New semi-supervised model handles multi-view data with incomplete views and missing labels

Real-world datasets often come with a practical challenge: different sources or "views" of the same object may be missing, and labeled examples are scarce. A new research paper from arXiv addresses this problem by developing a semi-supervised generative model that makes better use of available data.

The Problem with Multi-View Learning#

Multi-view learning — where a dataset is described from multiple perspectives or modalities — is common in applied settings. However, two key obstacles usually get in the way: some views may be absent for certain samples, and labeling data is expensive, leaving most samples without annotations. Earlier work tackled the missing-view problem using a product-of-experts approach combined with the information bottleneck (IB) principle, but these methods were fully supervised and couldn't take advantage of unlabeled data.

A New Semi-Supervised Framework#

The proposed method overcomes this limitation by combining labeled and unlabeled samples within a single framework. The model maximizes the likelihood of unlabeled data to discover a shared latent space that aligns with the information bottleneck learned from labeled examples.

The approach also incorporates two refinements:

• Modality-specific information is preserved in the likelihood model, ensuring view-specific details aren't lost
• Cross-view mutual information maximization is applied in the shared space to better extract common signals across different data sources

Performance Gains#

On complex datasets with both missing views and limited labeled examples, the model demonstrates improvements in both predictive accuracy and generation quality compared to existing approaches, according to the paper.