Exploration of Various Approaches for Photometric Redshift Estimation of Quasars

We explore various approaches to predict photometric redshifts of quasars with SDSS, WISE and UKIDSS photometric data, compare the effect of various distances and different indexes of k-nearest neighbors (KNN) on the performance of estimating photometric redshifts. Moreover we compare the performance of KNN, partial least squares regression (PLS), ridge regression, Extra-Trees and these methods with lasso for feature selection on the same sample. The results show that KNN with different distance functions have comparable accuracy in terms of percents in different redshift ranges and scatter, but the speed of KNN with Euclidean distance function is the fastest. The performances of KNN with different indexes also have similar performances while KD-Tree KNN is faster. Except ridge regression, the accuracy of other methods with Lasso for feature selection on this issue becomes better. Extra-Trees show its superiority on photometric redshift estimation. Usually feature selection is necessary when dealing with high dimensional data. Lasso used for feature selection is helpful to improve the performance of regressors.