Research map / Graph learning / Graph self-supervised learning
AutoG: Towards automatic graph construction from tabular 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 eight tabular datasets for graph construction evaluation and AutoG, an LLM-based framework that generates graph schemas via close-ended function-call actions with chain-of-augmentation prompting and GNN-based oracle feedback, reaching near-expert performance.
Problem
Constructing suitable heterogeneous graphs from multi-table tabular data is manual, understudied, and lacks datasets and automatic methods; heuristic methods like Row2Node miss implicit relationships.
Approach
Use an LLM as a decision maker that picks from four close-ended schema transformation actions (connect columns, generate new table, add/remove primary key, unfold multi-category columns) guided by chain-of-thought demonstrations and column statistics, with an oracle that scores candidate schemas by averaged GNN performance.
Key results
- AutoG achieves average ranking 2.0 (RGCN), tying the expert-designed schema and beating JTD (3.1) and TabGNN (4.7) on the eight datasets.
- On RetailRocket CVR, AutoG reaches 74.78 AUC versus 50.82 for the best JTD schema and 84.70 for the expert schema.
- On Diginetica Purchase, AutoG reaches 36.91 MRR, matching the expert schema and far above the JTD schema's 15.47.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AutoG graph construction instead of the strongest listed baseline | RGCN downstream evaluation · AVS Repeat | AUC: 55.08 -> 56.03 | Helped |
| Use AutoG graph construction instead of the strongest listed baseline | RGCN downstream evaluation · RetailRocket CVR | AUC: 84.70 -> 74.78 | Hurt |
| Use AutoG graph construction instead of the strongest listed baseline | RGCN downstream evaluation · MAG Citation | MRR: 81.50 -> 80.84 | Hurt |
| Use AutoG graph construction instead of the strongest listed baseline | RGCN downstream evaluation · Diginetica CTR | AUC: 75.07 -> 75.07 | No clear effect |
| Use AutoG-A oracle selection rather than AutoG-S final state (ablation) | AutoG on MAG Year · MAG Year | accuracy: 50.99 -> 54.09 | Helped |
| Use AutoG-A oracle selection rather than AutoG-S final state (ablation) | AutoG on IEEE-CIS Fraud · IEEE-CIS Fraud | AUC: 87.28 -> 89.17 | Helped |
| Remove chain-of-thought demonstrations (ablation) | AutoG prompt generation · Synthetic MAG challenge 1, original columns | successful actions: 3/3 -> 1/3 | Hurt |
| Remove column statistics from the prompt (ablation) | AutoG prompt generation · Synthetic MAG challenge 1, original columns | successful actions: 3/3 -> 1/3 | Hurt |
| Remove in-context demonstrations (ablation) | AutoG prompt generation · Synthetic MAG challenge 1, original columns | successful actions: 3/3 -> 0/3 | Hurt |
Surprising findings
- Expert-designed graph is much worse than the original schema for MAG year prediction, with adjusted homophily of 0.02 versus 0.156 for venue prediction.
- Open-ended generation without close-ended function calling produced invalid graph schemas, and removing chain-of-thought led LLMs to only find trivial same-name relations.
Limitations
- Semantic column names are vital; performance drops when columns are anonymized.
- Graph oracle is needed when expert or generated schemas are harmful, and LLM alone cannot predict benefit of a graph.
- Currently addresses only moderately complex scenarios; some datasets (e.g., IEEE-CIS) show no construction method beats the original.
Experimental setup
Datasets: MAG Movielens AVS IEEE-CIS Outbrain Diginetica RetailRocket Stackexchange
Benchmarks: IEEE-CIS RetailRocket Movielens Outbrain AVS MAG Diginetica Stackexchange
Metrics: AUC accuracy MRR ranking
Baselines: XGBoost DeepFM TabGNN Vanilla schema JTD schema (Row2Node/Edge) Expert schema
Builds on: Row2Node DeepJoin 4DBInfer RGCN HGT
Closest papers by meaning
- TabGLM: Tabular Graph Language Model for Learning Transferable Representations Through Multi-Modal Consistency Minimization (AAAI 2025 · similarity 0.86)
- RDB2G-Bench: A Comprehensive Benchmark for Automatic Graph Modeling of Relational Databases (NeurIPS 2025 · similarity 0.86)
- Can Large Language Models Analyze Graphs like Professionals? A Benchmark, Datasets and Models (NeurIPS 2024 · similarity 0.86)
- Flatten Graphs as Sequences: Transformers are Scalable Graph Generators (NeurIPS 2025 · similarity 0.85)
- TEG-DB: A Comprehensive Dataset and Benchmark of Textual-Edge Graphs (NeurIPS 2024 · similarity 0.85)
- How Do Large Language Models Understand Graph Patterns? A Benchmark for Graph Pattern Comprehension (ICLR 2025 · similarity 0.85)
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.