Searching Method Study On M Subdwarfs Based On Ensemble Learning

M subdwarfs are metal-poor stars with low-mass and low-luminosity,known as the old populations in the Milky Way.The study of M subdwarfs is critical for probing the evolution and composition of the Galaxy.However,M subdwarfs have similar spectral morphology with M dwarfs,which often leads to the confusion between them.In recent years,with the development of machine learning,some new approaches are introduced for the classification of M subdwarfs and M dwarfs.We apply the ensemble learning algorithm based on decision tree built the classification models of M dwarfs and M subdwarfs.The ensemble learning can not only produce the final prediction results,but also effectively evaluate the important features that distinguish the two stars.Furthermore,the experiment showed that our LightGBM model in LAMOST classified M subdwarfs and M dwarfs very well,with an accuracy of 97.22%.And we also applied this prediction model to SDSS successfully.This article is mainly divided into the following four parts.(1)Data preprocessingThe data preprocessing process includes the acquisition of spectral data,interpolation,normalization and removal of atmospheric absorption lines,as well as dataset partitioning.(2)Model constructionIn the experiment,we built the models of SVM,RandomForest,XGBoost and LightGBM in the data from LAMOST DR4 respectively.Among these methods,the LightGBM has the best classification result and the highest efficiency.The first contribution of this paper is to apply LightGBM,an advanced ensemble learning algorithm,to discriminate M dwarfs and M subdwarfs.And the experimental results manifest that the method is competent to deal with this problem(3)Features analysisThe second contribution of this paper is to quantify the important classification features of M dwarfs and M subdwarfs.By analyzing the characteristics evaluated by the LightGBM model,the experimental results show that Ti05,CaHl,CaH2,and CaH3 are very important for distinguishing M subdwarfs from M dwarfs,which is consistent with previous research.In addition,we also find that gravity-sensitive atomic lines(K I and Na I)and metallicity-sensitive atomic line(Ca I)also have an important influence on the classification results.(4)Searching the M subdwarfsIn the end,we also applied the LightGBM model constructed in this paper to 5306 M-type spectra from SDSS DR7 and detected 2538 M subdwarfs successfully.This work provides a new method for the problem of searching M subdwarf dwarfs from large modern spectroscopic surveys.