缺少数据
计算机科学
人工智能
光学(聚焦)
特征(语言学)
机器学习
训练集
特征学习
面子(社会学概念)
趋同(经济学)
半监督学习
图形模型
模式识别(心理学)
深度学习
结构化预测
培训(气象学)
数据挖掘
数据点
基于实例的学习
数据建模
特征向量
监督学习
在线学习
特征选择
作者
Jun-Tong Wang,Cun-Yuan Xing,Wei Gao
出处
期刊:Proceedings
[Institute of Electrical and Electronics Engineers]
日期:2025-11-12
卷期号:: 1593-1602
标识
DOI:10.1109/icdm65498.2025.00172
摘要
Multi-label learning concentrates on the scenario where an instance is associated with multiple labels. Most previous studies focus on the design of effective algorithms over training data with full information. In some real applications, we may face with the situation where features and labels are missing simultaneously. In this work, we develop the one-pass multi-label learning with missing features and labels, which requires going through the training data only once without keeping training data in memory. Our basic idea is to recover missing features from feature correlations via a graphical model, and recover missing labels by exploiting label correlations implicitly and explicitly from learning models and multiple labels, respectively. We take the follow-the-regularized-leader algorithm to get sparse and stable solution. Theoretically, we establish the relationship among optimal classifiers, recovered features and labels, and present the convergence analysis for our method. Finally, we conduct extensive experiments to validate the effectiveness of our method.
科研通智能强力驱动
Strongly Powered by AbleSci AI