Research map / Graph learning / Graph self-supervised learning
Adaptive Feature Imputation with Latent Graph for Deep Incomplete Multi-View 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 AGDIMC, a deep incomplete multi-view clustering method that imputes missing embedded features adaptively using cross-view soft cluster assignments and latent kNN graphs with GCN refinement and a nearest-neighbor graph constraint.
Problem
Incomplete multi-view clustering, where views are missing for some samples; existing methods impute missing data inaccurately and ignore latent graph structure within views.
Approach
Impute missing embedded features from global cluster centroids and cross-view soft assignments, accept imputation only when it yields more consistent cluster assignments than unimputed features, and refine embeddings with per-view latent kNN graphs fused into a global graph via a GCN, plus a nearest-neighbor graph constraint.
Key results
- On BDGP at missing rate 0.1, AGDIMC reaches ACC 0.979, NMI 0.937, ARI 0.949, outperforming the compared IMVC baselines.
- On Reuters at missing rate 0.1, AGDIMC achieves ACC 0.588 versus 0.575 for the next best (PGP).
- Ablation on three datasets shows the full model (all three components) gives the best ACC/NMI/ARI across BDGP, Reuters and HW.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add local graph convolution and global graph fusion (ablation) | AGDIMC · BDGP | ACC: 0.932 -> 0.979 | Helped |
| Add local graph convolution and global graph fusion (ablation) | AGDIMC · Reuters | ACC: 0.527 -> 0.588 | Helped |
| Add local graph convolution and global graph fusion (ablation) | AGDIMC · HW | ACC: 0.907 -> 0.973 | Helped |
| Add adaptive imputation module (ablation) | AGDIMC · BDGP | ACC: 0.789 -> 0.979 | Helped |
| Add adaptive imputation module (ablation) | AGDIMC · Reuters | ACC: 0.502 -> 0.588 | Helped |
| Add adaptive imputation module (ablation) | AGDIMC · HW | ACC: 0.880 -> 0.973 | Helped |
| Add nearest neighbor graph constraints (ablation) | AGDIMC · BDGP | ACC: 0.913 -> 0.979 | Helped |
| Add nearest neighbor graph constraints (ablation) | AGDIMC · Reuters | ACC: 0.462 -> 0.588 | Helped |
| Add nearest neighbor graph constraints (ablation) | AGDIMC · HW | ACC: 0.938 -> 0.973 | Helped |
Limitations
- Performance declines as missing rate rises; clustering structure of embeddings becomes less apparent at high missing rates.
- Imputation quality is hard to assess without ground truth for missing data; the method relies on a consistency check between imputed and unimputed assignments.
- Reuters results remain low in absolute terms (ACC around 0.5).
Experimental setup
Datasets: BDGP Handwritten Numerals Reuters
Benchmarks: BDGP Handwritten Numerals Reuters
Metrics: ACC NMI ARI
Baselines: GIMC-FLSD CDIMC-net DCP HCP-IMSC IMVC-CBG DSIMVC LSIMVC PGP
Builds on: deep embedded clustering (DEC) Student's t-distribution soft assignment k-nearest neighbor graph
Closest papers by meaning
- Deep Variational Incomplete Multi-View Clustering: Exploring Shared Clustering Structures (AAAI 2024 · similarity 0.94)
- Incomplete Contrastive Multi-View Clustering with High-Confidence Guiding (AAAI 2024 · similarity 0.93)
- Attribute-Missing Multi-view Graph Clustering (CVPR 2025 · similarity 0.92)
- Incomplete Multi-view Clustering via Diffusion Contrastive Generation (AAAI 2025 · similarity 0.92)
- KOALA: Kernel Coupling and Element Imputation Induced Multi-View Clustering (AAAI 2025 · similarity 0.91)
- Fast Incomplete Multi-view Clustering with Adaptive Similarity Completion and Reconstruction (AAAI 2025 · similarity 0.91)
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.