Research On Rare Features Recognition Method For Astronomical Big Data

As one of the main tasks in data mining and machine learning.Efficient classifications have been concerned by scholars all over the world,and successfully applied in various fields,such as medicine,meteorology,finance,etc.With the worldwide attention to the space strategy,more and more spectral data are accumulated by advanced observation technology.Searching and identifying the astronomical objects with rare features is significant for humans to know the universe deeply.And the spectra with double-peaked profiles can provide important evidence to study the science tasks,including binary AGNs,galactic pairs,etc.The double-peaked profiles presented in spectra are characterized by rarity,complexity and diversity,which greatly increase the difficulty of finding,searching and identifying in massive spectral data.In this paper,a classification method including the feature extraction,search and recognition,result evaluation of double-peaked profiles is studied,based on the relevant subspace,SVM,and concept lattice to address the recognition of these data observed by LAMOST telescope.The main contents are as follows:1.A feature extraction method of double-peaked profiles based on relevant subspace is proposed,which resolves the problem of massive,high-dimensional,and feature sparseness of spectra data.The sparse factor is used to describe the sparseness of the local dataset on dimensions,so as to further analyze the frequency of different dimensions in the dataset.The recognized double-peaked profiles spectra of LAMOST are selected to verify the correctness of the method by experiments.The results show that the dimensions are reduced effectively.The 8 feature subspaces extracted with high frequency are important feature lines confirmed by expects,in which the double-peaked profiles are presented probably.In addition,the certified feature subspace and its basic properties are described formally,which is the foundation for further searching and identifying of double-peaked profiles.2.A double-peaked profiles recognition method DoPS based on SVM is proposed.First,on the basis of double-peaked profiles extraction method,theApriori algorithm is used to mine the frequent item sets between the feature subspaces and they are analyzed by the up/down approximation theory,and then the feature subsets are grouped according to their inherent correlation.On this basis,the classifier based on hyperplane is built using the support vector thresholds by SVM.Second,five datasets with different sizes are selected from LAMOST to be applied in the experiments.The result shows that the DoPS algorithm has great advantages over other similar methods in terms of efficiency,accuracy,recall rate,etc.Finally,a spectrum with P-Cygni profiles found during the certification process is analyzed theoretically,which has great significance for the study of rare celestial bodies and the improvement of the universe evolution theory.3.A recognition and evaluation frame SVM-Lattice for double-peaked profiles is proposed,based on DoPS and concept lattice,aiming at the feature correlation of DoPS and its result evaluation.The DoPS is used to define the intent and extent of the node.The intent is the support vectors trained by the double-peaked profiles data,and the extent is the positive sample upper the hyperplane.Each node in SVM-Lattice presents a DoPS classifier,and its result is evaluated according to the relationship between layers.On the basis,the SVM-Lattice build and evaluation algorithms are proposed.Different datasets of LAMOST and similar methods are selected for the experiment analysis.The result shows that the SVM-Lattice has better performance in efficiency,accuracy,etc,which verifies the effectiveness and feasibility of the method.