缺失数据下极限学习机改进算法及其应用

发布时间：2019-02-28 14:05

【摘要】：二十一世纪以来，人类加快了对海洋资源探索的步伐，在大洋深海的中存在的大量的海洋资源，因此各个国家都在积极争取各自的海洋利益。在探索海底矿产资源的过程中，通过各种手段获得的大量数据存在缺失的情况，而处理缺失数据的矿产资源评价算法研究较为滞后。本文旨在研究能够处理带缺失数据的机器学习算法，以实现对海底矿产资源数据的利用和矿产资源的评价。极限学习机（ELM）是近年来机器学习算法中较为成功的一个，其具有速度快和泛化能力强的特点。稀疏贝叶斯ELM（SBELM）算法和TROP-ELM算法是标准ELM算法最新的两个变种，，SBELM算法能够解决多分类问题，TROP-ELM能够解决回归问题。与标准ELM算法相比，以上两种ELM算法的变种不仅保留了ELM算法的优点，还能够得到更加紧凑的网络结构。然而，以上两种算法无法处理带缺失数据的数据集。为了解决这个问题，本文提出了两种基于ELM的处理缺失数据的方法，分别是：处理缺失数据的加性模型方法（Additivemodels for missing data，AM-MD）和处理缺失数据的自调节神经元状态模型（Self-adjusting neuron state models for missing data，SNSM-MD）。在UCI公开数据集上的实验结果说明了改进的SBELM算法和TROP-ELM算法与其它处理缺失数据的方法比，具有更好的泛化能力和效果。不仅如此，即使数据集中属性值缺失率达到50%，本文提出的ELM改进算法仍然具有较好的效果。最后，本文将改进的TROP-ELM算法用于带缺失数据的大洋海底钻孔数据的回归分析。
[Abstract]:Since the 21 century, mankind has accelerated the pace of exploration of marine resources, in the deep ocean of the existence of a large number of marine resources, so each country is actively striving for their own marine interests. In the process of exploring seabed mineral resources, a large number of data obtained by various means are missing, while the research on the evaluation algorithm of mineral resources which deals with missing data is lagged behind. The purpose of this paper is to study the machine learning algorithm which can process the missing data in order to realize the utilization of seabed mineral resources and the evaluation of mineral resources. Extreme Learning Machine (ELM) is one of the most successful machine learning algorithms in recent years. It has the characteristics of fast speed and strong generalization ability. Sparse Bayesian ELM (SBELM) algorithm and TROP-ELM algorithm are two new variants of standard ELM algorithm. SBELM algorithm can solve multi-classification problem and TROP-ELM algorithm can solve regression problem. Compared with the standard ELM algorithm, the variants of the above two ELM algorithms not only retain the advantages of the ELM algorithm, but also obtain a more compact network structure. However, the above two algorithms cannot deal with data sets with missing data. In order to solve this problem, two methods of processing missing data based on ELM are proposed in this paper. They are: additive model method (Additivemodels for missing data,) to deal with missing data. AM-MD and the self-regulated neuron state model (Self-adjusting neuron state models for missing data,SNSM-MD) for processing missing data. The experimental results on UCI open datasets show that the improved SBELM algorithm and TROP-ELM algorithm have better generalization ability and effect than other methods for processing missing data. Moreover, even if the missing rate of attribute value in data set is up to 50%, the improved ELM algorithm presented in this paper still has a good effect. Finally, the improved TROP-ELM algorithm is applied to regression analysis of ocean bottom borehole data with missing data.
【学位授予单位】：清华大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：P744;TP181

【相似文献】