Abstract:

Extreme Learning Machine（ELM） is an effective classification learning algorithm which has the advantages of fast learning speed, high generalization performance, and good approximation ability. With the development of unsupervised learning, integrating ELM with autoencoders has become a new perspective for extracting features using unlabeled data sets, for example, Extreme Learning Machine Autoencoder（ELM-AE）, which is an unsupervised neural network. The main components representing the original sample and its learning process can be found without iteration. It reconstructs the input signal to obtain the main characteristics of the original sample, and considers the global information of the original data to avoid the loss of information. However, this type of method does not consider the inherent Manifold structure of the data, that is, the neighbor structure relationship between the samples. This paper draws on the idea of extreme learning machine autoencoder, and proposes a Manifold-based Extreme Learning Machine autoencoder algorithm （M-ELM）. The algorithm is a nonlinear unsupervised feature extraction method, which combines manifold learning to maintain local information of the data, and simultaneously learns the similarity matrix during feature extraction instead of using a specific formula to calculate the similarity between samples. By conducting experiments on IRIS data set, EEG data set and gene expression data set, the accuracy of this algorithm and other unsupervised learning methods including PCA, LPP, NPE, LE and ELM-AE algorithm after [k]-means are compared, to show the effectiveness of this algorithm.

Key words: extreme learning machine, extreme learning machine autoencoder, manifold learning, unsupervised learning, feature extraction

摘要：

极限学习机（ELM）作为一种无监督分类方法，具有学习速度快、泛化性能高、逼近能力好的优点。随着无监督学习的发展，将ELM与自动编码器集成已成为无标签数据集提取特征的新视角，如极限学习机自动编码器（ELM-AE）是一种无监督的神经网络，无需迭代即可找到代表原始样本和其学习过程的主要成分。其重建输入信号获取原始样本的主要特征，且考虑了原始数据的全局信息以避免信息的丢失，然而这类方法未考虑数据的固有流形结构即样本间的近邻结构关系。借鉴极限学习机自动编码器的思想，提出了一种基于流形的极限学习机自动编码器算法（M-ELM）。该算法是一种非线性无监督特征提取方法，结合流形学习保持数据的局部信息，且在特征提取过程中同时对相似度矩阵进行学习。通过在IRIS数据集、脑电数据集和基因表达数据集上进行实验，将该算法与其他无监督学习方法PCA、LPP、NPE、LE和ELM-AE算法经过[k]-means聚类后的准确率进行了比较，以表明该算法的有效性。

关键词: 极限学习机, 极限学习机自动编码器, 流形学习, 无监督学习, 特征提取