Incomplete Multi-View Multi-Label Learning via Label-Guided Masked View- and Category-Aware Transformers

As we all know, multi-view data is more expressive than single-view and multi-label annotation enjoys richer supervision information single-label, which makes learning widely applicable for various pattern recognition tasks. In this complex representation problem, three main challenges can be characterized as follows: i) How to learn consistent representations of samples across views? ii) exploit utilize category correlations guide inference? iii) avoid the negative impact resulting from incompleteness views or labels? To cope with these problems, propose a general framework named label-guided masked view- category-aware transformers in paper. First, design two transformer-style based modules cross-view features aggregation classification, respectively. The former aggregates different process extracting view-specific features, latter learns subcategory embedding improve classification performance. Second, considering imbalance power among views, an adaptively weighted view fusion module proposed obtain view-consistent features. Third, impose label manifold constraint sample-level maximize utilization supervised information. Last but not least, are designed under premise incomplete labels, our method adaptable arbitrary data. Extensive experiments on five datasets confirm that has clear advantages over other state-of-the-art methods.