Photometric Determination of Quasar Candidates

We describe an efficient and fast method for the detection and classification of quasars using a machine learning tool, making use of photometric information from SDSS DR7 data release. The photometric information used are the ten independent colours that can be derived from the 5 filters available with SDSS and the machine learning algorithm used is a difference boosting neural network (DBNN) that uses Bayesian classification rule. An adaptive learning algorithm was used to prepare the training sample for each region. Cross validations were done with SDSS spectroscopy and it was found that the method could detect quasars with above 96.96% confidence regarding their true classification. The completeness at this stage was 99.01%. Contaminants were mainly stars and the incorrectly classified quasars belonged to a few specific patches of redshifts. Color plots indicated that the colors of some stars and quasars in those redshits were indistinguishable from each other and was the major cause of their incorrect classification. A confidence value (computed posterior Bayesian belief of the network) was assigned to every object that was classified. Most of the incorrect classifications had a low confidence value. This information may be used to filter out contaminants and improve the classification accuracy at the cost of reduced completeness.