Research map / Graph learning / Knowledge and temporal graphs
TGB 2.0: A Benchmark for Learning on Temporal Knowledge Graphs and Heterogeneous Graphs
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 TGB 2.0, a benchmark of eight large-scale multi-relational temporal graph datasets (four temporal KG, four temporal heterogeneous graph) with a reproducible evaluation pipeline using edge-type-aware negative sampling, and benchmarks TKG and THG methods against heuristic baselines.
Problem
Evaluating future link prediction on multi-relational temporal graphs (temporal knowledge graphs and temporal heterogeneous graphs) lacks large-scale standardized datasets and reproducible, realistic evaluation protocols.
Approach
Extends the Temporal Graph Benchmark to multi-relational graphs with eight new large datasets, pre-generated negative samples, time-aware filtered MRR, and edge-type-aware negative sampling for large graphs, plus heuristic baselines (RecurrencyBaseline, EdgeBank) to expose competitive simple methods.
Key results
- On tkgl-smallpedia, CEN reaches test MRR 0.612 and TLogic reaches 0.228 on tkgl-polecat, while RecB is competitive across most TKG datasets.
- On thgl-software, STHN achieves test MRR 0.731, but requires 185 GB memory for subgraph computation and cannot scale to other datasets.
- Most methods (RE-GCN, CEN, TLogic, TGN, STHN) run out of memory or time on the large datasets, while EdgeBank is the only method scaling to tkgl-wikidata.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add edge-type information to TGN (ablation) | TGN · thgl-software | MRR: 0.324 -> 0.424 | Helped |
| add edge-type information to TGN (ablation) | TGN · thgl-forum | MRR: 0.649 -> 0.729 | Helped |
| sample negatives using edge-type information rather than random sampling that omits edge types (ablation) | TGB 2.0 evaluation protocol · pedia | MRR | Helped |
Surprising findings
- Simple heuristic baselines (RecurrencyBaseline, EdgeBank) are often competitive with complex learned methods.
- EdgeBank, which ignores edge types, is the only method that scales to tkgl-wikidata.
Limitations
- Only continuous-time setting for THG datasets; discrete-time methods not evaluated on them.
- Datasets cover only five domains; biological and citation networks absent.
- Most learned methods run out of memory or time on medium and large datasets, limiting comparison.
Experimental setup
Datasets: tkgl-smallpedia tkgl-polecat tkgl-icews tkgl-wikidata thgl-software thgl-forum thgl-github thgl-myket
Benchmarks: tkgl-smallpedia tkgl-polecat tkgl-icews tkgl-wikidata thgl-software thgl-forum thgl-github thgl-myket
Metrics: MRR time-aware filtered MRR runtime GPU memory
Baselines: EdgeBank RecurrencyBaseline RE-GCN CEN TLogic TGN STHN
Builds on: Temporal Graph Benchmark (TGB) EdgeBank TGN RE-GCN TLogic
Closest papers by meaning
- TGB-Seq Benchmark: Challenging Temporal GNNs with Complex Sequential Dynamics (ICLR 2025 · similarity 0.91)
- DTGB: A Comprehensive Benchmark for Dynamic Text-Attributed Graphs (NeurIPS 2024 · similarity 0.90)
- TEILP: Time Prediction over Knowledge Graphs via Logical Reasoning (AAAI 2024 · similarity 0.89)
- TEG-DB: A Comprehensive Dataset and Benchmark of Textual-Edge Graphs (NeurIPS 2024 · similarity 0.89)
- Unifying Text Semantics and Graph Structures for Temporal Text-attributed Graphs with Large Language Models (NeurIPS 2025 · similarity 0.89)
- Can Large Language Models Analyze Graphs like Professionals? A Benchmark, Datasets and Models (NeurIPS 2024 · similarity 0.88)
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.