一种基于聚类的图卷积多示例学习算法
A multi-instance learning algorithm of graph convolution based on clustering
投稿时间:2020-07-30  修订日期:2020-07-30
DOI:
中文关键词: 多示例学习  聚类  包图  图卷积
英文关键词: multi-instance learning  clustering  bag graph  graph convolution
基金项目:湖北省技术创新专项重大项目(2019ABA101);中央高校基本科研业务费专项资金项目(CZY18015)
作者单位E-mail
王江晴 中南民族大学 计算机科学学院 wjqing2000@mail.scuec.edu.cn 
毕建权 中南民族大学 计算机科学学院  
帖军 中南民族大学 计算机科学学院  
孙翀 中南民族大学 计算机科学学院  
艾勇 中南民族大学 计算机科学学院  
摘要点击次数: 160
全文下载次数: 0
中文摘要:
      基于图结构的多示例学习可用于解决挖掘包中示例间相关性问题.然而,现有的方法大多随机选择包中示例构建图结构,忽略了包中具有代表性示例对图结构的影响;同时都是间接在包图结构上建立分类器,造成了模型运行效率低下问题.针对上述问题,提出了一种基于聚类的图卷积多示例学习算法MIL-CGC,首先通过聚类方法获取每个包中的超示例,作为包图结构中的节点;然后通过挖掘超示例间关系构建包图的边,确定包图结构;最后利用图卷积对包图节点重要度分数进行学习,直接在包图结构上建立分类器.实验表明,MIL-CGC可以充分表示包图结构,有效提高分类模型的质量.
英文摘要:
      Multi-instance learning based on graph structure can be used to solve the problem of correlation between instances in mining bag.However,most of the existing methods randomly select the instances in the bag to construct the graph structure,ignoring the influence of the representative instances in the bag on the graph structure,and all of them build classifiers on the bag graph structure indirectly,resulting in the low efficiency of the model.Aiming at these problems,a multi-instance of graph convolution based on clustering is proposed (abbrev. MIL-CGC). Firstly,the super-instances in each bag are obtained by clustering method as the nodes in the bag graph structure.Then,the edge of the bag graph is constructed by mining the relationship between super-instances to determine the structure of the bag graph.Finally,graph convolution is used to learn the node importance score of the bag graph,and the classifier is directly established on the bag graph structure.Experiments show that MIL-CGC can fully represent the bag graph structure and effectively improve the quality of the classification model.
View Fulltext   查看/发表评论  下载PDF阅读器
关闭