Correcting the negativity of high-order kernel density estimators