Relational Deep Learning: Graph Neural Networks on Snowflake Data

Watch the full session · 43 min

In a Data Science Connect webinar, Kumo AI co-founder Hema Raghavan showed how relational deep learning treats a Snowflake database as a graph. A predictive query language and graph neural networks replace the hand-written feature pipelines behind churn, recommendation and forecasting models.

Most enterprise prediction problems, from churn to lifetime value to fraud, run on relational data stored in warehouses like Snowflake. In a Data Science Connect expert webinar brought to you by Kumo AI and Snowflake, Hema Raghavan, Vice President of Engineering and Co-Founder of Kumo AI, argued that data scientists should stop hand-building features from that data and let graph neural networks learn them instead. Data Science Connect co-founder and president Brian Mink hosted.

Raghavan, who previously worked on AI at LinkedIn, IBM and Yahoo!, called the approach relational deep learning (RDL) and demonstrated it running inside Snowflake.

Why is feature engineering the bottleneck in predictive AI?

Raghavan traced the problem to her own career. Early question-answering systems relied on years of handcrafted features, until transformers learned those patterns directly from raw text. Predictive machine learning on company data never had that leap, she said.

At LinkedIn, her team spent most of its time on data prep, feature engineering and model development, and building a model took about a quarter. Pipelines were also fragile. When a front-end engineer turned off a single tracking event, it broke several models, and "it took us three months to debug that," she said.

Much of that work is what she called "time travel SQL": code that builds labels, such as whether a customer churned in a given month, and then lines up features to the right time window without leaking future information. "Why did we all get into data science? Was it to write the tons and tons of SQL that does time travel and training data prep?" she asked.

What is relational deep learning?

Relational deep learning treats a database as a graph, Raghavan said. Each table is a node type, each row is a node, each primary key and foreign key relationship forms an edge, columns become node attributes, and time is "a first class citizen" on each edge.

A graph neural network can then learn patterns a data scientist would otherwise engineer by hand. Using a three-table retail example drawn from an H&M dataset of customers, articles and transactions, she said the model can pick up on signals such as a customer's affinity for red items or the products bought by people with similar purchase histories.

She compared it to computer vision, where models moved from hand-engineered features to learned embeddings. "Graph neighborhood is what defines what a node or an edge is about," she said. In her view, most machine learning problems reduce to classification or regression on either nodes or links in that graph.

How does a predictive query language replace training pipelines?

The second idea behind Kumo is that "all predictive AI tasks can be described in a domain specific language," Raghavan said. Kumo's predictive query language looks like SQL with a predict clause. Users name the target, such as the sum of a customer's transactions over the next 30 days, choose the entity to predict for, and add filters.

Because every classification, regression or forecasting task can be expressed as a training table with three columns, the platform generates labels automatically, she said. The graph supplies the features. The same language also covers recommendation and ranking problems.

What did the Kumo demo on Snowflake show?

Raghavan installed Kumo from the Snowflake Marketplace, where it runs on provisioned compute and GPUs inside the customer's Snowflake account through Snowpark Container Services. A Snowflake admin grants Kumo access to specific databases, schemas and views, which she said works well with security teams.

In the demo she connected tables, marked primary keys and date columns, let the platform infer the links between tables, and typed a churn query. The platform recognized it as a binary classification task, and the model trained in about 10 minutes. Results included standard training and evaluation metrics plus explainability showing which columns in which tables mattered most. Some users find data quality bugs that way, she said.

Users can also bring in embeddings generated elsewhere, including from Snowflake Cortex, and work through a notebook interface instead of the UI.

How well does graph learning scale in the enterprise?

Raghavan said the platform scales to about 15 billion rows, 2,000 columns and two terabytes of data per table, and currently up to 12 tables, which has been enough for most customers. Building graph neural networks in-house is expensive. She estimated it takes $20 million or more and noted LinkedIn spent about seven years on its work.

She shared two examples. In a five-day internal hackathon on an anime recommendation dataset, 12 participants who finished wrote 353 predictive queries, triggering more than 2,000 experiments. And an auto insurance customer predicting bodily injury trained 62 predictive queries in eight days, with a 51% improvement over its in-house model. Each model would have taken the customer 46 days to build internally, she said.

Asked whether teams need a full data science staff, she said no. "Kumo doesn't need you to have an advanced DS degree," she said, adding that the platform serves engineers, data scientists and business analysts. Some customers also feed Kumo's node embeddings into existing XGBoost models.

Learn more in Data Science Connect webinars

Data Science Connect hosts 24 free live webinars on its 2027 schedule across four series: the AI × Enterprise Panel Series, AI Demo Days, Agentic AI Showcases and the Plug & Play Series. Sessions like this one give data science teams a look at new modeling approaches and the platforms behind them. See the full lineup on the webinars page.

Watch: Scalable Graph Learning for your Enterprise on YouTube
Event: Webinars · More coverage: Webinars insights