Abstract
The problem of missing values in multivariable time series is a key challengein many applications such as clinical data mining. Although many imputationmethods show their effectiveness in many applications, few of them are designedto accommodate clinical multivariable time series. In this work, we proposemultiple imputation models that capture both cross-sectional information andtemporal correlations. We integrate Gaussian processes with mixture models andintroduce individualized mixing weights to handle the variance of predictiveconfidence of Gaussian process models. The proposed models are compared withseveral state-of-the-art imputation algorithms on both real-world and syntheticdatasets. Experiments show that our best model can provide more accurateimputation than the benchmarks on all of our datasets.