Figure 3: Synthetic data: ISI (lower is better).

Each row represents the ground-truth subspace structure used to generate the data and each column represents the test subspace structure used to fit the model. If a workflow could correctly identify all ground-truth subspace structures, the lowest ISI values would align along the main diagonal. The unimodal initialization workflow (PCA+ICA) and the MSIVA initialization workflow (MGPCA+ICA) led to the lowest ISI values (≤ 0.02) along the main diagonal, indicating that these two approaches successfully identified the correct ground-truth subspace structures from the incorrect ones. However, the multimodal initialization workflow (MGPCA+GICA) failed to detect the subspace structure $S_4$ with a high ISI value (0.065) in the main diagonal. Thus, MSIVA and the unimodal baseline are considered better than the multimodal baseline.