基于类标签关联度与缓变原理的有序判别回归研究及应用

发布时间：2018-04-28 20:13

本文选题：有序回归 + 核方法　；参考：《南京航空航天大学》2016年硕士论文

【摘要】：有序回归(Ordinal Regression,OR)是模式识别中一种特殊的有监督学习范例,其目的是基于一组给定的输入和离散有序的输出数据建立一个回归器,使其能实现对相应有序离散测试样本类标的预测。在现实应用中,此种类标离散而有序的场景广泛存在,这些类标值通常是基于人们的偏好而被赋予的。此类标号不仅体现了有序标号间的等级差异,同时也反映了具有不同类标样本之间的联系。因而不同于单纯的回归与分类学习,OR兼具分类和回归的双重特性。相比于普通回归方法,有序回归模型融合利用了类标号的有序性这一先验信息,使得其具有比普通回归更优的回归性能。近年来,因OR在人脸识别、信用评级、脑机接口以及年龄估计等领域的广泛应用,其得到了研究者们越来越多的关注和研究。然而,现有OR方法在问题先验信息的利用方面仍存在某些遗漏的不足。为此,本文针对性地在以下几方面展开了研究:鉴于现有OR方法对有序标号间临近关联信息利用不足的问题,我们设计出了一类有序标号间关联度的量化表示,然后通过将该关联度表示与经典且有效的有序回归方法KDLOR(Kernel Discriminant Learning for Ordinal Regression)相结合为有序学习建模,并将构建出的方法称为结合类标签关联度的有序线性判别回归学习(Linear Discriminant Learning for Ordinal Regression using Label Membership,LM-LDLOR),为应对非线性有序学习问题,我们还对LM-LDLOR进行核化从而得到结合类标签关联度的有序核判别回归学习,简记为LM-KDLOR。最后,在8个标准有序回归数据集上的对比实验验证了本文所提策略对有序回归性能提升的有效性。缓变原理是基于人类的视觉特性而构建的一种学习原理,并且在模式识别中已有了较广泛的应用。然而,据我们所知缓变学习原理尚未与有序回归学习相结合进行研究。受此启发,本文通过最近邻法对每个样本类构建多个类内输出有序序列计算缓变类内散度矩阵,同时通过有序约束保证类标号的有序性,然后根据线性判别准则进行映射实现有序学习,最终形成了基于缓变原理的判别有序回归方法SP-DLOR(Slowness Principle based Discriminant Learning for Ordinal Regression)。最后,在8个标准有序回归数据集以及人脸数据集FG-NET上的对比实验验证了所提算法在回归和分类性能上的优越性。
[Abstract]:Ordinal regression order is a special supervised learning paradigm in pattern recognition, which aims to build a regression based on a given set of input and discrete ordered output data. It can realize the prediction of the corresponding ordered discrete test samples. In practical applications, this kind of label is widely used in discrete and orderly situations, and these class values are usually given based on people's preferences. This kind of label not only reflects the grade difference among the ordered labels, but also reflects the relationship between the samples with different classes. Therefore, different from the simple regression and classification learning OR has the dual characteristics of classification and regression. Compared with ordinary regression methods, the integration of ordered regression model makes use of the prior information of class labeling, which makes it have better regression performance than ordinary regression. In recent years, due to the wide application of OR in face recognition, credit rating, brain-computer interface and age estimation, it has attracted more and more attention and research. However, there are still some omissions in the use of prior information in the existing OR methods. For this reason, this paper focuses on the following aspects: in view of the insufficient use of the adjacent correlation information between ordered labels by existing OR methods, we design a kind of quantitative representation of the correlation degree between ordered labels. Then, the relation degree representation is combined with the classical and effective ordered regression method KDLOR(Kernel Discriminant Learning for Ordinal Regression to model ordered learning. The method is called Linear Discriminant Learning for Ordinal Regression using Label regression Learning (LM-LDLORA) combined with class label correlation degree. In order to deal with the problem of nonlinear ordered learning, We also nucleate LM-LDLOR and obtain an ordered kernel discriminant regression learning combined with class label correlation, which is abbreviated as LM-KDLOR. Finally, a comparative experiment on eight standard ordered regression datasets verifies the effectiveness of the proposed strategy in improving the performance of ordered regression. The principle of slow change is a learning principle based on human visual characteristics, and has been widely used in pattern recognition. However, as far as we know, the principle of slow learning has not been combined with ordered regression learning. Inspired by this, this paper constructs multiple in-class output order sequences for each sample class by nearest neighbor method to calculate the slowly varying intra-class dispersion matrix, and guarantees the ordering of class labels through ordered constraints. Then, according to the linear discriminant criterion, the ordered learning is realized by mapping, and finally a discriminant ordered regression method, SP-DLOR(Slowness Principle based Discriminant Learning for Ordinal Regression, is formed based on the principle of slow variation. Finally, the comparison experiments on eight standard ordered regression data sets and face dataset FG-NET show the superiority of the proposed algorithm in regression and classification performance.
【学位授予单位】：南京航空航天大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.4

【参考文献】