---
title: "MetaSieve: Faster Relational Deep Learning through SQL-Based Metapath Selection"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.25903v1"
permalink: "https://feed.astralor.com/items/itm_39ea06aff2748c98b58123db"
published: "2026-08-26T15:22:52.000Z"
score: 71
channel: "ai"
tags: ["MetaSieve","RelBench","SQL","Relational Deep Learning","Graph Neural Network"]
---

# MetaSieve: Faster Relational Deep Learning through SQL-Based Metapath Selection

## AI commentary

> Training GNNs on relational databases is often slowed by subgraph sampling, and MetaSieve's decision to hand metapath selection to SQL statistics is refreshingly practical. It is worth reading for researchers at the intersection of databases and graph learning, and for algorithm teams working with multi-table data. Because it leaves GNN parameters untouched, it can be dropped into existing pipelines at low cost.

## At a glance

Relational Deep Learning (RDL) models multi-table databases as graphs and trains graph neural networks, yet subgraph sampling costs are high. MetaSieve leverages SQL join and aggregation capabilities to compute statistics for each candidate metapath extension and prunes uninformative metapaths using a novel scoring function. Since it is independent of GNN parameters, it integrates with various GNN backbones for classification and regression. On the RelBench benchmark, MetaSieve consistently reduces per-epoch training time by large margins while maintaining or often improving accuracy. This makes relational deep learning more practical for real-world multi-table data.

## Key points

- MetaSieve uses SQL joins and aggregation queries to score metapath usefulness and prunes uninformative paths to reduce subgraph sizes.
- Metapath selection relies only on database statistics and task labels, independent of GNN parameters, easing integration across GNN backbones.
- On RelBench, MetaSieve dramatically reduces per-epoch training time while maintaining or often improving accuracy.
- The method offers an efficient and general acceleration route for GNN training over multi-table relational databases.

## Full text

Relational Deep Learning (RDL) is an effective approach to machine learning over multi-table relational databases. In RDL, a database is modeled as a graph in which each row is a node and each foreign-key relation is an edge, and a graph neural network (GNN) is trained on this graph. Training a GNN requires sampling a subgraph around every seed node in the training set, and the cost of training is largely determined by the size of these subgraphs. This paper aims to reduce subgraph size by leveraging the join and aggregation capabilities of relational database systems. We observe that sampled subgraphs are obtained by following metapaths composed of foreign-key links, and that many of these metapaths can be pruned without loss of accuracy. We present MetaSieve, a metapath selection layer that determines which metapaths to retain and which to prune. For each candidate metapath extension, MetaSieve computes statistics via SQL join and aggregation queries and evaluates the extension based on a novel scoring function that prefers lightweight but informative candidates. Metapaths whose scores fall below a threshold are deemed uninformative and pruned. Metapath selection in MetaSieve is lightweight since it relies only on database statistics and task labels, and it is independent of GNN parameters, so it integrates with diverse GNN architectures for classification and regression. Our evaluation on the RelBench benchmark with multiple GNN backbones shows that MetaSieve consistently reduces per-epoch training time by large margins while maintaining and often improving accuracy.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_39ea06aff2748c98b58123db