Spatial clustering of spatial correlated data, multi-valued data, and domain-driven spatial clustering ensemble

空间聚类根据数据对象在空间中的距离、密度和可达性将相似的数据对象划分为同一组(称为簇)。本研究的主要目标是设计、开发和改进一套空间聚类方法和过程，从大量的空间相关数据、多值空间数据和多特征空间数据中发现有意义和有用的聚类。具体地说，本研究将解决以下空间聚类问题：1)空间数据集中普遍存在空间相关性。它表明了空间属性和非空间属性之间的依赖关系，有可能是某种因果关系导致的。本研究将确定非空间属性相似性和空间相关性的不同无偏度量。然后将这些度量集成到不同的聚类方法中。2)在许多应用中，空间对象的非空间属性可以用d维空间中的一组多个值来表示。该对象称为多值空间对象。如何发现多值空间对象的聚类是一项具有挑战性的任务。本研究的一个成果是为多值空间对象定义了不同的相似性度量。将根据这些措施开发不同的空间聚类方法。3)现有的大多数聚类算法在聚类过程中没有考虑领域知识，这使得用户无法准确地描述自己的目标和理解聚类结果。此外，由于空间数据的多重特性，现有的聚类方法都不适用于许多应用。每种聚类方法可以对来自特定特征的数据进行聚类，或者可以适合于空间区域的一部分。如何借助领域知识提高多特征空间数据的聚类效果是一个有趣的研究课题。本研究将首先研究如何完善空间聚类本体。然后将提出新的领域驱动的空间聚类方法和聚类集成方法，以组合多个聚类结果来产生有用的和有意义的结果。

Spatial clustering partitions similar data objects into the same group (called clusters) based on their distance, density and reachability in space. The primary goal of this study is to design, develop and improve a set of spatial clustering methods and processes that discover meaningful and useful clusters from large amounts of spatial correlated data, multi-valued spatial data and multi-featured spatial data. Specifically, the research will address the following spatial clustering problems: 1) Spatial correlation generally exists in spatial datasets. It indicates the dependency between the spatial and non-spatial attributes, with the chance that some cause and effect lead to it. The research will identify different unbiased measures of non-spatial attribute similarity and spatial correlation. Then the measures will be integrated into different clustering methods. 2) In many applications, the non-spatial attribute of a spatial object can be represented by a set of multiple values in the d-dimensional space. The object is called a multi-valued spatial object. How to discover the clusters of multi-valued spatial objects is a challenging task. One outcome of this research is to define different similarity measures for multi-valued spatial objects. Different spatial clustering methods will be developed based on the measures. 3) Most existing clustering algorithms do not consider the domain knowledge during the clustering process, which prevents users from precisely describing their goals and understanding the clustering results. In addition, for many applications, none of the current clustering methods is suitable due to the multiple features of the spatial data. Each clustering method may cluster the data from a particular feature or may be suitable for part of the spatial area. How to improve the clustering result for spatial data with multiple features with the aid of domain knowledge is an interesting research topic. This research will first study how to improve the spatial clustering ontology. The new domain-driven spatial cluster methods and cluster ensemble methods will then be proposed to combine multiple clustering results to produce useful and meaningful results.