Fig 4. MVIB cross-study generalisation and benchmark.

Values are test ROC AUC computed by first training the models on a source dataset and then testing it on a target dataset. RF: Random Forest. All datasets used for cross-study experiments belong to the joint collection. For MVIB, the J_MVIB−T objective has been adopted for the optimisation (Eq 8). The datasets reported on the x-axis shall be interpreted as: train →test. For the Random Forest, the error bars represent the standard error over five repeated experiments and account for the stochasticity of the Scikit-learn implementation. The standard error is missing for the MVIB results, as our PyTorch implementation has been made deterministic.