基于近邻集成保持策略的降维和分类方法研究

发布时间：2021-12-23 19:56

　　降维（DR）和数据分类是两个最重要的机器学习任务,用于许多模式识别应用,如人脸识别,医学成像,指纹识别,语音识别等。邻域保留策略应用在许多著名的算法中,例如邻域保持嵌入（NPE）,局部保留投影（LPP）和k最近邻规则（KNN）。但是这些算法对参数设置非常敏感。例如NPE和LPP对邻域大小的参数非常敏感,这降低了降维的性能。此外,现有的多种DR方法通常利用单个图来保持邻域关系,这种区分不适合于多视图数据集的降维。此外KNN的分类性能受邻域大小k和现有异常值的影响很大。因此本文设计了基于近邻集成保持策略的降维和分类方法研究,旨在减少NPE,LPP和KNN中的上述近邻约束。在第一种DR方法中,我们提出了一种称为加权邻域保持集成嵌入（WNPEE）的新型DR方法。与NPE不同,所提出的WNPEE构造了多个近邻图的集成。通过近邻图的集成构建,WNPEE可以通过联合优化方式获得最优嵌入图的低维投影。对ORL,Georgia Tech,CMU PIE和Yale四种人脸数据集的实验表明,WNPEE实现了比NPE和其他实验对比的DR方法更高的识别率。此外,与NPE和其他相关的DR算法相比,所提出的WNPE... 

【文章来源】：江苏大学江苏省

【文章页数】：146 页

【学位级别】：博士

【文章目录】：
ABSTRACT
摘要
Chapter1 Introduction
    1.1 Background
    1.2 Dimension reduction
    1.3 Challenges in neighborhood related DR techniques
    1.4 Data classification
    1.5 Challenges in nearest neighbor classifiers
    1.6 Research contributions
    1.7 The organization of thesis
Chapter2 Related Work
    2.1 Related DR techniques
        2.1.1 Traditional linear DR techniques
            2.1.1.1 Principal components analysis
            2.1.1.2 Linear discriminant analysis
        2.1.2 Local nonlinear DR approaches
            2.1.2.1 Local linear embedding
            2.1.2.2 Laplacian eigenmaps
        2.1.3 Linear approximations to local nonlinear DR techniques
            2.1.3.1 Linearity preserving projection
            2.1.3.2 Neighborhood preserving embedding
        2.1.4 Multi-view learning techniques
    2.2 Nearest neighborhood based classifiers
        2.2.1 A local mean-based nonparametric classifier
        2.2.2 Nearest centroid neighbor classifier
        2.2.3 The k-nearest centroid neighbor classifier
        2.2.4 A local mean-based k-nearest centroid neighbor classifier
        2.2.5 Harmonic mean distance
        2.2.6 A new k-harmonic nearest neighbor classifier based on the multi-local means
Chapter3 Weighted Neighborhood Preserving Ensemble Embedding
    3.1 Introduction
    3.2 The proposed WNPEE method
        3.2.1 Obtaining weight parameter in proposed WNPEE
    3.3 Experimental results and analysis
        3.3.1 Experiments on ORL face database
        3.3.2 Experiments on GT face database
        3.3.3 Experiments on CMU PIE face database
        3.3.4 Experiments on YALE face database
        3.3.5 Analysing the sensitivity to neighborhood size parameter
        3.3.6 Computation complexity
    3.4 Brief summary
Chapter4 Generalized Multi-manifold Graph Ensemble Embedding for Multi-View Dimensionality Reduction
    4.1 Introduction
    4.2 Proposed Methods
        4.2.1 Obtaining Parameter and in proposed methods
        4.2.2 Classification difference between LPP,EGLPP and MLGEE
    4.3 Experimental Results
        4.3.1 Dataset description
        4.3.2 Experimental settings
        4.3.3 Result discussions and comparisons
            4.3.3.1 Comparison between EGLPP and LPP
            4.3.3.2 Parameter selection in MGLPP
            4.3.3.3 Experimental results for MLGEE
        4.3.4 Computational complexity
    4.4 Brief summary
Chapter5 A New Nearest Centroid Neighbor Classifier Based on K Local Means Using Harmonic Mean Distance
    5.1 Introduction
    5.2 Description of LMKHNCN
    5.3 Comparison with traditional KNN based classifiers
    5.4 Experiment results and discussion
        5.4.1 Performance evaluation
        5.4.2 Description of the datasets
        5.4.3 Experimental procedure
        5.4.4 Analysing the error rates results with corresponding k value
        5.4.5 Results of the sensitivity to the neighborhood size k
        5.4.6 Analysing the effect of distance on classification performance
        5.4.7 Analysing the computational complexity
        5.4.8 Evaluation of experimental results
    5.5 Brief summary
Chapter6 General Conclusions and Future Works
    6.1 General conclusions
    6.2 Future works
References
Acknowledgements
Publications


【参考文献】：
期刊论文
[1]一种有监督的稀疏保持近邻嵌入算法[J]. 郑豪,金忠.  计算机工程. 2011(16)



本文编号：3549089