Regularized Maximin Correlation Approach (RMCA) can provide a significantly improved solution method that overcomes these limitations of the original MCA. Geometric interpretation of regularization is described as follows: (a) MCA finds a vecto whose direction minimizes the worst (i.e., maximum) angle between the vector and the class members. (b) Adding outliers (the shaded region) causes an abrupt swing in the traditional maximin that MCA returns. In contrast, the r-maximin that R-MCA finds is more robust to outliers. The objects on the dotted line from the origin have the minimum correlation with the template vector $u$. (c) The character A represented in 10 different fonts. (d) The three aggregate templates of (c).
Taehoon Lee, Taesup Moon, Seung Jean Kim, and Sungroh Yoon, "Regularization and Kernelization of the Maximin Correlation Approach," IEEE Access, vol. 4, pp. 1385-1392, April 2016. [link]
If you have any questions, bug reports, or suggestions regarding RMCA, please do not hesitate to contact us.