归纳型安全半监督分类学习及其扩展研究

Semi-supervised classification learning is one of the machine learning tasks attracted much attention recently, which aims to use both the labeled and unlabeled data so as to achieve better classification performances than supervised classification learning based on the labeled data alone. Though with many achievements, there is still an essential problem in semi-supervised classification: the insecurity of semi-supervised classification, i.e., the semi-supervised classification methods may gain even worse performances than the corresponding supervised ones in some cases, which seriously limits their real applications. As a result, safe semi-supervised classification naturally becomes an extremely important learning task. But as far as we known, there are few related works, and are all transductive, i.e., aiming to obtain only the class labels for the given unlabeled data, thus unable to predict unseen data. While in many real classification tasks, one needs to predict unseen data, thus needs inductive classification methods. The purpose of the program is to directly develop an inductive safe semi-supervised classification method for, on one hand, filling the blank of current research, and on the other hand, further improving the safety and applicability of semi-supervised classification methods. The entire work is carried out with several processes including modeling, algorithm design and implementation, theoretical analysis and experimental verification. Furthermore, the idea is also extended to extremely semi-supervised classification learning for addressing more challenging applications.

半监督分类学习是当前备受关注的机器学习任务之一，旨在同时利用标号和无标号样本进行学习，期望获得比仅利用标号样本的监督分类学习更好的分类性能。尽管已取得众多进展，但其中仍存在一个本质问题：半监督分类不安全问题，即半监督分类方法在某些场合可能产生比对应监督方法更差的分类性能，这严重限制了其实际应用。因此，安全半监督分类便成为一项极其重要的研究任务。然而据我们所知，目前相关的研究工作相当有限，且均为直传型学习方法，即通过学习获得给定无标号样本的类标号，无法预测未见样本。但真实分类任务常需对未见样本进行预测，因而需要归纳型学习方法。本项目旨在直接提出一种归纳型安全半监督分类方法，一方面力图填补现有研究的空白，另一方面进一步提升半监督分类方法的安全性和可应用性。整个工作围绕建模、算法设计与实现、理论分析与实验验证等诸方面系统展开。同时尝试扩展该理念至极端半监督分类学习以解决更具挑战性的应用问题。

半监督分类学习是机器学习中的一个热点研究领域，同时利用少量有标签样本和大量无标签样本进行分类学习。然而，无标签样本的利用并不总能提升分类性能，有时甚至导致性能下降，造成不安全半监督分类。本项目致力于安全半监督分类学习研究，围绕提升半监督分类的学习性能开展了一系列深入和广泛的研究。主要包括：1）在半监督分类器设计上，提出一系列安全半监督分类学习模型，以提升半监督分类的安全性。首先，从无标号样本的不安全性考虑，提出基于改进聚类假设的安全半监督分类学习方法；其次，从流形结构图的不合理性出发，分别提出基于判别型图的MR分类学习框架、图自适应的MR分类学习框架和点状MR分类学习框架；最后，直接寻求监督和半监督分类学习间的合理折衷，提出归纳型安全半监督分类方法，并提出归纳型半监督universum学习模型。2）核技术是提升分类学习性能的有效方法之一。本项目针对不定核学习，分别提出求解不定核SVM的凸差函数规划方法和基于不定核的判别型正则化分类学习框架；同时，针对多核学习问题，提出lp-范数多类多核学习方法。3）在进行分类学习前，利用特征选择或加权技术可进一步提升学习性能。本项目在特征选择方面分别提出稳定的和面向隐私保护的特征选择方法。4）此外，本项目还针对分类学习的一些相关研究提出了一些改进方法，包括同时聚类和分类的流形学习框架和基于局部相似性学习的成对约束传播等。