Feature-embedding neural processes for missing value prediction

Abstract

This paper addresses the missing value prediction problem in data. Existing missing value prediction studies apply their algorithms after replacing missing values with a constant value, which does not represent the missing values clearly and has disadvantages in processing data with a high missing rate. Also, it is needed to consider the uncertainty in missing value prediction which is important for model reliability. We propose a model that generalizes a single data sample from its observations and predicts the distribution of the value of a query feature to measure uncertainty. Our model can deal with an incomplete data sample independently of missing values, and the prediction uncertainty can be measured by the output variance. In this process, the feature embedding matrix and feature importance weights are introduced to represent each feature and learned jointly with the proposed model. Through experiments in various aspects, we verify that the proposed model effectively predicts the missing values of data and offers reliable uncertainty measures.