| 摘要: |
| 近些年来, 作为一种新的有监督学习范式, 标记分布学习(LDL)已被应用到多个领域, 如人脸年龄估计、头部姿态估计、电影评分预测、公共视频监控中的人群计数等, 并且在这些领域的相关任务上取得了一定性能上的进展. 最近几年, 很多关于标记分布学习的算法在解决标记分布学习问题时考虑到了标记之间的相关性, 但是现有方法大多将标记相关性作为先验知识, 这可能无法正确刻画标记之间的真实关系. 此外, 标记相关性通常用于在训练阶段调整假设空间, 而最终的标记预测并未显式利用标记间的相关性.因此, 提出一种新的标记分布学习方法——考虑标记间协作的标记分布学习(LDLCL).该方法旨在训练期望模型的同时, 显式地考虑标记间的相关预测.具体来讲, 首先提出假设: 对于每个标记, 最终的预测结果涉及到它自己的预测和其他标记的预测之间的协作. 基于这一假设, 提出一种通过标记空间中的稀疏重构来学习标记相关性的新方法; 然后, 将学习到的标记相关性无缝集成到模型训练中; 最终, 在标记预测时使用学习到的标记相关性. 大量的实验结果表明, 该方法优于近期的同类方法. |
| 关键词: 标记分布学习 标记相关性 样本相似性 标记分布 |
| DOI：10.13328/j.cnki.jos.006139 |
| 分类号:TP18 |
| 基金项目:江苏省自然科学基金(BK20191287) |
|
| Label Distribution Learning with Collaboration among Labels |
|
LI Rui-Yu, ZHU Ji-Hua, LIU Xin-Yuan
|
|
School of Software Engineering, Xi'an Jiaotong University, Xi'an 710049, China
|
| Abstract: |
| In last few years, as a new supervised learning paradigm, label distribution learning (LDL) has been applied to many fields and shown good results in these fields, such as face age estimation, head posture estimation, movie score prediction and crowd count in public video surveillance. Recently, the correlations between labels have been considered in some algorithms when solving the problem of label distribution learning. However, most of the existing methods take label correlations as a prior knowledge, which may not be able to correctly describe the real relationship between labels. In addition, label correlations are usually used to regularize the hypothesis space in the training phase, while the final label distribution prediction does not use these correlations explicitly. Therefore, this study proposes a new label distribution learning method, label distribution learning with collaboration among labels (LDLCL), which aims to explicitly consider the correlated predictions of labels while training the expected model. Specifically, the hypothesis is first proposed: for each label, the final prediction involves the cooperation between its own prediction and other labels' predictions. Based on this assumption, a new method is proposed to learn label correlations by sparse reconstruction in label space. Then, the learned label correlations are seamlessly integrated into model training, and finally the learned label correlations are used in label prediction. Sufficient experimental results show that the proposedapproach is superior to other similar methods. |
| Key words: label distribution learning label correlations instance similarity label distribution |
