A New Neural Network to Process Missing Data without Imputation

This paper introduces the cosine neural network (COSNN) and shows how it can be used to process data with missing components without imputation. It uses a cosine basis function with a weighted norm which can be trained to match the input data, or it can be set to zero to 'ignore' missing data components. The COSNN is compared to feedforward neural networks using deletion and imputation. The COSNN is shown to be superior in both a function approximation and a classification test set.