Most machine learning models expect to receive tables with samples in rows and attributes in columns as input. In reality, the vast majority of structured data is stored in relational databases, where information is scattered across multiple tables. For example, a cancer patient can have information scattered in radiotherapy, chemotherapy and surgery tables depending on his treatment. This data representation, although natural, creates variable-sized inputs, with heterogeneous subsets of attributes. This poses difficulties for machine learning, as a majority of models require fixed-size vector representations. An alternative data representation can be obtained by operating a join on the tables of interest, creating a single table with [Not Applicable] missing values. This project aims at developing appropriate and theoretically grounded neural network architectures for [Not Applicable] and relational data.