Deep Double Incomplete Multi-view Multi-label Classification via Graph-Constraint Learning

The utilization of multi-view multi-label data holds extensive practical implications across various domains. Recent advancements in multi-view multi-label learning have consistently faced two formidable challenges: incomplete views and missing labels. Such obstacles hinder the capability of the model to adequately harness the rich semantic diversity inherent in each sample. Despite considerable scholarly efforts, numerous approaches inadequately capture the semantic information of each data. To counter these challenges, our research introduces the Graph-Constraint Learning (GCL) model, carefully designed to address incomplete views and missing labels. Given that prevailing approaches, including stacked autoencoders, predominantly concentrate on shallow-level feature extraction, potentially undermining the coherent representation of semantic structures, we have formulated the semantic structure preservation framework based on specific views. The framework is tailored to maintain the consistency of semantic structure throughout the feature extraction phase for each view. To address the challenge of missing labels, we introduce the weighted missing labels classification module, which leverages a label missing indicator matrix as the weighting mechanism, selectively disregarding the uncertain information to efficiently mitigate the negative impact. Assessment across five real-world datasets corroborates the efficacy of the method.