Divergence-guided disentanglement of view-common and view-unique representations for multi-view data

In the field of multi-view learning (MVL), it is crucial to extract both common (consistent) and unique (complementary) information across different views. While the focus has traditionally been on acquiring common information, there has been a recent shift towards exploring unique information as well. However, developing an MVL model that can simultaneously capture both common and unique information, thereby facilitating a comprehensive understanding of multi-view data, remains a significant challenge. To address this, we propose the Divergence-guided Multi-view Learning framework (DG-MVL), inspired by information-theoretic learning theory, specifically the generalized divergence measure. This framework employs multi-view autoencoders to disentangle the features obtained from each view into coarse common and unique components. By minimizing the divergence between the coarse common features learned from the common encoder of each view and maximizing the divergence between the coarse unique features from unique encoders simultaneously, we optimize the extraction of both common and unique information. Subsequently, we merge these features to generate a comprehensive and concise representation of the multi-view data, which can be easily utilized for various downstream tasks. We validate our framework on a synthetic multi-view dataset, demonstrating its effectiveness in disentangling common and unique information. Further experiments on various real-world datasets confirm the effectiveness of DG-MVL in capturing common and unique information from multi-view data, resulting in superior classification performance compared to existing methods. Code is available at https://github.com/LMFLRB/DG-MVL.git.