A new Sammon algorithm for sparse data visualization

Abstract

Sammon's mapping is an important non-linear projection technique that has been widely applied to the visualization of high dimensional data. However when dealing with sparse data, the object relations induced by the map become often meaningless. In this paper, we present a new Sammon algorithm (SSammon) that overcomes this problem by previously transforming the dissimilarity matrix in an appropriate manner. The connection between our algorithm and a kernelized version of Sammon's mapping is also studied. The new model has been applied to the high dimensional and sparse problem of word relation visualization. We report that SSammon outperforms two widely used alternatives proposed in the literature.