Gaussian processes for missing value imputation

A missing value indicates that a particular attribute of an instance learning problem is not recorded. They are very common in many real-life datasets. In spite this, however, most machine methods cannot handle values. Thus, they should be imputed before training. Gaussian Processes (GPs) non-parametric models with accurate uncertainty estimates combined sparse approximations and stochastic variational inference scale to large data sets. Sparse GPs (SGPs) can used get predictive distribution for We present hierarchical composition predict the values at each dimension using observed from other dimensions. Importantly, we consider input attributes GP prediction may also have The those replaced by predictions previous hierarchy. call our approach (MGP). MGP impute all It outputs then imputation evaluate on one private clinical set four UCI datasets different percentage Furthermore, compare performance state-of-the-art imputing values, including variants based deep GPs. Our results show significantly better.