Exploration of Approach to Mining WDMS Spectra based on Laplacian Eigenmap and Neural Network

Abstract

For the purpose of discovering White Dwarf +Main Sequence (WDMS) from massive spectra, in this paper, an unsupervised learning algorithm for Nonlinear Dimensionality Reduction named Laplacian Eigenmap is discussed. It turns out that, comparing with Principle Component Analysis (PCA), Laplacian Eigenmap maintains the information of nonlinear structure of high dimensional spectral data, which leads to a higher classification accuracy. In the feature space, backpropagation neural network is used to classify WDMS and non-WDMS spectra. Furthermore, Particle Swarm Optimization (PSO) is implemented to increase the classification accuracy via optimizing the parameters of the network. The results shows that the method in this paper can discover WDMS efficiently and accurately after training the neural network with low-dimensional data from Sloan Digital Sky Survey Data Release 10 (SDSS-DR10). Introduction With the rapid development in data science, the concept of big data attracts increasingly more attention. As the limitations of traditional algorithm used in data analysis arise, how to apply machine learning algorithm to big data has become a popular topic. Sloan Digital Sky Survey (SDSS), started in 2000, is a large redshift survey project, using a 2.5 m diameter telescope which is sited in Apache Peak Observatory in New Mexico to observe, recording nearly two millions spectral data, which include more than 80 million galaxies and more than 10 million quasar spectra data. WDMS is a very special binaries system, which is the progenitor star of Ia supernovae and cataclysmic variable star and is worthy of being studied. However, template matching methods based on the physical parameters, which generally were used to classify the spectrum, bring about great artificial intervention. Since astronomical spectra belong to high-dimensional data, how to find its structural features from the high dimensional data and furtherly use appropriate algorithms to reduce data dimensionality becomes a key problem in data preprocessing of the machine learning algorithm. Previously, Tan Dong-mei [1] used Principle Component Analysis (PCA) to classify stellar spectra rapidly. Connolly et al. [2] used PCA to extract the feature of known redshift galaxy spectra, discovering that some former main component spectra of galaxies have strong linear relation. Madgwick et al. [3] used PCA to classify emission line and absorption line spectra. However, although PCA can reconstruct linear-independent components from the data, it still uses the Euclidean distance to measure the sample space in essence. But the high dimensional spectral data have a strong non-linear structure so that the dimensionality reduced by PCA cannot accurately describe the distance between samples. In this paper, Laplacian Eigenmap is used to analysis and process WDMS spectral data and reconstruct them in a low dimension space, then backpropagation neural network is used to classify the date. It turns out that Laplacian Eigenmap performs better than PCA. Furthermore, the initial weights and threshold value of neural network have a great impact on the results. Using Particle Swarm Optimization to optimize the parameters of the neural network, the result shows that the accuracy of BP neural network reaches 88.97%, greatly improving the accuracy of discovering WDMS. 2nd International Conference on Electrical, Computer Engineering and Electronics (ICECEE 2015) © 2015. The authors Published by Atlantis Press 986 The rest of this paper is organized as follows. Section 1 introduces the fundamental idea and implementation tricks about Laplacian Eigenmap. Then demonstrates how to optimize controlling arguments of backpropagation neural network (BPNN) using particle swarm optimization (PSO) algorithm. Experimental results and analysis are presented in Section 2. Finally, conclusion and the future extension of the model are drawn at last. Section