tabular representation learning has recently gained a lot of attention.
However, existing approaches only learn a representation from a single table,
and thus ignore the potential to learn from the full structure of relational
databases, including neighboring tables that can contain im