计算机科学
人工智能
算法
模式识别(心理学)
数据挖掘
算法设计
特征提取
特征(语言学)
统计分类
机器学习
作者
Xinyu Xiao,Shuhan Qi,Xuan Wang,Jie Wen,Wenjian Luo,Bin Chen
标识
DOI:10.1109/tmm.2026.3753178
摘要
Multi-view multi-label classification (MVMC) leverages multi-view information from samples to accurately predict multiple labels. In real-world scenarios, Incomplete Multi-view Multi-label Classification (IMvMLC) faces severe challenges due to missing views and incomplete labels. Existing methods, especially those based on instance-level contrastive learning, often suffer from the false negative problem, where potentially similar samples are erroneously repelled, disrupting the intrinsic local manifold structure. To overcome this limitation, we propose an Adaptive Weighted Granular-ball Computing framework (AWGBC). AWGBC first constructs a unified latent space using a deep autoencoder and encapsulates the local topological structure using granules as basic units. To address the issue that incomplete data often leads to the retention of non-uniform viewpoint-specific noise in latent features, we introduce a weighted granularity calculation module. By iteratively optimizing feature weights based on local data distribution, we perform adaptive soft feature selection, thereby suppressing residual noise and refining granular boundaries. Subsequently, AWGBC established a reliable two-level association by quantifying the precise number of common samples between granularities, thus avoiding spurious correlations common in coarse-grained settings. Finally, granular-ball contrastive learning is performed to pull associated granules closer and push unassociated ones apart, learning highly discriminative feature representations while preserving local structural information. Extensive experiments on multiple benchmark datasets demonstrate that AWGBC outperforms state-of-the-art methods on IMvMLC tasks. Ablation studies further validate the effectiveness of each module.
科研通智能强力驱动
Strongly Powered by AbleSci AI