From: CVPR.2006.90
核心创新:simultaneously co-propagates multiple labels,明确的探索了high-order correlation between labels,在这篇文章提出的框架下有一个优势——考虑了the label correlation of any order
Abstract
标准的标记传播算法是将训练数据上的一个标记传播到测试数据上,独立的看待每个标记,CLP与之不同,CLP同时传播多个标记(CLP simultaneously co-propagates multiple labels. )。
submodular functions 是求解目标函数的方法
1 Introduction
Multi-label learning 和 multi-class learning的区分:前者类label间是correlated,后者label间是互斥的。Since many words are semantically related, class labels are correlated. Label间的correlation有助于帮助预测测试数据的类别标记。例子:sky&ocean is both related to blue, but if an image has green and blue and be annotated with ‘grass’, then it is more likely to be annotated as ’sky’ rather than ‘ocean', 这个例子进一步形式化就是下图
CLP simultaneously co-propagates multiple labels
Eg:按频次统计的,X0会被标为a,然后是b和f,因为它们的频次是一样的
但是多标记传播后,会rank b before f,因为b-a出现的频次比f-a高
2 Related work
Key wor |