A Study on Symbolic Data Analysis.

点击翻译按钮获取中文摘要

A main theme in the term 1997-1998 is to establish the feature selection algorithm in classification problems. In our desired feature selection method, we have to evaluate a given feature subset simultaneously from the two viewpoints : the effectiveness of class discrimination ; and the effectiveness of the generality for class descriptions. For this purpose, we introduced a new graph concept so called the Interclass Mutual Neighborhood Graph (IMNG) and we constructed several new feature selection algorithms.Another research theme is to establish a feature selection method which is able to detect "geometrically thin structure in global sense" imbedded in multidimensional symbolic data In this research term, we found a sufficiently effective feature selection method which is based on a simple and intuitively clear principle that "If the given symbolic data has a functional structure, then the data has a geometrically thin structure". This method is useful and powerful as a preprocessing tool for functional identification problems by neural networks. A part of our results was reported in the international conference held at Luxembourg (KESDA-98) and was remarked as a new powerful engine for data mining.