Sequential Regression with Missing Data Using LSTM Networks

Abstract

We study regression for variable length sequential data suffering from missing samples and introduce a long shortterm memory (LSTM) based sequential regression algorithm. In most sequential regression studies, one considers data sequence is complete, i.e., does not contain any missing data. However, the missing data problem appears in a large number of areas such as finance and medical imaging. The remedies to resolve this problem depends on certain statistical assumptions and imputation techniques. However, the statistical assumptions does not hold in real life and the imputation of artificially generated inputs results in sub-optimal solutions. In our experiments, we achieve significant performance gains with respect to the classical algorithms.