An Improved k-Nearest Centroid Neighbor Classification Method for Incomplete Data

Missing values often exist in scientific datasets. Therefore, practical methods for missing data imputation and classification are necessary for machine learning, data analysis. The k-Nearest Neighbor (KNN) algorithm is a simple and effective algorithm in missing data imputation and classification. This paper focuses on the missing data classification problem and proposes a new classification method based on the local mean k-nearest centroid neighbour. When making classification judgments, the proposed method examines the closeness and symmetrical arrangement of the k neighbours and adopts the local mean-based vector of the k centroid neighbours for each class. We run classification error experiments on six UCI datasets to see how well the proposed method performs when there is missing data. Experimental results show that the performance of our proposed method obtains a significant improvement compared to the most advanced KNN-based algorithms.