Combining granular clustering and classifications in knowledge based networks

点击翻译按钮获取中文摘要

Knowledge acquired from databases may be fragmented for two principal reasons: the databases may be physically stored in different locations distributed on a network, or it may be computationally advantageous to apply a data mining technique to subsets of a massive database. In such cases, combining fragmented knowledge is an important issue. Our long term goal is to combine granular clustering and classification results in a knowledge-oriented network. Granular computing typically encompasses algorithms based on fuzzy, rough, and interval sets, which were initially used in classification problems. Recently, researchers have shown that these three representations are useful for modeling overlapping clusters. Fuzzy clustering makes it possible to specify the degree with which a given object belongs to a cluster, instead of a binary description of the membership. The interval and rough set representations allow us to state that an object may belong to more than one cluster. Our proposals that use interval set representations of clusters have been well received by the research community. These approaches were based on Genetic Algorithms, K-means, and Kohonen Self-organizing Maps. We are also collaborating with other researchers on support vector interval set clustering. These interval set representations are more flexible than the conventional crisp clusters and less verbose than the fuzzy clusters. The distributed pieces of databases in a knowledge-oriented network may differ from each other because they contain different features and/or they contain different objects. For our short term goal, we will focus on databases that differ in features used to represent the objects. These features may be semantically different from each other, such as customer spending and customer visits. Or, they may be the same attributes from different time periods, for example, spending in summer versus spending in winter. When two or more significantly different criteria are used in the classification or clustering, it may be desirable to create separate models for each criteria, and then combine the results. We plan to extend this concept by combining conventional and granular classifications and clustering from different time periods.