Research map / Graph learning / GNN architectures
GraphLand: Evaluating Graph Machine Learning Models on Diverse Industrial Data
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces GraphLand, a benchmark of 14 diverse industrial node-property-prediction graph datasets with random, temporal, and inductive splits, comparing GNNs, GBDTs with neighborhood feature aggregation, and graph foundation models.
Problem
Popular graph ML benchmarks for node property prediction cover narrow domains (mostly academic citation networks) and rely on text-attributed or random splits, leaving realistic industrial settings, temporal distribution shifts, and inductive evaluation under-explored; graph foundation models are al
Approach
A collection of 14 real-world industrial graph datasets with rich numerical/categorical node features, unified RL/RH/TH/THI splits (including temporal and inductive settings), and a neighborhood feature aggregation (NFA) trick that lets GBDTs use one-hop graph information, used to benchmark GNNs, GBDTs, and graph foundation models.
Key results
- LightGBM with neighborhood feature aggregation achieves the best results on four datasets, including three regression ones (e.g., city-roads-M R^2 61.00, city-roads-L 55.26, artnet-views 56.55 under RL).
- Temporal (TH) splits are significantly harder than random RH splits for nearly all models (except avazu-ctr), and inductive THI performance drops further.
- Current graph foundation models (OpenGraph, AnyGraph, TS-GNN, GCOPE) produce very weak results, e.g., OpenGraph reaches 9.49 accuracy on hm-categories under RL versus GAT's 67.96.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add one-hop neighborhood feature aggregation to graph-agnostic models (ablation) | ResMLP and LightGBM · GraphLand RL datasets | task score | Helped |
| add neighborhood feature aggregation to LightGBM (ablation) | LightGBM · city-roads-M RL | R²: 57.76 ± 0.10 -> 61.00 ± 0.05 | Helped |
| add neighborhood feature aggregation to LightGBM (ablation) | LightGBM · city-roads-L RL | R²: 52.65 ± 0.08 -> 55.26 ± 0.04 | Helped |
| add neighborhood feature aggregation to LightGBM (ablation) | LightGBM · hm-prices RL | R²: 65.44 ± 0.09 -> 70.46 ± 0.09 | Helped |
| use graph-aware GNNs instead of graph-agnostic ResMLP | GraphLand models · GraphLand RL datasets | task score | Helped |
| use attention-based GNNs rather than classic GNNs | GAT and GT versus GCN and GraphSAGE · GraphLand datasets | task score | Helped |
| use temporal rather than random high-label-rate splits | GraphLand models · GraphLand datasets except avazu-ctr | task score | Hurt |
| use inductive rather than transductive evaluation | GraphLand models · temporal high split | task score | Hurt |
| adapt with current general-purpose graph foundation models | OpenGraph, AnyGraph, TS-GNN, and GCOPE · GraphLand node classification datasets | task score | Hurt |
Surprising findings
- Graph foundation models with in-context learning perform far below simple GNNs and even graph-agnostic GBDTs on the proposed datasets.
- Neighborhood feature aggregation lets GBDTs match or beat GNNs on several datasets, especially regression tasks.
Limitations
- Many GFMs cannot run on large datasets and do not support node regression in their official implementations, limiting their evaluation coverage.
- Temporal and inductive splits are unavailable for some datasets (city-roads, city-reviews, web-traffic).
- NFA provides only one-hop neighbor statistics, limiting the graph information available to GBDTs.
Experimental setup
Datasets: web-fraud web-topics web-traffic artnet-views artnet-exp city-roads-M city-roads-L city-reviews avazu-ctr hm-categories hm-prices pokec-regions
Benchmarks: GraphLand
Metrics: accuracy average precision R^2
Baselines: ResMLP XGBoost LightGBM CatBoost GCN GraphSAGE GAT OpenGraph AnyGraph TS-GNN GCOPE
Builds on: GraphGPS XGBoost LightGBM CatBoost
Closest papers by meaning
- Beyond Real-world Benchmark Datasets: An Empirical Study of Node Classification with GNNs (NeurIPS 2022 · similarity 0.88)
- No Metric to Rule Them All: Toward Principled Evaluations of Graph-Learning Datasets (ICML 2025 · similarity 0.88)
- 4DBInfer: A 4D Benchmarking Toolbox for Graph-Centric Predictive Modeling on RDBs (NeurIPS 2024 · similarity 0.86)
- Can Graph Neural Networks Expose Training Data Properties? An Efficient Risk Assessment Approach (NeurIPS 2024 · similarity 0.86)
- RDB2G-Bench: A Comprehensive Benchmark for Automatic Graph Modeling of Relational Databases (NeurIPS 2025 · similarity 0.86)
- Equivariance Everywhere All At Once: A Recipe for Graph Foundation Models (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.