A multi-task learning-based generative adversarial network for red tide multivariate time series imputation

Abstract

Red tide data are typical multivariate time series (MTS) and complete data help analyze red tide more conveniently. However, missing values due to artificial or accidental events hinder further analysis of red tide phenomenon. Generative adversarial network (GAN) is effective in capturing distribution of MTS while the imputation performance is far from satisfactory, especially in conditions of high missing rate. One of the remaining open challenges is that common GAN-based imputation methods usually lack the ability to excavate implicit correlations between different attributions and downstream tasks, from which advanced latent information about missing values can be mined to improve imputation performance. To deal with the problem, a novel multi-task learning-based generative adversarial imputation network (MTGAIN) is proposed by introducing the prediction task into GAN to unearth more detailed information about missing values to better model distribution of red tide MTS. Furthermore, the homoscedastic uncertainty of multiple tasks is exploited to balance the weights of losses between generation and prediction tasks. The experiments conducted on a real-world dataset demonstrate that MTGAIN outperforms existing methods in terms of imputation and post-imputation performances, especially in conditions of high missing rate.