Research map / Graph learning / Graph self-supervised learning
Homophily-Related: Adaptive Hybrid Graph Filter for Multi-View Graph Clustering
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 AHGFC, a multi-view graph clustering method that builds a graph joint aggregation matrix and an adaptive hybrid low/high-pass graph filter whose weight is set by a homophily ratio, evaluated on homophilous and heterophilous graphs.
Problem
Multi-view graph clustering methods based on GNNs act as low-pass filters and perform poorly on heterophilous graphs, where high-frequency information matters; the graph type is unknown so fixed filter mixing is suboptimal.
Approach
Builds a graph joint aggregation matrix from autoencoded features and adjacency (Z_a Z_x^T and its square) to separate low/high frequencies more clearly than A, then applies an adaptive hybrid filter weighted by a pseudo-label-based homophily ratio hr, with weighted view fusion into a consensus embedding.
Key results
- On ACM, AHGFC reaches NMI 90.1 and ARI 91.4 versus DuaLGR's 72.5 and 78.8.
- On Texas, AHGFC improves ACC by 14.9% and ARI by 21.2% over the best compared baseline (reported in text).
- On Wiki-cooc, AHGFC reports ACC 77.3 vs DuaLGR 32.2 in the table; text claims ACC improvement of 5.6% over best baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use AHGFC instead of the strongest listed baseline | multi-view graph clustering · Texas | ACC: 55.4 \pm 2.1 -> 70.3 \pm 1.0 | Helped |
| use AHGFC instead of the strongest listed baseline | multi-view graph clustering · Chameleon | ACC: 42.2 \pm 0.2 -> 43.2 \pm 0.3 | Helped |
| use AHGFC instead of the strongest listed baseline | multi-view graph clustering · ACM | ACC: 92.4 \pm 0.2 -> 93.0 \pm 0.5 | Helped |
| use AHGFC instead of the strongest listed baseline | graph clustering · Wiki-cooc | ACC: 71.7 \pm 0.4 -> 77.3 \pm 0.4 | Helped |
| use AHGFC instead of the strongest listed baseline | graph clustering · Minesweeper | ACC: 69.7 \pm 3.0 -> 80.0 \pm 0.0 | Helped |
| use AHGFC instead of the strongest listed baseline | graph clustering · Workers | ACC: 78.0 \pm 0.0 -> 78.3 \pm 0.1 | No clear effect |
| add reconstruction loss (ablation) | AHGFC · Chameleon | ACC: 41.5 -> 43.2 | Helped |
| add reconstruction loss (ablation) | AHGFC · Minesweeper | ACC: 63.7 -> 80.0 | Helped |
| add KL loss (ablation) | AHGFC · Chameleon | ACC: 43.2 -> 43.2 | No clear effect |
| add KL loss (ablation) | AHGFC · Minesweeper | ACC: 80.0 -> 80.0 | No clear effect |
| use graph joint aggregation matrix S instead of adjacency matrix A (ablation) | AHGFC · Chameleon | ACC: 35.0 -> 43.2 | Helped |
| use graph joint aggregation matrix S instead of adjacency matrix A (ablation) | AHGFC · Minesweeper | ACC: 69.4 -> 80.0 | Helped |
Surprising findings
- On Minesweeper, AHGFC's ACC of 80.0 is reached while NMI and ARI are near zero, and VGAE gets the best NMI and ARI there.
- The paper reports that the KL clustering loss has almost no impact on performance, while the reconstruction loss dominates.
Limitations
- Results on some datasets (e.g., Minesweeper NMI/ARI, Wiki-cooc F1) are weaker than baselines, and standard deviations are large on some datasets.
- The homophily ratio relies on pseudo-labels from the current consensus embedding, which may be noisy early in training.
Experimental setup
Datasets: Texas Chameleon ACM Wiki-cooc Minesweeper Workers
Benchmarks: Texas Chameleon ACM Wiki-cooc Minesweeper Workers
Metrics: NMI ARI ACC F1
Baselines: VGAE O2MAC MvAGC MCGC DuaLGR
Builds on: VGAE MCGC MvAGC O2MAC
Closest papers by meaning
- HeroFilter: Adaptive Spectral Graph Filter for Varying Heterophilic Relations (NeurIPS 2025 · similarity 0.91)
- PC-Conv: Unifying Homophily and Heterophily with Two-Fold Filtering (AAAI 2024 · similarity 0.90)
- What Is Missing For Graph Homophily? Disentangling Graph Homophily For Graph Neural Networks (NeurIPS 2024 · similarity 0.89)
- SURER: Structure-Adaptive Unified Graph Neural Network for Multi-View Clustering (AAAI 2024 · similarity 0.88)
- Unifying Homophily and Heterophily for Spectral Graph Neural Networks via Triple Filter Ensembles (NeurIPS 2024 · similarity 0.88)
- Multi-View Graph Clustering via Node-Guided Contrastive Encoding (ICML 2025 · 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.