A Probabilistic Graph Coupling View of Dimension Reduction
Title : A Probabilistic Graph Coupling View of Dimension Reduction
Asbtract : Dimension reduction is a long-standing problem for which many algorithms have been proposed. Most popular approaches include spectral (PCA-like algorithms) and pairwise similarity coupling methods (tSNE-like). Deciphering which approach is best suited to a particular case is tedious as these cannot be easily compared. In this talk, we will show that they can be unified as instances of a latent graph coupling model. These graphs induce a Markov random field dependency structure among the observations in both input and latent spaces. Interestingly, what distinguish each method are the priors considered for the latent structuring graphs. Then we will show that methods relying on shift-invariant kernels (e.g. tSNE) suffer from a statistical deficiency that explains poor performances in preserving large scale dependencies and focus on mitigating this effect with a new initialization of the embeddings.
More information : en thèse avec Aurélien Garivier (UMPA) et Franck Picard (LBMC)