Correlation-aware label-guided consistent network for partial multi-view incomplete multi-label classification

Given the uncertainties in manual labeling and data collection, advances in multi-view multi-label learning often face challenges posed by partial views and incomplete labels. Although some methods have been proposed to address these challenges, most of them do not fully leverage the inter-view and the inter-label correlations, which are crucial for improving the quality of the fused representations. To cope with this issue, we propose a partial multi-view incomplete multi-label network, named correlation-aware label-guided consistent network(CALGC-Net). Different from existing multi-view fusion methods, we design a correlation-aware sub-network that dynamically allocates fusion weights to views based on their inter-view essential correlations, improving the quality of fusion representation. Additionally, we leverage the semantic dependencies between labels to constrain the feature representations, ensuring the semantic structure in the embedding space aligns with that in the label space. Extensive evaluations on five benchmark datasets demonstrate that CALGC-Net consistently outperforms several existing methods, including CDMM, iMvWL, NAIM3L, DD-IMvMLC-net, SSP, DICNet, LMVCAT, and MTD, across multiple evaluation metrics. These results underscore the effectiveness of CALGC-Net in handling partial multi-view incomplete multi-label classification tasks. The code is available at https://github.com/yanglq2003/CALGC-Net.