With the increase of known RR Lyrae stars, it is reliable to create classifiers of RR Lyrae stars based on their photometric data or combined photometric and spectroscopic data. Nevertheless the total number of known RR Lyrae stars is still too small compared with the large survey databases. So classification of RR Lyrae stars and other sources belongs to imbalanced learning. Based on Sloan Digital Sky Survey (SDSS) photometric and spectroscopic data, we apply cost-sensitive Random Forests fit for imbalanced learning to preselect RR Lyrae star candidates. Only with photometric data, u-g,g-r,r-i,i-z is the best input pattern. While also considering physical parameters (Teff, [Fe/H], log(g)), the optimal input pattern is Teff, [Fe/H], log(g), u-g,g-r,r-i,i-z, at this moment for cost-sensitive Random Forests, the performance metrics of completeness, contamination, and Matthews correlation coefficient are 0.975, 0.019, and 0.975, respectively. It indicates that adding stellar physical parameters is helpful for identifying RR Lyrae stars from other stars. We apply the best classifiers on the SDSS photometric data and combined photometric data with physical parameters to select RR Lyrae star candidates. Finally 11,041 photometric candidates with spectral type A and F are obtained, and then 304 candidates with physical parameters are selected out. Among the 304 candidates, a small part are HB stars, BS stars, RGB stars, and peculiar stars, and the rest are unknown in the Simbad database. These candidates may be used as the input catalog for time-series follow-up observations.