生物组学数据的集成特征选择研究

发布时间：2017-12-31 10:39

本文关键词：生物组学数据的集成特征选择研究　出处：《深圳大学》2017年博士论文　论文类型：学位论文

【摘要】：高通量技术的发展产生了大量与基因、蛋白质和代谢相关的生物组学数据。从生物组学数据中发现和提炼与疾病相关的信息一直是生物信息学领域的热点问题。通过模式识别中的特征选择和分类技术,对高维生物组学数据中蕴藏的重要信息进行提取、筛选、识别和分类是分析生物组学数据的常用方法。生物组学数据具有的高维小样本、多类间样本分布不平衡等多种特征和样本统计分布特性,对特征选择算法在分类泛化性和稳定性等方面的表现提出巨大挑战,为了精确和稳定地提取与分类目标高度相关的特征子集,需要我们深入地研究和设计针对生物组学数据特征选择问题的新方法。本文针对生物组学数据的不同特性,提出了一系列集成特征选择算法,主要工作概括如下:1.在研究多种特征相关性度量的基础上,提出一种基于多种相关性度量的集成最大相关最小冗余(maximum Relevance and Minimum Redundancy,mRMR)特征选择算法。我们分析了最大信息系数、皮尔逊相关系数和互信息量在特征相关性度量方面的差异,并对前向搜索算法进行改进,增加可以设置所选择特征个数和搜索范围的机制,基于3种特征相关性度量和改进的搜索算法获得三组最优特征子集并进行集成分类学习。对不同类型的多组生物组学数据集的分类对比实验结果表明,集成m RMR特征选择算法能够针对各种类型的生物组学数据有效地选择最优特征子集,并有助于分类算法获得良好的识别性能。2.针对生物组学数据普遍存在多个最优或次优特征子集的情况,提出基于小生境二进制粒子群优化的集成特征选择算法。基于单一特征子集构建的分类模型在小样本生物组学数据上容易产生过拟合而影响分类泛化性能。为了尽量避免该问题,本文通过小生境二进制粒子群优化算法作为特征子集搜索算法,获得多个差异度最大且分类性能最优的特征子集并进行集成分类学习,由投票技术所集成的强分类器体现出优良的稳定性和泛化性。3.为了克服生物组学数据类间样本不平衡特性对特征选择和分类过程的影响,本文提出一种迭代式集成特征选择算法。该算法利用2种样本平衡预处理方法和3种过滤式生物组学数据特征选择算法,通过样本平衡和特征选择不断迭代的方式,使特征选择在一个趋于平衡的样本分布中迭代完成。对多组具有类间样本不平衡特性的生物组学数据进行分类实验,结果表明,由于克服了样本不平衡特性对特征选择的影响,所设计的迭代集成特征选择算法可以进一步提高分类性能。4.针对迭代集成特征选择所具有的局限性,本文提出一种基于粒子群优化的集成特征和模型选择算法。该算法将候选样本平衡模型、特征选择模型和分类模型的选择以及相应模型的超参数编码到粒子中。通过粒子种群的优化,自适应搜索具有最优分类性能的模型组合以及对应模型的超参数设置。实验结果表明,基于粒子群优化的集成特征和模型选择算法能够根据不同生物组学数据的样本和特征分布特性,自适应选择样本平衡模型和特征选择模型的最优组合,避免人为选择和设置模型所引入的主观偏差。本文所提出的一系列集成特征选择算法有助于解决生物组学数据复杂的样本和特征分布特性对特征选择和分类所造成的困难,所使用的集成特征选择思路可为后续相关生物组学数据分析提供借鉴。
[Abstract]:Produced a large number of genes and the development of high-throughput technologies, biological data and metabolism related proteins. Proteomics data from biological discovery and refining and disease related information has been a hot issue in the field of bioinformatics. By using feature selection in pattern recognition and classification technology, the important information of high dimensional biological group the data contained in the extraction, screening, identification and classification is used to analyze biological omics data. Littlescale bio omics data with the multi class samples of unbalanced distribution of a variety of features and the statistical distribution characteristic, the challenge for the feature selection algorithm in the classification generalization and stability other aspects of the performance, in order to accurately and stably extract feature subset is highly correlated with the classification of the target, we need to further research and design for the new bio omics data feature selection problem Method. In this paper, according to different characteristics of biological data set, put forward a series of integrated feature selection algorithm, the main work is summarized as follows: 1. based on the correlation of various features measurement, and put forward a multiple correlation measure integrated optimization based on (maximum Relevance and Minimum Redundancy, mRMR) feature selection algorithm. We analysis of the maximum information coefficient, Pearson correlation coefficient and mutual information measure of the difference in the characteristics of correlation, and improvement of the forward search algorithm, which can increase the set number of feature selection mechanism and the scope of the search, the search algorithm measure and improve the 3 kinds of features based on the correlation of three groups are obtained and the best subset of features the integrated classification study. The classification and comparison of experimental results of different types of multi omics data sets show that the integrated m RMR feature selection algorithm can pin The biological group various types of data to select the optimal feature subset, and contribute to the classification algorithm to obtain good performance of.2. identification data for biological ubiquitous multiple optimal or sub feature subset selection algorithm is proposed, the integrated characteristics of niche binary particle swarm optimization based on classification generalization performance the classification model of single feature subset construction in small samples of biological omics data on prone to overfitting. Based on in order to avoid this problem, this paper through the niche binary particle swarm optimization algorithm for feature subset search algorithm, obtained a lot of difference and the maximum classification performance of optimal feature subset and ensemble learning, strong the classifier integrated by voting technology showed excellent stability and generalization of.3. in order to overcome the omics data between the sample imbalance characteristics of special Effect of feature selection and classification process, this paper proposes an iterative ensemble feature selection algorithm. 2 kinds of sample pretreatment methods and the balance of 3 kinds of filtering biological omics data feature selection algorithm using the algorithm, the sample balance and iterative feature selection, the feature selection iteration in a balance the distribution of samples. Data classification experiment is performed on groups of biological groups between class imbalance characteristics showed that due to overcome the unbalanced samples characteristics of feature selection, the limitations of the iterative ensemble feature selection algorithm can further improve the classification performance of.4. in accordance with the iterative ensemble feature selection in this paper, a particle swarm optimization model and ensemble feature selection algorithm based on the algorithm. The candidate sample balance model, feature selection and classification model selection model And the corresponding model to super parameter encoding particles. By optimizing the particle population, set parameter has the optimal performance of the classification model combination and the corresponding model of adaptive search. The experimental results show that the particle swarm optimization algorithm for model selection and integration features according to different biological omics data samples and distribution characteristics based on adaptive to select the optimal combination sample balance model and feature selection model, avoid subjective bias introduced by artificial selection and setting of the model. This paper proposed a series of integrated feature selection algorithm helps to solve biological omics samples and feature distribution characteristics of complex data caused by feature selection and classification problems, using integrated features select the ideas for the follow-up biological omics data analysis to provide reference.

【学位授予单位】：深圳大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：Q811.4;TP18

【相似文献】