关键词: 朴素贝叶斯（NB）, 期望最大值（EM）算法, 连续属性, 核密度估计, 蛋白质纯化

Abstract: When dealing with classification problem, Naive Bayes（NB） usually assumes that the numerical continuous attributes follow normal distribution, the classification accuracy is also affected by the integrity of training data. But the actual sampled data are difficult to meet the above requirements. For missing data, the Naive Bayesian classifier uses existing incomplete data to implement parameter learning based on the Expectation-Maximum（EM） algorithm; for non-
normal numerical continuous attributes, distribution density based on kernel density estimation and a new method are used to calculate the maximum posterior probability, meanwhile, the classification experiment using standard data sets verifies the effectiveness of the improvement. Finally, the improved algorithm（EM-DNB） is applied to the prediction of the protein purification technologies in biological engineering. The experimental results show that the accuracy is improved.

Key words: Naive Bayes（NB）, Expectation-Maximum（EM） algorithm, continuous attributes, kernel?density?estimation, protein purification