{
  "title": "RelGNN: Composite Message Passing for Relational Deep Learning",
  "authors": [
    "Tianlang Chen",
    "Charilaos Kanatsoulis",
    "Jure Leskovec"
  ],
  "abstract": "Predictive tasks on relational databases are critical in real-world\napplications spanning e-commerce, healthcare, and social media. To address\nthese tasks effectively, Relational Deep Learning (RDL) encodes relational data\nas graphs, enabling Graph Neural Networks (GNNs) to exploit relational\nstructures for improved predictions. However, existing heterogeneous GNNs often\noverlook the intrinsic structural properties of relational databases, leading\nto modeling inefficiencies. Here we introduce RelGNN, a novel GNN framework\nspecifically designed to capture the unique characteristics of relational\ndatabases. At the core of our approach is the introduction of atomic routes,\nwhich are sequences of nodes forming high-order tripartite structures. Building\nupon these atomic routes, RelGNN designs new composite message passing\nmechanisms between heterogeneous nodes, allowing direct single-hop interactions\nbetween them. This approach avoids redundant aggregations and mitigates\ninformation entanglement, ultimately leading to more efficient and accurate\npredictive modeling. RelGNN is evaluated on 30 diverse real-world tasks from\nRelBench (Fey et al., 2024), and consistently achieves state-of-the-art\naccuracy with up to 25% improvement.",
  "pdf_url": "http://arxiv.org/pdf/2502.06784v1",
  "entry_id": "http://arxiv.org/abs/2502.06784v1",
  "categories": [
    "cs.LG",
    "cs.AI",
    "cs.DB"
  ]
}