Fig. 3

Performance comparison of dimensionality-reduction techniques. a Toy simulated data example illustrating the performance of ZIFA compared to standard dimensionality-reduction algorithms. b Performance on simulated data sets based on correlation score between the estimated and true latent distances as a function of λ (larger λ, lower dropout rate), number of genes and latent dimensions, and noise level used in the simulations. c Plots showing the divergence between the predictive and empirical data distributions as a function of dropout rate and mean expression level for FA, PPCA and ZIFA. Illustrative predictive performance and model fits (red, color online) on the T-cell single-cell data set (black) [3]