We present a novel framework that can combine multi-domain learning (MDL),\ndata imputation (DI) and multi-task learning (MTL) to improve performance for\nclassification and regression tasks in different domains. The core of our\nmethod is an adversarial autoencoder that can: (1) learn to produce\ndomain-invariant embeddings to reduce the difference between domains; (2) learn\nthe data distribution for each domain and correctly perform data imputation on\nmissing data. For MDL, we use the Maximum Mean Discrepancy (MMD) measure to\nalign the domain distributions. For DI, we use an adversarial approach where a\ngenerator fill in information for missing data and a discriminator tries to\ndistinguish between real and imputed values. Finally, using the universal\nfeature representation in the embeddings, we train a classifier using MTL that\ngiven input from any domain, can predict labels for all domains. We demonstrate\nthe superior performance of our approach compared to other state-of-art methods\nin three distinct settings, DG-DI in image recognition with unstructured data,\nMTL-DI in grade estimation with structured data and MDMTL-DI in a selection\nprocess using mixed data.\n