高维序列数据的核学习方法及应用研究

Research on kernel learning methods is one of the most important forefront issues in the field of machine learning. Targeted high-dimensional sequence data, which are typically generated by the real-world applications such as Clinical medical nursing, we propose in this project to study new models and algorithms for kernel learning on the complex structural data from a new perspective called Sequential Feature View, and to develop efficient methods for computer-aided clinical diagnosis and treatment based on non-linear analysis of the sequential features. The major studies include: (1) proposing a hyperkernel model and its learning methods with embedded feature selection for high-dimensional sequence data, which allow the supervised or unsupervised algorithms to learn adaptive kernels from the reduced data; (2) proposing a descriptive probability model, incorporating with its evaluation methods, for describing the patterns existing in the Hilbert space; (3) proposing non-linear subspace clustering algorithms and multi-class classifiers for high-dimensional sequence data; (4) the application of pattern recognition on patient conditions and disease-related risk factors, based on kernel learning of clinical time series. In this project, we will concentrate efforts on the integration of kernel density estimates and Mercer kernel methods, which results in a creative probability models-based method. Using the new method, we aim at investigating solutions for the bottleneck problems arising from the classical kernel learning theory and its applications such as data-driven learning of kernels for complex data and explicitly characterizing of the patterns in the kernel spaces, and at establishing the featured clinical decision aided system. Therefore, the output of this proposal will obtain a significant progress in the foundational research of the statistical learning theory and methods.

核学习方法研究是机器学习领域一项重要的前沿课题。本项目以临床医疗护理等应用产生的高维序列数据为对象，基于序列特征新视角，研究复杂结构数据核学习的新模型和新算法，以序列特征非线性分析为主要工具开发临床辅助诊疗的有效方法。主要研究:(1)嵌入序列特征选择的高维序列数据超核模型及其学习方法，使得监督或无监督算法可以从约简数据中学习自适应的核函数；(2)Hilbert空间数据模式的描述性概率模型及模型评价方法；(3)高维序列数据非线性子空间聚类算法和多类分类器；(4)基于临床时序数据核学习的患者病情模式及疾病相关风险因子识别。项目致力于两类核方法(核密度估计和Mercer核)的有机结合，以创新性的概率模型方法为基础，探索数据驱动的复杂数据核函数学习、核空间数据模式显式刻画等经典核学习理论和应用中若干瓶颈问题的解决方案，构建独具特色的医用辅助决策系统，对推动统计学习理论和方法的基础研究具有科学意义。

高维序列是一种广泛存在于工业控制、医疗健康等领域的复杂数据形式。本项目以序列特征模型为基础，聚焦其变长性、时序性和异质性等特点，围绕申请书提出的研究目标，以医院电子病历、图象序列、生物序列等实际应用领域产生的数据为主要研究对象，开展这种复杂类型数据的非线性机器学习新模型、新算法研究，取得了若干成果。. 首先，提出了概率空间嵌入方案，其基本思路是将数据对象嵌入到特定概率空间中，继而在嵌入空间中进行核学习等，并将自适应核函数构造问题转换为数据驱动的概率模型参数学习问题。项目研究中，我们基于核密度估计、马尔可夫模型和隐马尔可夫模型等提出并验证了多种嵌入方法，为其他复杂类型数据特别是非结构化数据的机器学习提供了一个具有参考意义的研究思路。其次，面向模式识别和重要特征识别应用，提出嵌入特征选择的核学习方法，多种嵌入型特征约简方法及嵌入空间中非线性表征学习和模式评价方法；特别地，我们建立了拓展经典模式评价方法至嵌入空间的有效途径，提出了多种面向符号数据的模式有效性指标。第三，应用理论成果于COPD（慢性阻塞性肺疾病）患者的生存分析中，基于患者临床诊疗及随访产生的真实时序数据，成功研发生存学习机(survival learning machine)等，用于患者生存概率和生存时间预测及生存状态相关关键风险因子（特征）识别，已被某医学研究中心成功运用于患者追踪学习和临床决策。. 项目组共发表基金标注论文33篇，其中被SCI/EI收录20篇，在IEEE TPAMI、TNNLS、KAIS和《软件学报》、IEEE ICDM等CCF推荐A/B类期刊或学术会议上发表7篇；获授权国家发明专利2项，培养研究生12名，获优秀学位论文奖3项，国际学术会议最佳应用论文奖1项。达到预期研究目标。.