距离度量学习在多示例多标记学习中的应用研究

发布时间：2018-06-21 01:25

本文选题：多示例多标记学习 + 距离度量学习　；参考：《南京邮电大学》2017年硕士论文

【摘要】：多示例多标记学习(MIMLL,Multi-Instance Multi-Label Learning)是机器学习的一个重要分支。MIML学习框架中,一个样本由多个示例即特征向量表示,并同时与多个标记相关联。因为自身的结构复杂和多语义特性,图像和文本的分类和近邻搜索问题可以归结为MIML学习问题。传统的MIML算法使用豪斯多夫距离(Hausdorff distance)来度量样本在特征空间的距离,无法反映出样本间的语义相关性。而距离度量学习通过利用样本的标记信息来优化样本间距离,达到距离度量的语义一致。本文在MIML学习架构下,通过距离度量学习优化样本的分类和搜索算法,主要贡献如下:提出了一种基于聚类策略的MIML距离度量学习算法,该算法在学习过程中引入了聚类思想与标记相关性,在简化了距离度量学习复杂度的同时探索了标记与示例之间的关系;设计了一种联合距离度量学习和k-最近邻技术的MIML分类算法,利用学习得到的距离度量计算样本近邻充分考虑了语义一致性,提高了分类算法的性能;设计了一种联合距离度量学习和锚图哈希(AGH,Anchor Graph Hashing)技术的MIML哈希算法,利用学习得到的距离度量构建样本与锚点的邻接矩阵,充分考虑语义一致性,提高了哈希算法性能;实验验证了本文提出的分类及哈希算法的有效性。
[Abstract]:Multi-example multi-label learning (MIMLL) is an important branch of machine learning. In the framework of MIML learning, a sample is represented by multiple examples, that is, eigenvector, and is associated with multiple tags at the same time. Because of its complex structure and multi-semantic features, the classification and nearest neighbor search of images and texts can be attributed to MIML learning problems. The traditional MIML algorithm uses Hausdorff distance (Hausdorff distance) to measure the distance of samples in feature space, which can not reflect the semantic correlation between samples. Distance metric learning optimizes the distance between samples by using the tag information of samples to achieve semantic consistency of distance measurement. Based on MIML learning architecture, this paper optimizes the classification and search algorithm of samples by distance metric learning. The main contributions are as follows: a distance metric learning algorithm based on clustering strategy is proposed. In the process of learning, the algorithm introduces the clustering idea and the correlation of label, which simplifies the learning complexity of distance measurement and explores the relationship between marking and examples. A MIML classification algorithm combining distance metric learning and k- nearest neighbor technique is designed. The semantic consistency is fully considered and the performance of the classification algorithm is improved. A MIML hash algorithm combining distance metric learning and Anchor Graph Hashingis designed. The distance metric is used to construct the adjacent matrix between the samples and the anchor points, which fully considers the semantic consistency and improves the performance of the hash algorithm. Experiments show the effectiveness of the proposed classification and hash algorithm.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP181

【相似文献】