Separating Quasars from Stars by Support Vector Machines

Based on survey databases from different bands, we firstly employed random forest approach for feature selection and feature weighting, and investigated support vector machines (SVMs) to classify quasars from stars. Two sets of data were used, one from SDSS, USNO-B1.0 and FIRST (short for FIRST sample), and another from SDSS, USNO-B1.0 and ROSAT (short for ROSAT sample). The classification results with different data were compared. Moreover the SVM performance with different features was presented. The experimental result showed that the accuracy with FIRST sample was superior to that with ROSAT sample, in addition, when compared to the result with original features, the performance using selected features improved and that using weighted features decreased. Therefore we consider that while SVMs is applied for classification, feature selection is necessary since this not only improves the performance, but also reduces the dimensionalities. The good performance of SVMs indicates that SVMs is an effective method to preselect quasar candidates from multiwavelength data.