面向基因表达数据的分类与特征选择方法研究

发布时间：2018-03-27 22:28

本文选题：基因表达数据　切入点：正则极限学习机　出处：《中国计量大学》2016年硕士论文

【摘要】：在生物信息学领域,DNA微阵列技术的出现是一个具有里程碑意义的重大技术突破。随着研究的不断深入,它已被广泛应用于药物研究、基因测序等众多领域,具有极高的应用价值和广阔的发展前景。然而在实际应用中,DNA微阵列技术研究的基因组对象规模越来越大,特征维数也越来越高,最终获得的基因表达数据不仅具有高维小样本的特点,还含有大量和样本分类无关或对样本分类作用很小的冗余基因和噪声基因。基因表达数据的这些特点会提高机器学习的时间及空间复杂度,降低分类精度,最终在疾病诊断等实际应用中会增加成本,降低疾病预测的准确度。因此,为提高基因表达数据的分类精度,本文从两方面着手进行研究:一方面对分类算法进行改进,另一方面通过提出有效的特征选择方法以筛选出关键基因,剔除冗余和噪声基因,降低基因特征维数,提高机器学习效率。研究的主要内容如下:(1)正则极限学习机(RELM)是在极限学习机(ELM)的基础上提出来的,具有简单易用、分类精度较高、泛化能力好等优点。然而RELM的输入层权值、隐含层偏差是随机给定的,会影响RELM的稳定性。另外,RELM为了获得较理想的分类精度,仍需设置较多的隐层节点。针对此问题,通过分析粒子群算法(PSO)的原理,把RELM初始产生的输入层权值、隐含层偏差作为粒子带入PSO进行寻优,提出了一种粒子群改进RELM(PSO-RELM)。在UCI数据集上的仿真实验表明,PSO-RELM相对BP神经网络、支持向量机(SVM)、RELM具有更好的分类精度和更佳的稳定性。(2)提出一种结合互信息最大化(MIM)和自适应遗传算法(AGA)的特征选择方法(MIMAGA-Selection),其中选择ELM作为分类器来计算样本的分类精度。首先根据源数据集中各基因和不同类别的互信息最大化进行分组和筛选,形成一个初选基因子集;然后运用以样本分类精度作为适应度函数的自适应遗传算法对初选基因子集进行寻优,最终得到一个最优基因子集。通过在3个UCI标准数据集上进行的实验表明,该方法能有效剔除冗余基因和噪声基因,显著提高样本分类精度。
[Abstract]:The emergence of DNA microarray technology in the field of bioinformatics is a landmark breakthrough. With the development of research, it has been widely used in many fields, such as drug research, gene sequencing and so on. It has high application value and broad development prospect. However, in the practical application, the genome objects studied by DNA microarray technology are becoming larger and larger, and the characteristic dimension is becoming higher and higher. The obtained gene expression data not only has the characteristics of high dimension and small sample, It also contains a large number of redundant and noisy genes which are independent of sample classification or have little effect on sample classification. These characteristics of gene expression data can increase the time and space complexity of machine learning and reduce the classification accuracy. Finally, the cost will be increased and the accuracy of disease prediction will be reduced in practical applications such as disease diagnosis. Therefore, in order to improve the classification accuracy of gene expression data, this paper studies from two aspects: on the one hand, the classification algorithm is improved. On the other hand, an effective feature selection method is proposed to screen out key genes, eliminate redundant and noise genes, and reduce the dimension of gene features. To improve the efficiency of machine learning. The main contents of the study are as follows: 1) the regular extreme learning machine (RELM) is put forward on the basis of the extreme learning machine (ELM), which has the advantages of easy to use, high classification accuracy, good generalization ability and so on. However, the input layer weight of RELM, The hidden layer deviation is given at random, which will affect the stability of RELM. In addition, in order to obtain the ideal classification accuracy, more hidden layer nodes still need to be set. In order to solve this problem, the principle of particle swarm optimization (PSO) is analyzed. In this paper, the input layer weights generated initially by RELM and hidden layer deviations are brought into PSO as particles for optimization. An improved particle swarm optimization (PSO) PSO-RELMN is proposed. The simulation results on UCI data sets show that PSO-RELM is relative to BP neural networks. Support Vector Machine (SVM) SVM / RELM has better classification accuracy and better stability. (2) A feature selection method combining mutual information maximization (mim) and adaptive genetic algorithm (AGA) is proposed, in which ELM is selected as the classifier to calculate the score of samples. Class accuracy. First, grouping and filtering based on the maximization of mutual information between genes and different categories in the source dataset, A subset of primary genes is formed, and then an adaptive genetic algorithm based on the accuracy of sample classification as fitness function is used to optimize the subsets of primary genes. Finally, an optimal subset of genes is obtained. Experiments on three UCI standard datasets show that the proposed method can effectively eliminate redundant genes and noise genes and improve the classification accuracy of samples.
【学位授予单位】：中国计量大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：Q811.4;TP18

【相似文献】