Missing Values Imputation Using Auto-associative Neural Network with Local Neighborhood Information

Missing values of real-world datasets bring challenges to data mining. Popular methods for handling missing values are to replace missing values in a dataset using predicted values, called imputation, which is a fundamental task in many research fields. In this paper, we propose an auto-associative neural network with local neighborhood information for missing values imputation. Specifically, we introduce local neighborhood information into auto-associative neural network to make the spatial relationships among samples work. With the spread of neighbor information, the similarity of samples is used to supplement the missing attributes of incomplete samples, which facilitates network analysis of attributes associations. Besides, we propose a training method to iteratively update estimation by treating missing values as variables, in which the imputation error and the incomplete model inputs are continuously optimized during training. The training method allows incomplete samples to participate in the network training as well, enabling all known attributes to be useful. Experiments based on several public datasets validate the effectiveness of the proposed method.