基于树结构的层次性多示例多标记学习

A Hierarchical Multi-Instance Multi-Label Learning for Tree Structure Among Labels

针对多示例多标记学习中标记间树结构的问题,将多示例学习、多标记学习和树结构标记优化方法有机融合,提出了基于树结构标记的层次性多示例多标记学习方法TreeMIML.TreeMIML先将样本中的多个示例转化为单示例,然后通过多标记学习得到新样本的标记,最后通过树结构标记优化方法学习样本的最终标记.实验结果证明,TreeMIML方法在G蛋白偶联受体的生物学功能预测上获得了很好的分类性能,优于目前最好的多示例多标记学习和多标记学习方法.

This paper proposed a novel hierarchical multi-instance multi-label learning algorithm named TreeMIML to solve the challenge of tree structure among labels in multi-instance multi-label learning(MIML),by integrating multi-instance learning,multi-label learning and tree-structure optimization scheme.TreeMIML first converts multiple instances in each sample into single instance,then obtains sample outputs by multi-label learning,and finally optimizes the outputs to obtain the labels of unseen samples by a tree-structure optimization method.The experimental results show that our TreeMIML algorithm achieves good classification performance in predicting biological functions of G protein-coupled receptors,which is superior to state-of-the-art multi-instance multi-label learning and multi-label learning methods.