Aiming at the deficiency of traditional data imputation methods in effectively using the label information and random characteristics of missing data, a particle swarm optimization based imputation method for mixed features was proposed. The value of continuous feature was modeled as Gaussian distribution, and the mean and standard deviation were used as optimization parameters. The value probability of categorical features was optimized as a parameter. The classification accuracy rate was used as the optimization target to make full use of random information of label information and missing data. Four statistical methods and two evolutionary algorithm based imputation methods were used to compare the results on six typical classification datasets. The results show that the proposed method significantly outperforms other comparison algorithms in terms of classification accuracy indicator, and has better time overhead at the same time, which can effectively solve the data missing problems of mixed features.