Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper re-evaluates simple baselines for relational deep learning by adapting fixed low-pass feature propagation to relational entity graphs and training an MLP, finding that neighborhood-filtered features significantly improve performance over target-entity features alone. Predictions are further refined using Correct-and-Smooth, a graph-based label propagation algorithm that leverages historical labels via task-specific meta-graphs. The study demonstrates that these simpler methods recover a substantial portion of heterogeneous GNN performance, highlighting the importance of meta-graph construction for label propagation.
Simple feature propagation and label smoothing on relational graphs match much of the performance of complex GNNs, suggesting architectural complexity might be over-emphasized.
Relational 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. However, the effectiveness of simpler relational baselines remains underexplored, making it difficult to assess how much architectural complexity is needed to achieve strong predictive performance. Motivated by the observation that predictive signals arise from target attributes, relational attributes, and historical labels, we revisit how effectively these signals can be exploited using simple propagation methods. We adapt fixed low-pass feature propagation to relational entity graphs and train an MLP for prediction. The resulting predictions are further refined using Correct-and-Smooth, a graph-based label propagation algorithm that incorporates historical labels over task-specific meta-graphs connecting target entities. Across classification and regression tasks on RelBench, an MLP trained on neighborhood-filtered features outperforms one using only target-entity features on most tasks and recovers a substantial portion of the performance of heterogeneous GNNs. Correct-and-Smooth provides further gains from historical labels, although applying it requires a separately constructed meta-graph because connections among target entities are not directly specified by the relational entity graph. Together, our studies revisit the effectiveness of established propagation methods as simple RDL baselines while identifying canonical meta-graph construction as a key challenge for effective label propagation.