Stable GNN Embeddings for Relational Data

Alan Gany, Paul Landrier, Bogdan Cautis, Laks V. S. Lakshmanan, Silviu Maniu

Keywords: GNN, stability, database embeddings

Abstract:

Graph neural networks (GNNs) are a valuable tool for extracting meaningful representations from graph-structured data. Graphs, like relational databases, represent relationships between entities. Recent research has explored the potential of using GNNs for downstream tasks on relational data, such as entity resolution and missing value imputation. However, applying GNNs to relational databases presents two challenges. The first challenge is data conversion: relational databases, organized as tables connected by key / foreign key constraints,
must be transformed into graphs without losing essential information. The second challenge is ensuring that the embedding technique can adapt to the dynamic nature of databases. When a database is updated, the embeddings of the resulting database should be recomputable efficiently. This requires that previously computed embeddings remain stable despite changes to the data.

Motivated by using GNNs for relational databases, we study stability, i.e., how much the embeddings generated by a GNN change when the input graph undergoes modifications. Building upon the work of Gama et al. (2020), which established a limit for the distance between embeddings of similar graphs, we focus on node-level stability for GNN embeddings, particularly when the graphs originate from relations. We propose several techniques for transforming relational databases into graphs. To assess the effectiveness of these methods, we conduct experiments using the TPC-E database benchmark and analyze their stability.

Primary Area: learning on graphs and other geometries & topologies

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 11620