Why not to use zero imputation? correcting sparsity bias in training neural networks

Abstract

Handling missing data is one of the most fundamental problems in machine learning. Among manyapproaches, the simplest and most intuitive way is zero imputation, which treats the value of a missingentry simply as zero. However, many studies have experimentally confirmed that zero imputation resultsin suboptimal performances in training neural networks. Yet, none of the existing work has explainedwhat brings such performance degradations. In this paper, we introduce thevariable sparsity problem(VSP), which describes a phenomenon where the output of a predictive model largely varies with respectto the rate of missingness in the given input, and show that it adversarially affects the model performance.We first theoretically analyze this phenomenon and propose a simple yet effective technique to handlemissingness, which we refer to asSparsity Normalization (SN), that directly targets and resolves the VSP.We further experimentally validate SN on diverse benchmark datasets, to show that debiasing the effectof input-level sparsity improves the performance and stabilizes the training of neural networks.