Beyond Graphs: Learning with Relational DBs

Jiaxuan You, Ge Liu

Keywords: graph machine learning, relational database, representation learning

TL;DR: This paper proposes a unified representation for diverse relational data beyond graphs using relational databases, as well as a novel deep learning architecture for learning with relational DBs.

Abstract:

Despite recent advancements in representation learning on graphs, there still lacks a unified framework that addresses the challenges of learning from real-world relational data, which can involve heterogeneous, dynamic, and multi-ary relationships. Existing efforts focus on extending graph learning to alternative graph representations, including heterogeneous graphs and hypergraphs; however, these approaches are highly specific to particular use cases, therefore introducing complexity into their application and deployment.
We propose to unify and extend existing graph learning research with relational databases (RDBs).
RDBs, characterized by their simplicity and versatility, consist of multiple tables linked by shared key columns. We show that diverse types of graphs can be unified as RDBs and different graph learning tasks can be formulated as predicting column values in RDB tables.
Furthermore, we introduce Relational Database Neural Networks (RDNNs), the first family of deep learning models that can holistically learn from multi-table information inside a relational DB, without the need for converting it to graphs. RDNNs provide a more flexible and comprehensive deep learning design space for modeling relational data, capable of solving problems beyond the scope of graph learning.
Through extensive experimentation on a range of graph and multi-table datasets, we demonstrate that the RDNNs offer competitive or superior performance in comparison to Graph Neural Networks (GNNs) on graph learning tasks and tabular machine learning methods on RDB prediction tasks.

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/2024/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: 7251