New incomplete data imputation based on k-nearest neighbor type framework

Data classification methods such as KNN and LMKNN have received extensive attention in machine learning, but these methods can only perform on complete datasets. However, incomplete data is very common in real life. Therefore, based on the classification method, this paper proposes five new methods to impute missing data: LMKNN_Imputation, LMPNN_Imputation, Weighted KNN_Imputation, Weighted LMKNN_Imputation, and Weighted LMPNN_Imputation. For the first two methods, based on the classification method, the k nearest neighbors / k-nearest pseudo neighbors' local mean vectors in the class closest to the missing value are to be found first and then used local mean vectors to impute the missing data. For the last three methods, a larger weight is assigned to the point / local mean vector of the class closest to the missing value. Then, use the weighted average or local mean vector to substitute the missing value. After completing the data, classification is performed in five different datasets which are imputed based on five new methods, mean substitution, and KNN imputation, and test the effectiveness of five new methods through the classification accuracy. Finally, the experimental results show that these five methods have an accurate prediction for missing values, the LMPNN_Imputation and the three weighted methods performed even better.