Computer Engineering and Applications ›› 2018, Vol. 54 ›› Issue (3): 109-114.DOI: 10.3778/j.issn.1002-8331.1608-0296

Previous Articles     Next Articles

Semi supervised pattern classification method based on Tri-DE-ELM

WU Mingsheng, DENG Xiaogang   

  1. College of Information and Control Engineering, China University of Petroleum, Qingdao, Shandong 266555, China
  • Online:2018-02-01 Published:2018-02-07

基于Tri-DE-ELM的半监督模式分类方法研究

吴明胜,邓晓刚   

  1. 中国石油大学(华东) 信息与控制工程学院,山东 青岛 266555

Abstract: When Extreme Learning Machine(ELM) is used, unlabeled samples are not fully utilized and the training accuracy is affected by the initial value of the network. Aiming at this problem, this paper proposes a modified ELM algorithm based on cooperative training and differential evolution(Tri-DE-ELM). Firstly, considering the issue that traditional ELM pattern classification techniques only use a small amount of labeled samples and ignore a large number of unlabeled samples, co-training mechanism based on Tri-training algorithm is applied to construct the semi-supervised Tri-ELM classification algorithm, which trains three base classifiers using small amount of labeled samples to achieve the labels of the unlabeled sample. Further, in view of the problem that the classification results are affected by random initialization of the ELM network input layer weights, Differential Evolution(DE) algorithm is used to optimize the initial value of the network. The optimization procedure considers two factors of network weights and the classification error, to avoid over fitting phenomenon. Experimental results on standard data sets show that the Tri-DE-ELM algorithm can utilize unlabeled data effectively, and has a higher classification accuracy than the traditional ELM.

Key words: Extreme Learning Machine(ELM), differential evolution, Tri-Training algorithm, semi-supervised learning

摘要: 针对极限学习机(ELM)未充分利用未标注样本、训练精度受网络权值初值影响的问题,提出一种基于协同训练与差分进化的改进ELM算法(Tri-DE-ELM)。考虑到传统的ELM模式分类技术只利用了少量标注样本而忽视大量未标注样本的问题,首先应用基于Tri-Training算法的协同训练机制构建Tri-ELM半监督分类算法,利用少量的标记样本训练三个基分类器实现对未标记样本的标注。进一步针对基分类器训练中ELM网络输入层权值随机初始化影响分类效果的问题,采用差分进化(DE)算法对网络初值进行优化,优化目标及过程同时包括网络权值和分类误差两方面的因素,以避免网络的过拟合现象。在标准数据集上的实验结果表明,Tri-DE-ELM算法能有效地利用未标注数据,具有比传统ELM更高的分类精度。

关键词: 极限学习机, 差分进化, Tri-Training算法, 半监督学习