This paper proposes a two-level fe ature selection to improves NaIve Bayes with kernel density estimation. The performance of the proposed fe ature selection is evaluated on question item set based on Bloom's cognitive levels. This two level fe ature selection contains of filter and wrapper based fe ature selection. This paper uses chi square and information gain as the filter based fe ature selection and fo rward fe ature selection and backward fe ature elimination as the wrapper based fe ature selection. The result shows that the two-level fe ature selection improves the NaIve Bayes with kernel density estimation. The combination of chi square and backward fe ature elimination give more optimal quality than the other combination.Keywords-bloom 's cognitive level; nai've bayes; kernel density estimation; filter based fe ature selection; wrapper based fe ature selection