Projects

Much of the world’s structured data lives in relational databases. In relational deep learning, a database is treated as a graph, with each row as a node and each foreign-key link as an edge, and a model learns directly from that graph instead of from hand-engineered features. Relational foundation models go a step further: they are pretrained once on many databases and then applied to new databases and tasks without retraining.

Our work in this area focuses on efficiency. Training or running these models means sampling a neighborhood of rows around every row of interest, and the cost grows with the size of that sample. We use SQL and the database schema to decide which parts of the neighborhood a model needs to see: pushing sampling into the database engine, pruning foreign-key paths that carry little signal, and focusing a foundation model’s context on the tables that matter for the task.

Knowledge graphs describe entities and the relations between them, and graph neural networks (GNNs) are a standard way to learn from them. Training a GNN on a large knowledge graph is expensive, because each node learns about the structure of the graph only through repeated rounds of message passing over sampled neighborhoods, which loses information along the way.

We study how to give GNNs that structural information up front. Our method, gHAWK, precomputes a compact summary of each node’s local neighborhood and of its position in the global graph before training starts. Standard GNNs can then train faster and use less memory while becoming more accurate, and gHAWK reports state-of-the-art results on benchmarks from the Open Graph Benchmark.

LLM-based agents increasingly need to read from and reason over data systems. Giving a model raw access to a data source is not enough: it can write queries that run without error but return wrong answers. We study how to design the interface between an agent and a data source so that the agent’s answers are grounded in the data.

Our GraSP-MCP connector lets an LLM query knowledge graphs such as Wikidata through SPARQL, using the Model Context Protocol. Our experiments show that the design of the tools the connector exposes matters more than the access itself. Breaking query construction into steps that ground entities and properties against the graph, and validating queries before running them, substantially improves answer quality compared with a connector that only forwards queries.