Probabilistic uncertainty-aware representation network for partial multi-view incomplete multi-label classification

doi:10.1016/j.neucom.2026.133292

Probabilistic uncertainty-aware representation network for partial multi-view incomplete multi-label classification

2026-03-07

0

PRE

AI

DOI：10.1016/j.neucom.2026.133292

Original

Original request for help

Save

Abstract

En 中文

Multi-view multi-label classification (MvMLC) has attracted widespread attention because it can assign multiple labels to samples using multiple perspectives. However, in real-world scenarios, it often faces issues of missing views and labels caused by insufficient data collection and unreliable annotations. In this situation, deterministic models struggle to support a reliable assessment of prediction confidence, while existing directed probabilistic models, although capable of sample-level uncertainty estimation, find it difficult to distinguish the contributions of different views and therefore lack the ability to perform view-level uncertainty attribution. To address these challenges, we propose the Probabilistic Uncertainty-aware Representation Network (PURN). This network employs a Variational Autoencoder (VAE) to explicitly model uncertainty at the view level through probabilistic representations. To better handle view-level uncertainty in multi-view fusion, we introduce the Confidence-Adjusted Product of Experts (CA-PoE) module, which performs confidence-aware fusion based on the product of experts and employs the Primary Preservation Mechanism (PPM) and Secondary Enhancement Mechanism (SEM) to regulate the contribution of each view. Furthermore, unlike conventional contrastive learning methods that rely on random or uniform negative sampling, we introduce the Hard Negative-aware Contrastive Learning (HNACL) module, which uses the PPM/SEM mechanisms along with top-k hard negative sample selection to direct the training signal toward the most easily confusable negative samples, thereby enhancing the discriminability between probabilistic representations at the view level. Experimental results on five benchmark datasets demonstrate that PURN achieves superior performance to a broad range of competitive methods on incomplete data.

Keywords:

Probabilistic Uncertainty

Multi-view Multi-label Classification

View-level Uncertainty Attribution

Confidence-Adjusted Product of Experts

Hard Negative-aware Contrastive Learning

Probabilistic uncertainty-aware representation network for partial multi-view incomplete multi-label classification

Abstract

Journal

Organization