Research map / Learning theory and algorithms / Clustering
EDISON: Enhanced Dictionary-Induced Tensorized Incomplete Multi-View Clustering with Gaussian Error Rank Minimization
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 EDISON, a scalable tensorized incomplete multi-view clustering method that combines enhanced dictionary representation (anchor graphs with hidden-effect recovery), Gaussian Error Rank minimization, and hyper-anchor graph Laplacian regularization.
Problem
Incomplete multi-view clustering where missing views/instances degrade graph construction, tensor low-rank methods are computationally expensive on large data, and the tensor nuclear norm uniformly shrinks singular values.
Approach
Replaces observed-data dictionaries with an enhanced dictionary (LatLRR-style hidden-effect term via a dual nuclear norm plus anchor-based factorization), replaces TNN with a non-convex Gaussian Error Rank that penalizes small singular values differently, and adds a hypergraph-based anchor Laplacian regularizer inside the tensor representation.
Key results
- On BBCSport at 30% missing rate, EDISON surpasses the second-best IMSR by 12.76% ACC, 21.93% NMI and 12.76% PUR (as claimed in the paper).
- EDISON achieves 100% ACC/NMI/PUR on NGs and BBCSport at 10% missing rate.
- On ALOI-100, EDISON runs in 193.28 seconds versus 1169.70 seconds for HCLS-CGL (Table 3).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EDISON instead of IMSR | Incomplete multi-view clustering · NGs, 10% missing | ACC (%): 97.80±0.00 -> 100.00±0.00 | Helped |
| Use EDISON instead of IMSR | Incomplete multi-view clustering · BBCSport, 10% missing | ACC (%): 92.46±0.00 -> 100.00±0.00 | Helped |
| Use EDISON instead of IMSR | Incomplete multi-view clustering · MSRCV1, 10% missing | ACC (%): 87.14±0.00 -> 99.05±0.00 | Helped |
| Use EDISON instead of IMSR | Incomplete multi-view clustering · HW, 10% missing | ACC (%): 92.10±0.07 -> 99.60±0.00 | Helped |
| Use EDISON instead of IMSR | Incomplete multi-view clustering · Scene15, 10% missing | ACC (%): 37.28±0.44 -> 87.78±0.00 | Helped |
| Use EDISON instead of HCLS-CGL | Incomplete multi-view clustering · ALOI-100, 10% missing | ACC (%): 76.60±1.11 -> 84.95±0.00 | Helped |
| Add EDR (ablation) | EDISON · NGs | ACC: 0.632 -> 1.000 | Helped |
| Add EDR (ablation) | EDISON · BBCSport | ACC: 0.632 -> 1.000 | Helped |
| Add reconstruction error term (ablation) | EDISON · NGs | ACC: 0.500 -> 1.000 | Helped |
| Add reconstruction error term (ablation) | EDISON · BBCSport | ACC: 0.805 -> 1.000 | Helped |
| Add HLR (ablation) | EDISON · NGs | ACC: 0.964 -> 1.000 | Helped |
| Add HLR (ablation) | EDISON · BBCSport | ACC: 0.991 -> 1.000 | Helped |
Surprising findings
- Ablation on NGs shows removing modules can sometimes lower performance drastically, e.g., one configuration yields ACC 0.526 vs 1.000 for the full model, with the paper noting that the full model is best on most settings.
Limitations
- Method is non-convex with no global optimality guarantee; only KKT stationarity convergence is shown.
- Performance fluctuates with the GER parameter delta and trade-off parameters, requiring fine-tuning in [1e-5, 10].
Experimental setup
Datasets: NGs BBCSport HW Scene15 MSRCV1 ALOI-100
Benchmarks: NGs BBCSport HW Scene15 MSRCV1 ALOI-100
Metrics: ACC NMI PUR runtime (seconds)
Baselines: BSV Concat DAIMC OPIMC IMSR HCLS-CGL CDIMC DSIMVC
Builds on: Low-Rank Representation LatLRR Tensor Nuclear Norm TGIMC ADMM
Closest papers by meaning
- Low-Rank Kernel Tensor Learning for Incomplete Multi-View Clustering (AAAI 2024 · similarity 0.91)
- From Dictionary to Tensor: A Scalable Multi-View Subspace Clustering Framework with Triple Information Enhancement (NeurIPS 2024 · similarity 0.90)
- KOALA: Kernel Coupling and Element Imputation Induced Multi-View Clustering (AAAI 2025 · similarity 0.89)
- Performance Gaps in Multi-view Clustering under the Nested Matrix-Tensor Model (ICLR 2024 · similarity 0.89)
- Adaptive Feature Imputation with Latent Graph for Deep Incomplete Multi-View Clustering (AAAI 2024 · similarity 0.89)
- Fast Incomplete Multi-view Clustering with Adaptive Similarity Completion and Reconstruction (AAAI 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.