Rare-Label-Oriented Discriminative Driven Feature Construction for Double Incomplete Multi-View Multi-Label Classification

ICLR 2026 Conference Submission16676 Authors

Keywords: multi-view multi-label datasets；double incomplete multi-view multi-label classification；Deep Learning；Contrastive Learning

TL;DR: A Rare-Label-Oriented Incomplete Multi-View Multi-Label Classification Model

Abstract: The double incomplete multi-view multi-label classification(DiMvMLC) task has attracted much attention due to the prevalence of missing views and sparse labels in real-world scenarios. However, existing methods over-rely on multi-view consensus information modeling, which results in view specificity being masked and weakens the ability to recognize rare labels. To this end, this paper proposes a model based on rare-label-oriented discriminative driven feature construction method. Through a view-specific label learning strategy, shared features and private features are decoupled to enable collaborative classification modeling of multi-view characteristics guided by commonalities. Specifically, a dual feature extraction encoder is designed to extract shared and private semantic information, respectively, and hierarchical contrastive learning loss function is introduced to enhance features separability: on the one hand, the embedding distance of the two types of features is expanded by cross-view negative sample comparison, and on the other hand, the semantic consistency of similar samples is constrained by using supervised labels. A multi-view shared feature discrimination mechanism is further proposed to strengthen the aggregation of consistent information, and the labels prediction is optimized by a rare-label-oriented decision level fusion strategy. Compared with other state-of-the-art methods, our method shows competitive experimental results on five widely used multi-view multi-label datasets.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 16676