By representing a multi-table database as a graph of entities and relationships, a graph neural network can learn directly on raw relational data, eliminating manual feature engineering and producing double-digit accuracy gains.
The core idea of relational deep learning: think of database tables as nodes and foreign-key relations as edges, then apply graph neural networks to skip manual feature engineering and boost accuracy. ✦ AI generated
Jure Leskovec · The TWIML AI Podcast · 2026-05-21 · original ↗
starts at this moment · 14:23
Just think of the database, think of these tables as a graph of relationships between the entities in the database... now that we have a graph, we can basically apply graph deep learning like graph neural networks... The first thing that happens is you don't have to do manual feature engineering... The second thing that happens is your models are more accurate... we see this double-digit increases in model accuracy.
verbatim transcript · starts at 14:23
14:16complexity. So our question was how could just how could I just learn directly with a neural network over this multitabular data? Um and the answer is you know kind of surprisingly simple is to say just think of the database think of these tables as a graph of relationships between the entities in the database. So this would mean in some in in my you know I I'm a graph person
14:43so I like to think in terms of graphs right so graphs are composed of vertices the nodes this would be pe my users would be my um would be my products would be my transactions and so on. So this would denow the nodes and then the connections are just saying this user ID was part of this transaction that was part of that product and now we have a
15:05path from a user to the transaction to the product and then you know another user or another transaction is another path in this very simplistic graph. Um and now that we have a graph, we can basically apply graph deep learning like graph neural networks. Um which is a way to generalize deep learning to graph structured data and just train over that to get to get an accurate prediction.
15:30And what happens is two things happen. The first thing that happens is you don't have to do manual feature engineering any right. So it's much faster. It requires much less effort to train these models. And the second thing that happens is your models are more accurate. And then you say why can my models be more accurate? And the answer is very similar to what happens in computer vision. Right? If you are
15:56saying I am a human, I know what a car is. So I will I will build perfect features that detect whether there is a car on the image or not. I know cars. I drive them. I'm such a car expert. I can build the best features for detecting cars. Nobody in the right mind claims that, right? But you know in in machine learning, data science prediction, people are still saying, you know, I'm
16:21the domain expert. I'll engineer the features. Your features are just some arbitrary human biased summary statistic of your data that you know you kind of dreamt up with. Put put it as a feature in the in in your training table. retrain the model and then you saw whether that increased the accuracy or not, right? And a neural network that trains with gradient descent is able to do so much