Find the intrinsic space for multiclass classification

Multiclass classification is one of the core problems in many applications. High classification accuracy is fundamental to be accepted as a valuable or even indispensable tool in the work flow. In the classification problem, each sample is usually represented as a vector of features. Most of the cases, some features are usually redundant or misleading, and high dimension is not necessary. Therefore, it is important to find the intrinsically lower dimensional space to get the most representative features that contain the best information for classification. In this paper, we propose a novel dimension reduction method for multiclass classification. Using the constraint of the triplet set, our proposed method projects the original high dimensional feature space to a much lower dimensional feature space. This method enables faster computation, reduce the space needed, and mostly importantly produces more meaningful representations that leads to better classification accuracy.