Revisiting Simple Baselines for Relational Deep Learning
Simple feature and label propagation recover much heterogeneous GNN performance on RelBench tasks.
The paper revisits simple baselines for relational deep learning on databases represented as temporal heterogeneous graphs. Fixed low-pass feature propagation plus an MLP outperforms target-entity features alone on most RelBench classification and regression tasks and recovers a substantial share of heterogeneous GNN performance. Correct-and-Smooth adds further gains from historical labels, but it requires a task-specific meta-graph because target-entity connections are not specified by the relational entity graph.
- Predictive signals come from target attributes, relations, and historical labels.
- An MLP on propagated neighborhood features beats target-only features on most RelBench tasks.
- Correct-and-Smooth needs a separately constructed meta-graph among target entities.
- Products
- RelBench
From the original article
1 min read at arxiv.orgRelational deep learning has emerged as a powerful paradigm for learning end-to-end predictive models over relational databases by representing them as temporal heterogeneous graphs. Recent advances based on graph neural networks and Transformers have achieved strong performance across diverse domains.
By Viet The Nguyen · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.