Research map / Graph learning / Graph self-supervised learning
Multi-Label Node Classification with Label Influence Propagation
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.
Proposes Label Influence Propagation (LIP), which quantifies positive/negative label influences in GNN propagation and gradient-conflict transformation steps, builds a label influence graph, and reweights per-label losses by PageRank importance for multi-label node classification.
Problem
Multi-label node classification on graphs ignores the mutual positive and negative influences between labels, which current methods only model as proximity or co-occurrence.
Approach
Decompose GNN message passing into propagation (PPR-based node influence, signed by label sets) and transformation (gradient-angle-based influence), combine them into a label influence matrix, convert it to a row-softmax label graph, and weight each label's loss by its PageRank importance score.
Key results
- Under node split (6:2:2), GCN+LIP reaches 94.38 macro AUC on DBLP vs 92.14 for the best prior method VariMul, and 74.82 on OGB-p vs 71.39 for GCN+Auto.
- Under label split (6:2:2), GCN+LIP gets 90.89 macro AUC on Delve-M and 94.91 on Abnormal, top among compared methods.
- Paper reports average AUC improvements of 3.06% (node split) and 3.42% (label split) over baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Combine propagation and transformation label influences as InfMat (ablation) | GCN+LIP · DBLP and EukLoc | performance | Helped |
| Use propagation influence only instead of transformation influence only (ablation) | GCN+LIP · EukLoc | performance | Helped |
| Use transformation influence only instead of propagation influence only (ablation) | GCN+LIP · DBLP | performance | Helped |
| Apply LIP to GCN instead of using the strongest baseline, VariMul | MLNC node split 6:2:2 · BlogCat | AP: 13.74 $\pm$ 0.83 vs 12.33 $\pm$ 0.91 | Hurt |
| Apply LIP to GCN instead of using the strongest baseline | MLNC node split 6:2:2 · DBLP, OGB-p, PCG, HumLoc, and EukLoc | Macro AUC | Helped |
| Apply LIP to GCN instead of using the strongest baseline | MLNC node split 6:2:2 · DBLP, OGB-p, PCG, HumLoc, and EukLoc | AP | Helped |
Surprising findings
- Simple loss-weighting with a learned auto coefficient (GCN+Auto) underperforms the proposed influence-based weighting, which the authors attribute to difficulty of searching loss coefficients.
- H2GCN performs poorly on BlogCat, a heterophily graph, contrary to expectations, with the authors attributing this to the use of Node2Vec features and multi-label heterophily.
Limitations
- Abnormal dataset is private and not publicly available, limiting reproducibility.
- Influence in the propagation step relies on a PPR approximation justified only in expectation, with theory restricted to k-layer GCN.
- Reported gains are AUC-focused; the paper itself notes AUC can be misleading for imbalanced datasets and some metrics (e.g., AP on BlogCat) are not best.
Experimental setup
Datasets: DBLP BlogCat ogbn-proteins PCG HumLoc EukLoc Delve-M Abnormal
Benchmarks: DBLP BlogCat ogbn-proteins PCG HumLoc EukLoc Delve-M Abnormal
Metrics: Macro AUC AP Macro F1
Baselines: Node2vec+BR Node2vec+CC GCN+ML-KNN GCN+PLAIN MLGW ML-GCN LARN LANC VariMul GCN+Auto
Builds on: APPNP GCN PLAIN PageRank VariMul
Closest papers by meaning
- Combining Label Propagation and Simple Models out-performs Graph Neural Networks (ICLR 2021 · similarity 0.89)
- ML-GOOD: Towards Multi-Label Graph Out-Of-Distribution Detection (AAAI 2025 · similarity 0.88)
- NLGT: Neighborhood-based and Label-enhanced Graph Transformer Framework for Node Classification (AAAI 2025 · similarity 0.87)
- HGDL: Heterogeneous Graph Label Distribution Learning (NeurIPS 2024 · similarity 0.86)
- Addressing Multi-Label Learning with Partial Labels: From Sample Selection to Label Selection (AAAI 2025 · similarity 0.86)
- Label-free Node Classification on Graphs with Large Language Models (LLMs) (ICLR 2024 · 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.