Abstract
Relational multi-table data is common in domains such as e-commerce,healthcare, and scientific research, and can be naturally represented asheterogeneous temporal graphs with multi-modal node attributes. Existing graphneural networks (GNNs) rely on schema-specific feature encoders, requiringseparate modules for each node type and feature column, which hindersscalability and parameter sharing. We introduce RELATE (Relational Encoder forLatent Aggregation of Typed Entities), a schema-agnostic, plug-and-play featureencoder that can be used with any general purpose GNN. RELATE employs sharedmodality-specific encoders for categorical, numerical, textual, and temporalattributes, followed by a Perceiver-style cross-attention module thataggregates features into a fixed-size, permutation-invariant noderepresentation. We evaluate RELATE on ReLGNN and HGT in the RelBench benchmark,where it achieves performance within 3% of schema-specific encoders whilereducing parameter counts by up to 5x. This design supports varying schemas andenables multi-dataset pretraining for general-purpose GNNs, paving the waytoward foundation models for relational graph data.