚类不是让用户手动设计距离度量和修改聚类算法，而是通过生成许多不同的数据聚类，并将这些聚类呈现给组织为Meta聚类的用户，来帮助用户找到良好的聚类。Meta聚类是对聚类的聚类，它使用户更容易导航聚类空间以找到对他们有用的聚类。该项目的一个重要目标是使用现有的聚类方法，而不是用新方法取代它们。通过将一些负担从用户转移到计算机，该项目将对集群的实践，研究和教育产生重大影响。这将使聚类更容易，特别是对于不擅长定义距离度量的用户，帮助有经验的用户更快地达到良好的聚类，并通过提供比较和评估聚类方法的框架来推进聚类研究和教育。除了产生基础研究成果外，该研究项目还将用于培养博士生。学生，将涉及本科生和优秀的高中生在研究中，将有助于创建一个新的课程，无监督学习，将产生Meta聚类软件，可通过网络，是兼容许多现有的聚类算法，并将形成一个新的研讨会和基于网络的比较聚类方法的基础上运行不同的聚类研究人员。项目网址http://www.cs.cornell.edu/~caruana将用于及时传播结果。

The goal of this research is to reduce the knowledge engineering required to apply clustering to real-world problems by changing how users interact with clustering systems. Instead of having users manually engineer distance metrics and modify clustering algorithms, meta clustering helps users find good clusterings by generating many different clusterings of the data, and presenting these to users organized as a meta clustering. A meta clustering is a clustering of clusterings that makes it easier for users to navigate the clustering space to find clusterings useful to them. An important goal of the project is to work with existing clustering methods instead of replacing them with new methods. By shifting some of the burden from users to the computer, the project will have significant impact on the practice, research, and education in clustering. This will make clustering easier, particularly for users not adept at defining distance metrics, help experienced users arrive at good clusterings faster, and will advance clustering research and education by providing a framework for comparing and evaluating clustering methods. In addition to producing fundamental research results, the research project will be used to train Ph.D. students, will involve both undergraduate and exceptional high school students in research, will help in creating a new course on unsupervised learning, will yield meta clustering software available through the web that is compatible with many existing clustering algorithms, and will form the basis of a novel workshop and web-based comparison of clustering methods run by different clustering researchers. The project Web site http://www.cs.cornell.edu/~caruana will be used for timely results dissemination.