Relational deep learning is an emerging approach that represents relational databases as temporal heterogeneous graphs, enabling end-to-end predictive modeling. Recent work has leaned heavily on graph neural networks (GNNs) to handle these structures.
A new arXiv preprint, "Revisiting Simple Baselines for Relational Deep Learning," takes a step back. According to the abstract, the authors revisit simple baselines in this context, likely comparing them against the more complex GNN-based methods that have become standard.
The abstract is brief, but the title and framing suggest that the paper aims to test whether the complexity of GNNs is always justified. If simple baselines perform competitively, it could have implications for model selection and computational efficiency in relational learning tasks.