Visualization of Non-vectorial Data Using Twin Kernel Embedding

Visualization of non-vectorial objects is not easy in practice due to their lack of convenient vectorial representation. Representative approaches are kernel PCA and kernel Laplacian eigenmaps introduced recently in our research. Extending our earlier work, we propose in this paper a new algorithm called twin kernel embedding (TKE) that preserves the similarity structure of input data in the latent space. Experimental evaluation on MNIST handwritten digit database verifies that TKE outperforms related methods