Improvement of non-linear mapping computation for dimensionality reduction in data visualization and classification

The projection of high-dimensional data by linear or nonlinear techniques is a well established technique in pattern recognition and other scientific and industrial application fields. Commonly, methods affiliated to multidimensional-scaling, projection pursuit or Sammons nonlinear distance preserving mapping are applied, based on gradient descent techniques. These suffer from well-known dependence on initial or starting value and their limited ability to reach only local minimum. In this paper, stochastic search techniques are applied to the NLM to achieve lower residual stress or error value in competitive time. Encouraging results have been obtained for a particular developed local algorithm both with regard to convergence time and residual error.