Research map / Graph learning / GNN architectures
Contrastive Auxiliary Learning with Structure Transformation for 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.
Proposes CALHG, a heterogeneous GNN trained with supervised loss plus a self-supervised auxiliary loss, where views come from edge perturbation and PPR graph diffusion and the loss combines invariance-covariance and category-guided terms without negative pairs.
Problem
Heterogeneous graph neural networks focus on neighbor aggregation and semantic fusion while neglecting graph structure and learning paradigms, and perform poorly when node features are absent, relying on a single supervised objective.
Approach
Combine graph structure transformation (type-based edge dropping, k-hop edge generation, and PPR graph diffusion with top-k sparsification) with a category-guided multi-view contrastive auxiliary loss (Barlow-Twins/VICReg-style invariance-covariance plus class-distance terms) that needs no positive/negative pairs.
Key results
- On node-feature-free datasets, CALHG reports Macro-F1 of 92.50 on ACM*, 95.25 on DBLP* and 61.15 on IMDB*, outperforming Simple-HGN by 0.8, 1.08 and 16.35 points.
- CALHG reaches Macro-F1 of 94.49 on ACM, 95.17 on DBLP and 66.54 on IMDB with node features, surpassing Simple-HGN by 1.22, 1.6 and 2.57 points.
- Ablation shows removing edge perturbation, graph diffusion or category-guided contrastive learning each degrades performance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add edge perturbation (ablation) | CALHG · ACM | Macro-F1: 92.49±0.80 -> 93.74±0.27 | Helped |
| Add graph diffusion (ablation) | CALHG · ACM | Macro-F1: 92.49±0.80 -> 92.51±0.44 | No clear effect |
| Add category-guided contrastive learning (ablation) | CALHG · ACM | Macro-F1: 94.14±0.25 -> 94.49±0.18 | Helped |
| Add edge perturbation (ablation) | CALHG · DBLP | Macro-F1: 91.86±0.35 -> 94.42±0.37 | Helped |
| Add graph diffusion (ablation) | CALHG · DBLP | Macro-F1: 91.86±0.35 -> 92.12±0.37 | Helped |
| Add category-guided contrastive learning (ablation) | CALHG · DBLP | Macro-F1: 94.57±0.50 -> 95.17±0.35 | Helped |
| Add edge perturbation (ablation) | CALHG · IMDB | Macro-F1: 64.23±1.21 -> 65.21±0.86 | Helped |
| Add graph diffusion (ablation) | CALHG · IMDB | Macro-F1: 64.23±1.21 -> 66.25±0.41 | Helped |
| Add category-guided contrastive learning (ablation) | CALHG · IMDB | Macro-F1: 66.48±0.39 -> 66.54±0.40 | Helped |
| Add edge perturbation (ablation) | CALHG · ACM* | Macro-F1: 82.93±17.00 -> 89.80±0.40 | Helped |
| Add graph diffusion (ablation) | CALHG · ACM* | Macro-F1: 82.93±17.00 -> 89.37±0.61 | Helped |
| Add category-guided contrastive learning (ablation) | CALHG · ACM* | Macro-F1: 91.50±0.41 -> 92.50±0.40 | Helped |
Limitations
- Baselines are described as incomplete multi-view clustering methods, an apparent inconsistency with the node classification task.
- Hyperparameter sensitivity and the edge-perturbation threshold choice are only briefly analyzed.
Experimental setup
Datasets: ACM DBLP IMDB Freebase AMiner
Benchmarks: ACM DBLP IMDB Freebase AMiner
Metrics: Macro-F1 Micro-F1
Baselines: GCN GAT HetGNN HGT Simple-HGN RGCN RSHN HAN MAGNN SeHGNN LDMLP RpHGNN
Builds on: Barlow Twins VICReg GRACE HGT HAN
Closest papers by meaning
- Bootstrapping Heterogeneous Graph Representation Learning via Large Language Models: A Generalized Approach (AAAI 2025 · similarity 0.90)
- HGOT: Self-supervised Heterogeneous Graph Neural Network with Optimal Transport (ICML 2025 · similarity 0.89)
- HeterSEED: Semantics–Structure Decoupling for Heterogeneous Graph Learning under Heterophily (NeurIPS 2026 · similarity 0.89)
- Graph Structure Refinement with Energy-based Contrastive Learning (AAAI 2025 · similarity 0.88)
- Heterogeneous Graph Neural Network on Semantic Tree (AAAI 2025 · similarity 0.88)
- An Efficient Subgraph-Inferring Framework for Large-Scale Heterogeneous Graphs (AAAI 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.