Scatter Balance based Semi-supervised Dimensional Reduction

Rui Yang, Xiangzhu Meng, Lin Feng

Abstract: Dimensionality reduction is an important research direction for machine learning. In recent years, semi-supervised learning has become a hot issue. Especially, SDA and SMMC are two representative semi-supervised dimensionality reduction methods, which mainly focus on the combination of local geometry structure and label information. For this reason, SDA and SMMC are deeply analyzed in this paper. However, the subspaces generated by these two methods only focus on within-class scatter or between-class scatter. Here we propose a new semi-supervised dimensionality reduction method that can balance them through the co-angle method. The experiment on real-world datasets can illustrate the efficiency of our method.