Research map / Graph learning / Graph self-supervised learning
Incomplete and Unpaired Multi-View Graph Clustering with Cross-View Feature Fusion
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 MGCCFF, a graph-based multi-view clustering framework that handles incomplete and unpaired multi-view data by using DDC-based pseudo labels for category-level cross-view alignment, cross-view feature fusion, and autoencoder-plus-self-expression graph learning.
Problem
Clustering multi-view data that is simultaneously incomplete (missing views for some samples) and unpaired (incorrect cross-view sample alignment), where prior methods handle only one of these issues.
Approach
Derive pseudo labels per view with a deep divergence-based clustering loss, use them to establish category-level cross-view correspondences (via k-means on cluster-overlap similarity) instead of sample-level pairing, fuse view features with cluster centroids, and learn a sparse affinity graph by integrating self-expression into an autoencoder, supervised by a pseudo-label affinity matrix.
Key results
- On Handwritten with 50% complete rate and 10% paired rate, MGCCFF reaches 67.61 NMI and 58.84 ARI, versus 51.45 NMI / 45.01 ARI for the best incomplete baseline CRTC.
- On BDGP at 50% complete and 70% paired, MGCCFF reports 80.54 NMI and 69.44 ARI, above the best baseline DGPPVC at 68.67 NMI and 59.21 ARI.
- Ablation on Caltech101-20 (50% complete, 50% paired) shows the full model reaches 31.74 ACC / 30.33 NMI / 19.52 ARI, with removing any component lowering performance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MGCCFF instead of IMRL-AGI | Handwritten clustering · Handwritten, 50% complete, 70% paired | NMI: 62.59±0.17 -> 79.29±0.10 | Helped |
| Use MGCCFF instead of IMRL-AGI | Handwritten clustering · Handwritten, 70% complete, 70% paired | NMI: 66.31±0.16 -> 88.84±0.13 | Helped |
| Use MGCCFF instead of DGPPVC | BDGP clustering · BDGP, 50% complete, 70% paired | NMI: 68.67±0.18 -> 80.54±0.03 | Helped |
| Use MGCCFF instead of DGPPVC | BDGP clustering · BDGP, 70% complete, 70% paired | NMI: 73.62±0.35 -> 81.06±0.08 | Helped |
| Add Lgl and Lgc to Lpl (ablation) | MGCCFF · Caltech101-20, 50% complete, 50% paired | NMI: 28.76 -> 30.33 | Helped |
| Add Lpl and Lgc to Lgl (ablation) | MGCCFF · Caltech101-20, 50% complete, 50% paired | NMI: 19.44 -> 30.33 | Helped |
| Add Lpl and Lgl to Lgc (ablation) | MGCCFF · Caltech101-20, 50% complete, 50% paired | NMI: 28.35 -> 30.33 | Helped |
Surprising findings
- Ablation on Caltech101-20 shows that using only the graph-constraint loss (Lgc) without the others yields lower scores than the full model but some configurations score higher on individual metrics, and the paper notes NMI on some Caltech settings is not uniformly best.
Limitations
- Graph affinity matrix is dense and noisy, which the method addresses only through a pseudo-label constraint.
- Some baselines (SURE) could not be run on certain Yale and BBCSport settings, so comparisons are incomplete.
Experimental setup
Datasets: Handwritten BDGP Caltech101-20 BBCSport Yale
Benchmarks: Handwritten BDGP Caltech101-20 BBCSport Yale
Metrics: ACC NMI ARI
Baselines: iCmSC CDIMC-net COMPLETER CRTC UAGCN IMRL-AGI MvCLN MVC-UM UPMGC-SM DGPPVC SURE KMeans
Builds on: DDC (Deep Divergence-based Clustering) self-expression SURE
Closest papers by meaning
- Incomplete Contrastive Multi-View Clustering with High-Confidence Guiding (AAAI 2024 · similarity 0.91)
- Sample-Level Cross-View Similarity Learning for Incomplete Multi-View Clustering (AAAI 2024 · similarity 0.91)
- A Non-parametric Graph Clustering Framework for Multi-View Data (AAAI 2024 · similarity 0.91)
- Attribute-Missing Multi-view Graph Clustering (CVPR 2025 · similarity 0.90)
- Fast Incomplete Multi-view Clustering by Flexible Anchor Learning (ICML 2025 · similarity 0.90)
- Adaptive Feature Imputation with Latent Graph for Deep Incomplete Multi-View Clustering (AAAI 2024 · similarity 0.90)
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.