Rejection via Learning Density Ratios