An Extension of COSSO Algorithm by Combining Variables

In the situation in which the number of variables(p) is comparatively larger than the number of sample(n) and there exist variables which have the same effects in the classifier, this paper presents a new variable selection algorithm extending COSSO (COmponent Selection and Smoothing Operator) by utilizing grouped variables. A grouped variable is composed of a component in the kernel expansion of a classifier. The construction of a kernel classifier is carried out by COSSO with the components made from grouped variables. From the results of simulations with synthetic and real data sets, the proposed method is verified to improve the performance of a kernel classifier and enhance the interpretability of the model.