多重代价失衡的机器学习技术研究

Cost risk minimization is one of the effective classification decision-making criteria recently, many researchers in the world focus on it in machine learning and pattern recognition research fields. Both dimensionality reduction and cost-sensitive learning are the effective strategies to improve the performance of the unbalanced cost-based classifier, but currently very little work has been done to simultaneously minimize multiple costs(misclassification cost, attribute cost or test cost, wait cost and etc) for classification, while multiple costs imbalance is widespread in many real classification issues such as the classification problem of imbalanced datasets. Therefore, this project aims to seek some new feature selection strategies, mutiple costs minimization-based classifier design and mutiple costs minimization-based integrated classifier design. Concretely, the project focuses on: 1)introducing new feature selection methods based on hypothesis-margin or information entropy or Filter-Wrapper model, and then proposing novel feature selection approaches based on those newly developed multiple costs minimization criteria; 2)giving new supervised(semi-supervised)feature selection algorithms by employing local structure preserving, and seeking new strategies that can effectively embed multiple costs minimization criterion into the newly proposed feature selection algorithms; 3)designing novel supervised(semi-supervised) classification models by incorporating cost-sensitive learning strategies; 4) introducing novel supervised (semi-supervised) classification models by simultaneously incorporating cost-sensitive learning and feature selection strategies. Further, based on studies above, our research aims at 1) designing a class of multiple costs minimization-based classifiers; 2) integrating the multiple costs minimization-based base classifiers into a classifier.

代价风险最小化是目前有效的分类决策判别准则之一，倍受国内外机器学习和模式识别研究者的关注。降维和代价敏感学习是改进代价失衡分类器性能的有效策略，但当前同时进行多重代价（错分代价、属性代价等）最小化的分类器研究还不多见，而多重代价失衡在数据不平衡等分类问题中普遍存在。为此，本项目旨在寻求特征选择的新策略、多重代价最小化的分类模型设计及多重代价最小化的集成分类器设计三个方面展开研究。侧重研究：1）提出基于假设间隔、信息熵、Filter-Wrapper模型的特征选择新方法，构建出多重代价最小化新准则下的特征选择新算法；2）提出局部结构保持的监督（半监督）特征选择新算法,探寻并入多重代价最小化的新策略；3）设计嵌入代价敏感学习策略的监督（半监督）分类模型；4）设计并入代价敏感学习和特征选择的分类器模型。以上述研究为基础，进而研究1）一类多重代价最小化的分类器设计；2）多重代价最小化的分类器的集成。

代价敏感学习是解决代价失衡问题的重要机器学习方法，具有理论和应用研究价值。项目组围绕代价敏感的降维和分类器设计、结合标记相关性的多标记学习、基于假设间隔的特征选择、字典学习和图像去噪展开了深入的研究。具体地说，项目组提出了一类嵌入代价敏感的降维和分类学习算法，论文发表在IEEE TIFS（CCF A)、Neurocomputing(CCF C)、NEURAL PROCESS LETT(CCF C)、软件学报等国内外学术期刊上。针对多标记数据的不平衡性，采用多任务学习框架，利用标记相关性，提出了一类面向多标记数据的学习算法，论文发表在软件学报、电子学报等国内学术期刊上。同时，项目组围绕特征表示和图像去噪进行了相关研究，论文发表在IEEE TCSVT(CCF B)、Pattern Recognition(CCF B)和Image Vision and Computing(CCF C)等国内外学术期刊上。此外，项目组在聚类分析、特征选择和SVM分类算法等方面进行了研究，论文发表在Knowledge-Based Systems(CCF C)和《控制与决策》等国内外学术期刊上。. 目前，有关嵌入代价敏感的降维算法研究相对较少，项目组的研究成果是该方面研究的有效补充，受到同行的关注。利用标记相关性，设计了多标记分类和特征选择联合学习框架，该框架也是解决数据不平衡引起的代价失衡问题的有效途径。同时，项目组进行了字典学习和图像去噪等相关研究，为未来进行面向多标记数据分类的判别型字典学习研究打下了必要的基础。