University of Houston

The goal of region discovery is to find interesting places in spatial datasets. Region discovery can be viewed as a clustering problem in which objects in the datasets are grouped into regions whose characteristics match a domain expert’s notion of interestingness. The use of clustering algorithms for region discovery faces several challenges: domain experts frequently look for clusters which exhibit additional characteristics that go far beyond the capabilities of traditional clustering algorithms. Other challenges include identifying arbitrary shaped clusters, selecting parameters of clustering algorithms to obtaining high quality clusters, and finding alternative clustering. Finally, it is desirable to have clustering algorithms that are capable of dealing with multiple objectives in parallel. This dissertation introduces novel spatial clustering frameworks and algorithms to address the aforementioned challenges in a highly automated fashion. It centers on developing clustering algorithms that support plug-in fitness functions to aid region discovery. First, we propose a generic agglomerative, spatial clustering framework for the discovery of clusters of arbitrary shape. Second, we propose a methodology, architecture and algorithms for multi-objective, multi-run clustering that derive interesting clusters regarding two or more of those objectives in parallel. The proposed framework and algorithms are evaluated in three case studies that centering co-location mining, agglomerative clustering, and comparing clustering algorithms.