A new method for one-dimensional linear feature transformations

摘要

We propose a method for finding a linear transformation of an initial pattern space into a one-dimensional new space, which optimizes the L2 distance between density functions. We use an orthogonal expansion with Hermite functions to compute the criterion; we discuss both the truncation and the statistical variation error for this expansion. To avoid false local optima caused by sample variation, we choose the starting point with a coarse step optimization method. Experimental results with two-class and multi-class, both unimodal and multimodal, are presented.