Classification of incomplete data using classifier ensembles

This paper proposes a method for classification of incomplete data using neural network ensembles. In the method, the incomplete data set is analyzed and projected into a group of complete data subsets that give a full description of the known values in the data set by joining together. Those complete data subsets are then used as the training sets for the neural networks. Base classifiers are selected and integrated according to their classification accuracies and the support degrees of their training data sets to give the final predication. Compared with other methods dealing with missing data in classification, the proposed method can utilize all the information provided by the incomplete data, maintain maximum consistency of the incomplete data set and avoid the dependency on distribution or model assumptions. Experiments on two UCI datasets showed the superiority of the algorithm to other two typical treatments of missing data in ensemble learning.