Large Margin Dimension Reduction for Sparse Image Classification
Ke Huang, Selin Aviyente · 2007 IEEE/SP 14th Workshop on Statistical Signal Processing · 2007
In this paper, a new dimension reduction algorithm called Large Margin Dimension Reduction (LMDR) is proposed for dimension reduction in classification. The formulation of LMDR incorporates the advantages of the L1-norm SVM [1] and distance metric learning [2] into one framework by using the idea of distance metric learning to search for an optimal linear transform on the original features and using the idea of L1-norm SVM to determine significant feature components. Experiments show that the proposed LMDR achieves better performance than the traditional linear discriminant analysis in certain cases.