Dimensionality Reduction with Evolutionary Shephard-Kruskal Embeddings

Oliver Kramer

Abstract: This paper introduces an evolutionary iterative approximation of Shephard-Kruskal based dimensionality reduction with linear runtime. The method, which we call evolutionary Shephard-Kruskal embedding (EvoSK), iteratively constructs a low-dimensional representation with Gaussian sampling in the environment of the latent positions of the closest embedded patterns. The approach explicitly optimizes the distance preservation in low-dimensional space, similar to the objective solved by multi-dimensional scaling. Experiments on a small benchmark data set show that EvoSK can perform better than its famous counterparts multi-dimensional scaling and isometric mapping and outperforms stochastic neighbor embeddings.