MDS algorithms are data analysis techniques that have been successfully applied to generate a visual representation of multivariate object relationships considering only a similarity matrix. However in high dimensional spaces the concept of proximity become meaningless due to the data sparsity and the maps generated by common MDS algorithms fail often to reflect the object proximities.
In this paper, we present a new MDS algorithm that overcomes this problem transforming the dissimilarity matrix in an appropriate manner. Besides a new dissimilarity is proposed that reflects better the local structure of the data. The connection between our model and a kernelized version of the Kruskal MDS algorithm is also studied.
The new algorithm has been applied to the challenging problem of word relation visualization. Our model outperforms several alternatives proposed in the literature.