[阅读笔记] Teaching-to-Learn and Learning-to-Teach for Multi-Label Propagation
Chen G, Tao D, Liu W, et al. Teaching-to-Learn and Learning-to-Teach for Multi-Label propagation[C]// Thirtieth AAAI Conference on Artificial Intelligence. AAAI Press, 2016:1610-1616.
Abstract
- 本篇文章中的标记是从 labeled examples propagate to unlabeled data based on a weighted graph
- 已有的多标记传播方法忽略了不同未标记示例间的传播的不同点导致the error-prone classification of some difficult examples with uncertain labels
- 为了解决上述问题提出了ML-TLLT算法,整个传播过程Teacher引导,从简单的examples到困难的examples
- 简单/困难的判断准则:
– 确定性: the definitiveness of each possible label of the unlabeled examples
– 依赖性: the dependencies between labels revealed by the labeled examples - Teacher根据反馈决定下一次传播中的simplest examples
Introduction
- MLL algorithm
– problem transformation methods
– algorithm adaptation methods
all focus on how to exploit the label correlations to optimize learning performance - graph is a simple yet powerful tool to model the relationship between labels or examples
– existing graph-based propagation的表现不佳,有三点原因:一是样本空间的噪声,如outliers;二是label search space过大;最重要的一点原因是 they propagate the labels to unlabeled examples in an unfavorable sequence without considering their individual propagation difficulty or relia- bility. - ML-TLLT
–WHY? This is beneficial to improving propagation quality because the previously attained simple knowledge eases the learning burden for the subsequent difficult examples. - Setting
– 基本设定与主流一致,graph是基于kNN构建的 - Algorithm简介
– Teach-to-learn
为q个label分配q个teacher,每个teacher为每个未标记数据计算每个标记的 label definitiveness M r ( x i ) M_r(\mathbf{x}_i)