Exploring Relational Database Foundation Models from a Graph Perspective

The transformative impact of foundation models has yet to be fully realized for Relational Databases (RDBs), largely due to their complex structures and diverse task requirements. We introduce Griffin, an initial attempt at a graph-centric foundation model for RDBs, which unifies data encoding and task decoding to handle heterogeneous data and a wide array of tasks. Griffin employs pretrained text and float encoders for categorical, numerical, and metadata features. Its novel architecture incorporates a cross-attention module for intra-row feature interaction and an enhanced message-passing neural network (MPNN) to capture inter-table relational complexities. Pretrained on extensive single-table and RDB datasets (over 150 million nodes), Griffin demonstrates superior or comparable performance to task-specific models. Importantly, it shows strong transferability to new datasets and tasks, particularly in low-data scenarios when pretraining data exhibits similarity or diversity to the target, underscoring its potential as a universally applicable foundation model for RDBs.