Research map / Graph learning / Graph self-supervised learning
Adversarial Graph Fusion for Incomplete Multi-view Semi-supervised Learning with Tensorial Imputation
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 AGF-TI, an incomplete multi-view semi-supervised learning method that fuses anchor-based bipartite graphs via adversarial (max-min-max) graph fusion and imputes missing-sample structure with tensor nuclear norm low-rank learning.
Problem
Graph-based multi-view semi-supervised learning with missing views and scarce labels: missing samples break the smoothness assumption of label propagation by fragmenting clusters into sub-clusters (the Sub-Cluster Problem).
Approach
A min-max adversarial graph fusion operator learns a robust consensus anchor graph with learnable view weights and anchor permutations, while stacking view bipartite graphs into a third-order tensor under a tensor nuclear norm recovers similarities of missing samples from cross-view consistency.
Key results
- On CUB at VMR=30%, LAR=5%, AGF-TI reaches 78.33 ACC versus 68.75 for AMSC (second best among compared methods).
- On OutScene, AGF-TI's ACC drops less than 6% as VMR rises from 30% to 70%, while SLIM and AMSC drop nearly 13%.
- Removing the tensorial imputation causes an average ACC drop of up to 17.8% at VMR=70%, versus 6.2% at VMR=30%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AGF-TI instead of the strongest reported baseline | Incomplete multi-view semi-supervised learning · Six datasets, VMR=30%, 50%, and 70%; LAR=5% | ACC | Helped |
| Add tensorial imputation of missing view graphs (ablation) | AGF-TI · Six datasets, VMR=70% versus VMR=30%; LAR=5% | ACC: average performance drop of up to 17.8% when VRM is 70%, compared to a 6.2% degradation at 30% VRM | Helped |
| Add permutation matrices for anchor alignment (ablation) | AGF-TI · Caltech101-20 and OutScene; different VMRs, LAR=5% | ACC | Helped |
| Add learnable view-quality weights (ablation) | AGF-TI · Six datasets under different VMRs, LAR=5% | ACC | Helped |
Surprising findings
- Several GMvSSL methods with deep matrix factorization imputation match or beat the IMvSSL baseline AMSC on some datasets such as CUB and MNIST-USPS.
Limitations
- Bipartite graphs and anchors are manually pre-constructed rather than jointly optimized with anchor selection.
- Running time is higher than AMSC and SLIM on several datasets.
- Hyperparameters (beta, rho) need tuning and performance is sensitive to beta.
Experimental setup
Datasets: CUB UCI-Digit Caltech101-20 OutScene MNIST-USPS AwA
Benchmarks: CUB UCI-Digit Caltech101-20 OutScene MNIST-USPS AwA
Metrics: ACC PREC F1
Baselines: SLIM AMSC AMMSS AMGL MLAN AMUSE FMSSL FMSEL CFMSC MVAR ERL-MVSC
Builds on: label propagation AMSC anchor graph construction t-SVD tensor nuclear norm deep matrix factorization
Closest papers by meaning
- Deep Incomplete Multi-View Learning Network with Insufficient Label Information (AAAI 2024 · similarity 0.87)
- Attention-Induced Embedding Imputation for Incomplete Multi-View Partial Multi-Label Classification (AAAI 2024 · similarity 0.85)
- Adaptive Feature Imputation with Latent Graph for Deep Incomplete Multi-View Clustering (AAAI 2024 · similarity 0.85)
- Incomplete Multi-View Multi-label Learning via Disentangled Representation and Label Semantic Embedding (CVPR 2025 · similarity 0.85)
- EDISON: Enhanced Dictionary-Induced Tensorized Incomplete Multi-View Clustering with Gaussian Error Rank Minimization (ICML 2024 · similarity 0.85)
- Global Graph Propagation with Hierarchical Information Transfer for Incomplete Contrastive Multi-view Clustering (AAAI 2025 · similarity 0.85)
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.