Recursive and iterative clustering in granular hierarchical, network, and temporal datasets

点击翻译按钮获取中文摘要

Clustering is one of the frequently used unsupervised data mining techniques for grouping similar objects. The proposed research program will investigate a novel iterative approach to clustering in a granular environment. An information granule represents an object. For example, a customer with certain purchasing patterns could be represented by an information granule. A granule is usually connected to other granules. For example, in a hierarchical environment, a customer granule will be connected to a number of product granules and vice versa. In a granular network, phone users are connected to other phone users. In a granular temporal environment, a daily pattern of events is connected to historical and future daily patterns. Traditionally, clustering of granules is done in isolation without any information on clustering of the connected granules. The primary theme of the proposed research is to simultaneously cluster all the granules iteratively. Each iteration will use results of previous clustering of connected granules, until a stable clustering of all the granules is achieved. In a hierarchical environment such as customers and products, it will mean that clustering of customers uses profiles of product clusters, and vice versa. For networked granules, a phone user is clustered using cluster profiles of the other connected users. In a temporal granular clustering, daily patterns will be clustered based on clustered profiles of historical and future patterns. These repeated applications of clustering are termed iterative in a hierarchy and are termed recursive in networks. The integrated meta-clustering of hierarchical, network, and temporal data is a multi-faceted project. Since clustering is unsupervised and we do not know the expected outcomes, it is important to study the quality of the resultant clustering. In addition to deriving quantitative evaluations, the notion of preference will be used to value a cluster based on how well-connected it is to more desirable objects. The iterative and recursive algorithms will be further modified for fuzzy and rough clustering, which allow an object to belong to multiple clusters. We plan to design, develop, implement, and test variations of the clustering algorithms for retail, mobile phone, engineering, and financial datasets.