Stop Engineering Features. Start Learning Graphs. - Stefan Andonov

Every ML pipeline hits the same wall: your data lives in a relational database, but your model wants a single flat table. So you join, aggregate, and hand-craft feature after feature just to recover the context — how a customer relates to their orders, how orders relate to products — that got lost the moment you flattened everything. This talk starts with the fundamentals: what graphs are and how graph neural networks learn from them. From there, we'll look at relational deep learning — an approach that applies these ideas directly to your existing database schema, learning from the relationships between tables instead of features engineered by hand. We'll cover how it works, how it stacks up against traditional pipelines, and how to try it on your own data.