Research map / Graph learning / Graph self-supervised learning
Revisiting Self-Supervised Heterogeneous Graph Learning from Spectral Clustering Perspective
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.
Revisits self-supervised heterogeneous graph learning as spectral clustering and proposes SCHOOL, which adds a rank-constrained affinity matrix and node- and cluster-level consistency losses.
Problem
Self-supervised heterogeneous graph learning (SHGL) message passing introduces inter-class noise into graph structures, and existing methods underuse cluster-level information, hurting downstream performance.
Approach
Theoretically shows that meta-path-based and adaptive-graph SHGL methods reduce to spectral clustering with regularization, then learns a rank-constrained affinity matrix (exactly c connected components) via a spectral loss on a projected orthogonal cluster assignment, combined with node-level and cluster-level consistency constraints.
Key results
- On heterogeneous graph node classification, SCHOOL reaches 92.7 Macro-F1 on ACM, 93.0 on Yelp, 94.0 on DBLP and 77.5 on Aminer, outperforming HERO.
- Reports an average 1.1% node classification gain over the best SHGL baseline (HERO) across four heterogeneous datasets.
- Reports an average 3.1% node clustering gain over the best SHGL baseline (HGMAE) across four heterogeneous datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SCHOOL instead of HERO | node classification · ACM, Yelp, DBLP, and Aminer | Macro-F1 and Micro-F1: 92.2$\pm$0.5 -> 92.7$\pm$0.6; 92.1$\pm$0.7 -> 92.6$\pm$0.5; 92.4$\pm$0.7 -> 93.0$\pm$0.7; 92.3$\pm$0.6 -> 92.8$\pm$0.4; | Helped |
| add spectral loss $\mathcal{L}_{sp}$ (ablation) | SCHOOL node classification · ACM | Macro-F1: 86.9$\pm$0.7 -> 92.7$\pm$0.6 | Helped |
| add node-level consistency loss $\mathcal{L}_{nc}$ (ablation) | SCHOOL node classification · ACM | Macro-F1: 89.0$\pm$0.5 -> 92.7$\pm$0.6 | Helped |
| add cluster-level consistency loss $\mathcal{L}_{cc}$ (ablation) | SCHOOL node classification · ACM | Macro-F1: 88.9$\pm$0.7 -> 92.7$\pm$0.6 | Helped |
| add all three objective components rather than the best variant (ablation) | SCHOOL node classification · four heterogeneous graph datasets | Macro-F1 and Micro-F1: 1.8% | Helped |
Limitations
- Relies on node features; heterogeneous graphs with missing node features are not handled.
- Theoretical analysis assumes orthogonal learned representations (H^T H = I).
Experimental setup
Datasets: ACM DBLP Aminer Yelp Photo Computers
Benchmarks: ACM Yelp DBLP Aminer Photo Computers
Metrics: Macro-F1 Micro-F1 clustering metrics
Baselines: DeepWalk GCN GAT Mp2vec HAN HGT DMGI DMGIattn HDMI HeCo HGCML CPIM
Builds on: HERO DMGI HeCo spectral clustering Info-NCE
Closest papers by meaning
- Self-Supervised Heterogeneous Graph Learning: a Homophily and Heterogeneity View (ICLR 2024 · similarity 0.91)
- Self-supervised Heterogeneous Graph Pre-training Based on Structural Clustering (NeurIPS 2022 · similarity 0.90)
- Every Node Is Different: Dynamically Fusing Self-Supervised Tasks for Attributed Graph Clustering (AAAI 2024 · similarity 0.89)
- Decoupled Self-supervised Learning for Graphs (NeurIPS 2022 · similarity 0.88)
- Pass-Efficient Algorithms for Graph Spectral Clustering (Student Abstract) (AAAI 2024 · similarity 0.87)
- Structure-Aware Spectral Sparsification via Uniform Edge Sampling (NeurIPS 2025 · similarity 0.87)
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.