基于多核表示和模糊近似的混合数据分类方法研究

In real-world applications, lots of classification problems can be expressed by the heterogeneous data, such as the symbols, numbers, time series, and images are commonly used for recognizing the diseases of patients in the medical diagnosis. Although the domain expert can easily distinguish the category of the patient's disease by fusing the information of all the modes, the existing techniques in machine learning and data mining fields cannot handle these complex data. In our project, the abundant kernel function family in the kernel methods are utilized to extract the relationships among the objects described by different information. Valuable information is extracted through different kernel functions. Moreover, the fuzzy rough set is used to conduct the approximate reasoning by fusing the information of different modes. A kind of kernel machines are thus developed to modelling the classification of the complicated data. The fuzzy relationships are computed based on different kinds of data, Therefore, the evaluate index of features and the theory of feature reduction are designed for the heterogeneous data through analysing the ability of the fuzzy relationships approximating the decision. Basing on the research results of the statistical learning theory, the modeling and reasoning approaches of the large margin fuzzy approximate reasoning based on the kernel learning are proposed. Therefore, the fuzzy rule learning and fuzzy approximate reasoning in the kernel space are implemented.

在实际应用中大量分类问题是由混合数据表示的，如医疗诊断中广泛使用符号、数值、时间序列甚至图片识别病人的疾病。尽管领域专家能够自如的融合各种类型的信息判断病人的疾病类别，但当前的机器学习和数据挖掘技术却难以处理此类复杂数据。本课题利用核方法中丰富的核函数族提取各种信息所描述对象间的关系，采用多种核函数分别抽取不同类型数据中的有效信息。然后利用模糊粗糙集融合不同类型的信息进行近似推理，从而开发出适合于混合数据分类建模的一类核机器。通过分析不同类型数据计算得到的模糊关系对决策的近似能力，设计了混合数据的属性评价指标和属性约简理论。基于统计学习理论的研究成果，本课题进一步提出基于核学习的大间隔模糊近似建模和推理方法，从而实现核空间的模糊规则学习和模糊近似推理。

本项目针对实际中大量学习问题面对的是复杂数据这一难题，如医疗诊断中广泛使用符号、数值、时间序列甚至图片识别病人的疾病。当前的机器学习和数据挖掘技术难以有效处理此类数据。本课题利用核方法中丰富的核函数族提取各种信息所描述对象间的关系，采用多个核函数分别抽取不同类型数据中的有效信息。然后利用模糊粗糙集技术融合各种信息进行近似推理，从而开发出适合于复杂数据分类建模的一类方法。给出一致函数下模糊逼近算子的基本性质，并提出了基于随机权网络的模糊非线性回归分析方法。提出了一般二元关系上不确定性度量，并提出了基于扩展优势关系的粗糙集模型及其属性约简方法。给出了基于领域模糊粗糙集的特征子集选取方法、多模态混合数据规则提取方法。提出了鲁棒光滑支持向量机和鲁棒光滑一类支持向量机，以及支持向量数据描述集成模型，进一步，给出了多核支持向量机和模糊多核支持向量机模型。。本项目研究成果必将提高和促进数据挖掘技术的发展，在复杂系统的海量数据处理等领域中具有广泛的应用前景。