Research map / Graph learning / Graph algorithms and clustering
Fast Incomplete Multi-view Clustering with Adaptive Similarity Completion and Reconstruction
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 ASCR, a fast anchor-based incomplete multi-view clustering method that jointly learns per-view anchor-sample similarity graphs, completes missing similarities adaptively, and factorizes them into latent embeddings with a t-SVD tensor nuclear norm for cross-view correlation.
Problem
Incomplete multi-view clustering with anchor graphs discards information about missing samples, learns only consensus graphs that miss complementary cross-view structure, and relies on post-hoc SVD that makes embedding learning suboptimal.
Approach
Learn a distinct anchor-sample similarity graph per view, fill missing similarity entries with a nonnegative completion matrix, and factorize the completed similarity matrix into anchor and sample embeddings (G_v H_v) jointly optimized with a t-SVD tensor nuclear norm over the stacked sample embeddings, removing the need for post-processing.
Key results
- On CCV, ASCR improves over the second-best method sFSR-IMVC by 34.21% in ACC on average across missing rates.
- On NGs at missing rate 0.1, ASCR reaches 100% ACC (ablation table), with average NGs ACC of 99.60%.
- ASCR runs in 0.1 s on NGs and 1.99 s on NHface, faster than most baselines such as SEC-IMVC and sFSR-IMVC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ASCR instead of DVSAI | NGs clustering · NGs | ACC: 86.37±0.16 -> 99.60±0.00 | Helped |
| Use ASCR instead of SEC-IMVC | HW clustering · HW | ACC: 93.24±0.00 -> 99.38±0.03 | Helped |
| Use ASCR instead of sFSR-IMVC | Scene15 clustering · Scene15 | ACC: 76.98±5.12 -> 86.01±3.78 | Helped |
| Use ASCR instead of LSIMVC | NHface clustering · NHface | ACC: 92.19±0.46 -> 97.48±0.01 | Helped |
| Use ASCR instead of sFSR-IMVC | CCV clustering · CCV | ACC: 30.25±1.45 -> 64.46±2.72 | Helped |
| Use ASCR instead of sFSR-IMVC | Caltech101 clustering · Caltech101 | ACC: 46.55±1.40 -> 58.76±2.15 | Helped |
| Use ASCR instead of sFSR-IMVC | SUNRGBD clustering · SUNRGBD | ACC: 22.75±1.04 -> 23.16±0.86 | Helped |
| Use ASCR instead of DVSAI | NUSWIDE clustering · NUSWIDE | ACC: 16.55±0.21 -> 16.42±0.29 | Hurt |
| Add adaptive similarity completion (ablation) | ASCR on NGs at p=0.1 · NGs, p=0.1 | ACC: 90.33 -> 100 | Helped |
| Use similarity reconstruction with latent embeddings instead of averaged-graph SVD (ablation) | ASCR on NGs at p=0.1 · NGs, p=0.1 | ACC: 98.62 -> 100 | Helped |
| Use tensor nuclear-norm regularization instead of matrix nuclear norm (ablation) | ASCR on NGs at p=0.1 · NGs, p=0.1 | ACC: 87.56 -> 100 | Helped |
Surprising findings
- The ablation variant that drops the similarity completion (ASCR-C) shows the sharpest ACC drop at high missing rates, while the variant without tensor regularization (ASCR-T) performs worst overall at all missing rates.
Limitations
- Evaluated only with a fixed anchor number k = c; sensitivity to k is only shown on two datasets.
- Hyperparameter alpha must be tuned per dataset (ranges given for NGs and NHface only).
- Experiments use a single missing-data generation protocol with random view dropping; no real-world missing patterns are tested.
Experimental setup
Datasets: NGs HW Scene15 NHface CCV Caltech101 SUNRGBD NUSWIDE
Benchmarks: NGs HW Scene15 NHface CCV Caltech101 SUNRGBD NUSWIDE
Metrics: ACC NMI PUR ARI F-score
Baselines: IMVC-CBG FIMVC-VIA SIMVC-SA DVSAI LSIMVC SEC-IMVC TDASC sFSR-IMVC
Builds on: anchor-based IMVC t-SVD tensor nuclear norm TDASC
Closest papers by meaning
- Fast Incomplete Multi-view Clustering by Flexible Anchor Learning (ICML 2025 · similarity 0.93)
- DVSAI: Diverse View-Shared Anchors Based Incomplete Multi-View Clustering (AAAI 2024 · similarity 0.91)
- Scalable One-Pass Incomplete Multi-View Clustering by Aligning Anchors (AAAI 2025 · similarity 0.91)
- Adaptive Feature Imputation with Latent Graph for Deep Incomplete Multi-View Clustering (AAAI 2024 · similarity 0.91)
- Alleviate Anchor-Shift: Explore Blind Spots with Cross-View Reconstruction for Incomplete Multi-View Clustering (NeurIPS 2024 · similarity 0.90)
- Collaborative Similarity Fusion and Consistency Recovery for Incomplete Multi-view Clustering (AAAI 2025 · 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.