A Concept-based Feature Extraction Approach

摘要

A concept has a perceived property and a set of constituents. The goal of this investigation is about extraction of meaningful relationships, if any, between the perceived property and the constituent's attributes. Such meaningful relationships (features) may be used as a prediction tool. The presented methodology for extracting the features is based on the concept expansion. To the best of our knowledge, feature extractions based on a concept expansion approach, for use in data mining, has not been reported in the literature. The goal was met by introducing the b-concept, conceptualizing a universe of objects using b-concept, and generating the complete gamma-expansion (CGE) of the b-concepts. The features were extracted from CGEs as anchor prediction (AP) rules. The AP rules were crystalized by a sequence of horizontal-vertical reductions. The prediction powers of the AP rules and their crystalized version were investigated by: (i) using 10 pairs of training and test sets, and (ii) comparing their performances with the performance of the well-known ID3 approach over the same training and test sets. The results revealed that the AP rules and ID3 have similar performances. However, the crystallized prediction rules have a superior performance over the AP rules and ID3. The average of the correct prediction is up by 17%, the average of the false positive is down by 13%, and the average of false negative is up by 3%. In addition, the number of test objects that cannot be predicted is down by 7%.