Fig 3. Effect of the triplet margin loss on the stochastic encodings of the microbiome samples.

The depicted curves are the 95% confidence intervals of the samples’ stochastic encodings $\mathit{z}\sim p\left(\mathit{z}\right|\mathit{x})=\mathcal{N}(\mathit{\mu },{\mathit{\sigma }}^2\mathit{I})$; the points are their means μ. The displayed encodings consist in a set of 22 test samples obtained from a random training-test split of the IBD dataset (i.e. the 20% of the dataset not used for training). The K dimension of the latent space has been set to 2 in order to allow a 2D visualisation. (A, B) optimisation of the J_MVIB objective (Eq 5). (C, D) optimisation of the J_MVIB−T objective (Eq 8).