Learning conditional preference networks