Automated Classification of Pointed Sources

Facing very large and frequently high dimensional data in astronomy, effectiveness and efficiency of algorithms are always the hot issue. Excellent algorithms must avoid the curse of dimensionality and simultaneously should be computationally efficient. Adopting survey data from optical bands (SDSS, USNO-B1.0) and radio band (FIRST), we investigate feature weighting and feature selection by means of random forest algorithm. Then we employ a kd-tree based k-nearest neighbor method (KD-KNN) to discriminate quasars from stars. Then the performance of this approach based on all features, weighted features and selected features are compared. The experimental result shows that the accuracy improves when using weighted features or selected features. KD-KNN is a quite easy and efficient approach to nonparametric classification. Obviously KD-KNN combined with random forests is more effective to separate quasars from stars with multi-wavelength data.