Most actionable data in enterprises are often relational data stored in tables connected by primary and foreign keys. Such data usually resides in transactional databases that run the production applications of the enterprise as well as in data lakes and data warehouses for analytics queries. Learning from such data is difficult due to the impedance mismatch between the relational schema and the training set, which is a single table with features, weights, and labels. The usual approach requires joining the normalized data and producing features for training and inference to produce the latter from the former. This is where most of the data science efforts for AI/ML efforts go today.

In this talk, I will present a representation learning approach for relational data that uses graph neural networks to learn directly from the raw relational data. Relational data comprises a graph based on the primary-foreign key relationships; thus, a graph view is more natural and simplifies the learning and the required infrastructure. I will present the benefits of this approach as implemented by kumo.ai to be enterprise-ready with reliability, scalability, and model performance.