Simple baselines match complex models in relational deep learning

Researchers have revisited simple baseline methods for relational deep learning, which represents relational databases as temporal heterogeneous graphs. Their study focused on exploiting target attributes, relational attributes, and historical labels using basic propagation techniques. By adapting fixed low-pass feature propagation and training a multilayer perceptron (MLP), they found that this approach outperformed models using only target-entity features on most tasks and achieved a significant portion of the performance of more complex graph neural networks and Transformers. Further refinements were made using Correct-and-Smooth, a label propagation algorithm that incorporates historical labels over meta-graphs, though constructing these meta-graphs remains a key challenge. AI

IMPACT This research suggests that simpler models can achieve comparable performance to complex ones in relational deep learning, potentially reducing computational costs and complexity for certain tasks.

RANK_REASON Academic paper detailing new research findings. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• Correct-and-Smooth
• graph neural networks
• Hugging Face
• multilayer perceptron
• RelBench
• Transformers

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →