Deep Regression Using Incomplete Data without Input-Output Correspondence

Abstract

The data to be analyzed in cases where it is difficult to collect comprehensive data, such as when privacy needs to be prioritized, or a non-centralized approach is used, are often given by incomplete data where the correspondence between input values (feature vectors) and output values (target values) is unknown. This study proposes a deep learning method for estimating regression functions from such incomplete data without input-output correspondence. We also develop a stochastic sparse EM algorithm that iteratively updates the discrete latent variables representing the input-output correspondence and the parameters of the deep model. Experiments on benchmark data confirm that the proposed method, which exploits the high expressive power of the deep model, outperforms existing methods based on linear models.