Neighbor Embedding by Soft Kendall Correlation (English)

Correlation-based embedding of complex data relationships in a Euclidean space is studied. The proposed soft formulation of Kendall correlation allows for gradient-based optimization of scatter point neighborhood relationships for reconstructing original data neighbors. The approach is able to handle asymmetric data relations provided in the form of a general scoring matrix. Scale and shift invariance properties of correlation help circumventing typical embedding distortion artefacts in dimension reduction and data embedding scenarios.