AutoClass is an unsupervised Bayesian classification approach which seeks a maximum posterior probability classification for determining the optimal classes in large data sets. Using stellar photometric data from the Sloan Digital Sky Survey (SDSS) data release 7 (DR7), we utilize AutoClass to select non-stellar objects from this sample in order to build a pure stellar sample. For this purpose, the differences between PSF (point spread function) magnitudes and model magnitudes in five wavebands are taken as the input of AutoClass. Through clustering analysis of this sample by AutoClass, 617 non-stellar candidates are found. These candidates are identified by NED and SIMBAD databases. Most of the identified sources (13 from SIMBAD and 28 from NED respectively) are extragalactic sources (e.g., galaxies, HII, radio sources, infrared sources), some are peculiar stars (e.g., supernovas), and very few are normal stars. The extragalactic sources and peculiar stars of the identified objects occupy 94.1%. The result indicates that this method is an effective and robust clustering algorithm to find non-stellar objects and peculiar stars from the total stellar sample.
We integrate k-Nearest Neighbors(kNN) into Support Vector Machine(SVM) and create a new method called SVM-kNN.SVM-kNN strengthens the generalization ability of SVM and apply kNN to correct some forecast errors of SVM and improve the forecast accuracy.In addition,it can give the prediction probability of any quasar candidate through counting the nearest neighbors of that candidate which is produced by kNN.Applying photometric data of stars and quasars with spectral classification from SDSS DR7 and considering limiting magnitude error is less than 0.1,SVM-kNN and SVM reach much higher performance that all the classification metrics of quasar selection are above 97.0%.Apparently,the performance of SVM-kNN has slighter improvement than that of SVM.Therefore SVM-kNN is such a competitive and promising approach that can be used to construct the targeting catalogue of quasar candidates for large sky surveys.
