rllm.nn.models.BRIDGE

class rllm.nn.models.BRIDGE(table_encoder: TableEncoder, graph_encoder: GraphEncoder)[source]

Bases: Module

The BRIDGE model introduced in the “rLLM: Relational Table Learning with LLMs” paper. BRIDGE is a simple RTL method based on rLLM framework, which combines table neural networks (TNNs) and graph neural networks (GNNs) to deal with multi-table data and their interrelationships, and uses “foreign keys” to build relationships and analyze them to improve the performance of multi-table joint learning tasks.

Parameters:

  • table_encoder (TableEncoder) – Encoder for tabular data.

  • graph_encoder (GraphEncoder) – Encoder for graph data.

Example

>>> from rllm.nn.models.bridge import BRIDGE, TableEncoder, GraphEncoder
>>> model = BRIDGE(TableEncoder(16, 32, metadata={}), GraphEncoder(32, 8))
forward(table: TableData, non_table: Tensor, adj: Tensor | List[Tensor]) → Tensor[source]

First, the Table Neural Network (TNN) learns the tabular data. Second, the learned representations are concatenated with the non-tabular data. Third, the Graph Neural Network (GNN) processes the combined data. along with the adjacency matrix to learn the overall representation.

Parameters:

  • table (Tensor) – Input tabular data.

  • non_table (Tensor) – Input non-tabular data.

  • adj (Tensor) – Adjacency matrix.

Returns:

Output table embedding features.

Return type:

Tensor