Fuzzy Variable Selection with Degree of Classification Based on Dissimilarity between Distributions of Variables
Mika Sato‐Ilic · International journal of intelligence technologies and applied statistics · 2008
This paper proposes a method of variable selection to reduce the redundant variables holding the classification structure of data. The classification structure is defined as the weights of objects with respect to variables and obtained by using fuzzy clustering. From the Kullback-Leibler divergence, under an assumption of variable normal distribution for the weights and data with respect to a variable, we use the derived dissimilarity between the distributions in which each distribution corresponds to each variable. Applying the dissimilarity to multidimensional scaling (MDS), configuration values for each variable in lower dimensional space are obtained. Then, from the distance of the configuration values, we select the useful variables that involve the classification structure of the data. Several numerical examples show the better performance of our proposed method.