A recursive dimension reduction algorithm based on marginal discriminant analysis is presented. A constraint
condition that the obtained margin discriminant vectors are orthogonal to the new normal vector of classification hyperplane
is added to the optimal problem of support vector machines (SVM), and margin discriminant vectors can be recursively
achieved by solving the modified SVM. The number of dimension of data can be reduced by projecting data in orthogonal
margin discriminant basis. The algorithm can not only overcome some drawbacks of the most existing algorithms, such as
unable to work effectively in small size sample case, easily affected by distribution of data etc, but also has better classification
performance. The simulation results show the effectiveness of the proposed algorithm.