面向鲁棒分类的半监督学习新算法及应用研究

Most real data X contain redundant information, unfavorable features, noise and outliers, thus virtually all the existing label propagation based classification performing learning in the original feature space X has the following drawbacks. First, the similarity measure and weight assignment of almost all existing label propagation methods are defined based on the original space X, but the included unfavorable features and noise in X will make the similarity measure inaccurate, and hence the subsequent label prediction capability and classification accuracy will be significantly affected. Second, in virtually all the frameworks of adaptive neighborhood based label propagation schemes, where the adaptive neighborhood of points is estimated using sparse coding, the original data X is usually defined as a dictionary for optimizations, so the learning and classification performances will be vulnerable to the density and type of noises in the dictionary.. To address the above shortcomings, we in this project will propose enhanced label propagation for semi-supervised learning and robust classification. The major contributions are three fold. First, semi-supervised feature learning is firstly used to remove unfavorable features and noises existed in the original data. The similarity measure and weight assignment are then performed in the optimal feature representation space. Finally, a new model called optimal representation space based label propagation is proposed for semi-supervised and robust classification. Second, we apply the popular low-rank recovery criterion to construct a descriptive and clean dictionary for sparse coding.The sparse codes for then used for weight assignment in label propagation. Finally, a new framework termed adaptive neighborhood based enhanced label propagation technique is also presented. Based on the outputted soft labels by label propagation, we further propose an enriched pairwise constraints setting based new semi-supervised subspace learning framework, where the pairwise constraints are formulated accoring to the outputted soft labels for obtaining more supervised infromation to enhance classification performance. Third, to improve the weight assignment and label prediction capability of the label propagation process, we introcduce an enhanced cluster assumption based semi-supervised reconstruction weight proposing to reconstruct a given data using the points of the same subject as much as possible for enhanced similarity measure. In addition, for the proposed semi-supervised subspace learning framework, a mixed soft-similarity measure metric based on sparse codes and the produced soft labels through label propagation is also presented with detailed theoretical analysis. In this project, theoretical analysis combined with the experimental results will be used to evaluate the the effectiveness of the proposed label propagation and semi-supervised subspace learning algorithms for robust pattern classification.

针对原始数据X中通常包含冗余信息、不利特征、噪音和异类点的事实，现有的基于标签传播用于分类的研究有如下不足:（1）标签传播框架中的相似性度量和权值分配被直接定义在原始空间X，因此相似性度量不准确，且标签预测能力和分类精度极易受到影响;（2）在基于稀疏编码的自适应紧邻式标签传播框架中，原始数据X常被作为字典进行优化，因此分类性能极易受到字典中的噪音浓度和类型影响。. 针对以上不足，本项目探讨增强型标签传播用于鲁棒分类的研究。贡献有:（1）利用半监督特征学习去除不利特征和噪音，提出基于最优空间的增强版标签传播新算法;（2）利用低秩恢复过程构造描述性字典用于稀疏编码和权值分配，提出基于自适应近邻的标签传播新框架，进而提出新型子空间学习算法;（3）为了有效增强相似性度量、加权和标签预测性能，提出了半监督型重构权和基于稀疏编码与软类别标签的混合型权分配方案。采用理论分析结合实验验证算法的有效。

在数据爆炸和大数据时代，大多数日常生活中产生的真实数据具有维度高，分布复杂，缺乏鉴别信息（如类信息）等原因难于分类和处理。 由于上述事实，可同时有标记和未标记数据进行数据表示和分类的半监督学习方法引起了越来越多的关注。本项目针对原始数据通常包含冗余信息、不利特征、噪音和异类点的事实，探索标签预测能力更强和分类精度对噪音等因素更加鲁棒的半监督学习与分类方法，具体研究内容包括：（1）围绕直推式标签传播模型，针对现有相关半监督分类方法因预先计算权重系数无法确保表示与分类误差同时最小化（通常在标签传播前引入额外的过程计算到权重系数）、Frobenius-norm约束的流形平滑重构误差项对于噪音等过于敏感、数据噪音等因素造成的预测结果（软类别标签）包含混合信号进而影响预测精度等不足，提出一系列鲁棒性更强、预测能力更强的直推式半监督分类方法；（2）围绕归纳式标签传播模型，针对相关半监督分类方法无法确保权重构造过程对于噪音的鲁棒性和权重系数对于分类问题联合最优、权重构造因未充分考虑数据类内相似性与类间不相似性而降低图像表示与分类性能、无法有效降低“浅层类别软标签”中混合信号对结果的影响、无法快速直接从图像二维矩阵数据获取标签等问题，提出一系列鲁棒性更强、预测能力更强的直推式归纳式分类方法。采用理论分析结合实验验了证算法的有效，实验结果表明本项目提出的新方法取得了更好的分类能力和标签预测能力，可潜在地应用于图像识别等任务。. 基于本项目内容，本项目已取得的科研成果如下：（1）已发表SCI期刊论文22篇（其中包括SCI一区或中国计算机学会CCF推荐A类顶级期刊论文8篇，SCI二区或CCF推荐B类权威期刊论文10篇）、EI检索论文18篇。共计发表学术论文41篇；（2）已申请或授权发明专利40余项、已申请或授权发明专利10余项；（3）独立指导硕士研究生7名，导师制本科生近10名；（4）相关研究获2014年江苏省科学技术二等奖、2015年江苏省计算机学会青年科技奖和2016年ACM南京分会（江苏地区）新星奖；（5）项目负责人获得2017年江苏省“六大人才高峰”高层次人才、2017年江苏省"青蓝工程"优秀青年骨干教师培养对象等人才项目支持。