Reducing variance in importance-weighted cross-validation under covariate shift