A Learning Method of Feature Selection for Rough Classification

In this paper, we present a new method of learning a feature selection dictionary for rough classification. In the learning stage, both the n-dimensional learning vectors and the n-dimensional reference vectors are transformed into an m(<n)-dimensional learning vector and the m-dimensional reference vector, respectively, using a current feature selection dictionary. The feature selection dictionary is then successively modified for each learning vector so as to decrease the distance between the learning vector and the m-dimensional reference vector corresponding to the correct category. Furthermore, the feature selection dictionary is modified for each learning vector so as to increase the distance between the learning vector and the m-dimensional reference vector that is the nearest incorrect reference vector of the learning vector. The experimental results showed that our method's processing time is 9 times faster than that without rough classification, even if the recognition rates are the same.