多示例学习的两阶段实例选择和自适应包映射算法

Two-stage instance selection and adaptive bag mapping algorithm for multi-instance learning

多示例学习(Multi-Instance Learning,MIL)研究对象的内部结构比单示例学习更加复杂.已有的MIL方法大都基于原始空间中的实例进行包映射,但这些方法通常忽略包的内部结构信息,难以保证所选实例与包在新特征空间中的关联性.提出一种多示例学习的两阶段实例选择和自适应包映射(TAMI)算法.首先,实例选择技术根据包中实例的密度值和关联性,挖掘包内结构特征,选取实例原型;其次,实例选择技术选取具有峰值密度的实例原型作为代表实例;最后,自适应包映射技术通过定义新的映射函数将包转换为单向量进行学习.实验利用显著性检验从统计学的角度验证了TAMI在图像检索、文本分类等基本数据集上的有效性.结果表明,TAMI在图像检索和医学图像数据集上取得了比其他MIL算法更好的效果,并在文本分类数据集上表现良好.

Compared with single-instance learning,multi-instance learning(MIL)has a more complex internal structure of its research objects.Most of the existing MIL methods map bags based on instances in the original space.They hardly consider the internal structure information of the bags.It is difficult to guarantee the affinity between the selected instance and the bag in the new feature space.In this paper,we propose a two-stage instance selection and adaptive bag mapping algorithm for multi-instance learning(TAMI)to handle this issue.Firstly,the first-stage instance selection technique excavates structural features and selects instance prototypes based on the density and affinity of the instances in the bag.Secondly,the second-stage instance selection technique chooses instance prototypes with the peak density as representatives.Finally,the new adaptive bag mapping technique converts each bag into a single vector.Experiment verifiy the effectiveness of TAMI on the basic dataset from a statistical point of view.The results show that TAMI has achieved better results than other MIL algorithms on image retrieval and medical image datasets,and it performs well on text classification datasets.