Margin preserving projection
Y. Lee, D. Ahn, K. Moon · Electronics Letters · 2006
Proposed is a novel feature extraction method of finding projection directions that preserve the distance between class boundaries, i.e. margin. The proposed method is a supervised version of locality preserving projection and overcomes the two main drawbacks of Fisher linear discriminant: failure to find useful projection directions when the means of classes coincide with each other and inability to find more features than the number of classes. Experimental results show the effectiveness of the proposed method.