Evolutionary Multidimensional Scaling for Data Visualization and Classification

摘要

Multidimensional Scaling (MDS) is a well established technique for the projection of high-dimensional data in pattern recognition, data visualization and analysis, as well as scientific and industrial applications. In particular, Sammons Nonlinear Mapping (NLM) as a common MDS instance, computes distance preserving mapping based on gradient descent, which depends on initialization and just can reach the nearest local optimum. Improvement of mapping quality or reduction of mapping error is aspired and can be achieved by more powerful optimization techniques, e.g., stochastic search, successfully applied in our prior work. In this paper, evolutionary optimization adapted to the particular problem and the NLM has been investigated for the same aim, showing the feasibility of the approach and delivering competitive and encouraging results.