Any-Way Independent Component Analysis with Reference

Kuaikuai Duan, Rogers F. Silva, Jingyu Liu, Oktay Agcaoglu, Vince D. Calhoun

Abstract: Multimodal fusion with reference allows a prior-guided exploration of coherent multimodal patterns to facilitate the interpretation. Multiset canonical correlation analysis with reference + joint independent component analysis (mCCAR + jICA) is one such method, and it is limited to decompose the same number of components for each modality, and two-step processing (i.e., mCCAR for correlation optimization, jICA for independence optimization) may hinder the optimization arriving at an optimal global solution. To overcome these limitations, we propose aNy-way ICA with reference, which simultaneously maximizes the independence of sources, cross-modality correlations, and correlations between reference and modalities without orthogonality requirements on the loadings, and allows different number of sources for different modalities. Simulation results demonstrated that aNy-way ICA with reference recovered the designed linkages between reference and modalities, and the designed cross-modality correlations, as well as recovered loading and component matrices with improved accuracy, compared to mCCAR+jICA under noisy conditions.