抄録
The nearest neighbor (NN) methods solve classification problem by storing examples as points in a feature space, which requires some means of measuring distances between examples. However, it suffers from the existence of noisy attributes. One resolution is to modify the distance of similarity degree using attribute weights, which can not only decrease the influence of noisy attributes, but also subset relevant attributes. In this paper,a rough genetic algorithm (RGA) proposed by Lingras and Davies is applied to the classification problem under an undetermined environment, based on a fuzzy distance function by calculating attribute weights. The RGA can complement the existing tools developed in rough computing. Computational experiments are conducted on benchmark problems, downloaded from UCI machine learning databases. Experimental results,compared with a usual GA[1] and the C4.5 algorithms, verify the efficiency of the developed algorithm. Furthermore, the weights learned by the proposed learning method is applicable to not only fuzzy similarity functions but also any similarity functions. As an application, a new distance metric, weighted discretized value difference metric (WDVDM), is proposed. Experimental results show that the WDVDM improves the discretized value difference metric (DVDM).
