Fractional-step max-min distance analysis for dimension reduction

Guowan Shao,Nong Sang
2012-01-01
Abstract:Dimensionality reduction has been regarded as a key step for high-dimensional data processing and analysis. Max-min distance analysis (MMDA) for dimension reduction is proposed to solve the class separation problem and the minimum pairwise distance between class centers is maximized in the low-dimensional subspace. However, the proposed algorithm ignores the distribution of class centers. Despite of the max-min pairwise distance, the nonuniform distribution of class centers may lead to a suboptimal classification rate. In this paper, we propose a novel method named fractional-step max-min distance analysis (FMMDA). The proposed method maintains excellent class separation and obtains a relatively uniform distribution of class centers by relaxing the max-min pairwise distance in fractional steps. Moreover, we present a dual form of the optimization problem in FMMDA to reduce the computational load of the optimization procedure. Empirical studies demonstrate that the proposed FMMDA significantly outperforms MMDA.
What problem does this paper attempt to address?