Research map / Graph learning / GNN architectures
GNNRank: Learning Global Rankings from Pairwise Comparisons via Directed Graph Neural Networks
Yixuan He, Quan Gan, David Wipf, Gesine D Reinert, Junchi Yan, Mihai Cucuringu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GNNRank-P instead of the strongest listed baseline | global ranking from pairwise comparisons · Animal | upset simple: 0.33±0.00 -> 0.25±0.00 | Improved |
| use GNNRank-P instead of the strongest listed baseline | global ranking from pairwise comparisons · Basketball (avg) | upset simple: 0.78±0.00 -> 0.73±0.00 | Improved |
| use GNNRank-P instead of the strongest listed baseline | global ranking from pairwise comparisons · Football finer (avg) | upset simple: 0.93±0.00 -> 0.82±0.01 | Improved |
| use GNNRank-P instead of the strongest listed baseline | ERO synthetic ranking · ERO(p=0.05, style=gamma, η=0.2) | Kendall tau: 0.74±0.00 -> 0.77±0.00 | Improved |
| use proximal ranking with learned similarity and Fiedler steps | GNNRank · ERO(p=1, style=uniform, η=0.5) | Kendall tau: 0.92±0.00 vs 0.92±0.00 | No clear effect |
| add margin loss to ratio loss (ablation) | GNNRank-P · real-world ranking datasets | upset simple | Improved |
| use margin loss for non-proximal GNNRank (ablation) | GNNRank-N · real-world ranking datasets | upset simple | Hurt |
| remove pretraining (ablation) | GNNRank-P · Faculty: CS | upset simple: 0.32±0.00 -> 0.33±0.00 | Hurt |
| make proximal step sizes trainable (ablation) | GNNRank-P · real-world ranking datasets | upset simple | Improved |
| reduce proximal steps from 5 to 3 (ablation) | GNNRank-P · Football (avg) | upset simple: 0.78±0.02 -> 0.77±0.01 | Improved |
| increase proximal steps from 5 to 7 (ablation) | GNNRank-P · Football (avg) | upset simple: 0.78±0.02 -> 0.78±0.02 | No clear effect |
| replace directed GNN embeddings with a two-layer MLP (ablation) | GNNRank · real-world datasets | upset simple: 2% worse | Hurt |
Closest papers by meaning
- Dynamic Ranking with the BTL Model: A Nearest Neighbor based Rank Centrality Method (JMLR 2023 · similarity 0.87)
- Learning to Rank: How GNNs Solve Max-Clique and Sparse PCA (AAAI 2026 · similarity 0.86)
- Diffusing to the Top: Boost Graph Neural Networks with Minimal Hyperparameter Tuning (ICLR 2025 · similarity 0.83)
- KAA: Kolmogorov-Arnold Attention for Enhancing Attentive Graph Neural Networks (ICLR 2025 · similarity 0.83)
- Learning to Reweight for Generalizable Graph Neural Network (AAAI 2024 · similarity 0.83)
- On the Interplay between Graph Structure and Learning Algorithms in Graph Neural Networks (ICML 2025 · similarity 0.83)
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.
