Research map / Label-efficient and robust learning / Recognition and long-tail
TIME-FS: Joint Learning of Tensorial Incomplete Multi-View Unsupervised Feature Selection and Missing-View 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 TIME-FS, a one-stage multi-view unsupervised feature selection method that jointly performs missing-view imputation, feature selection, and tensor CP-based anchor graph learning for incomplete multi-view data.
Problem
Unsupervised feature selection on unlabeled multi-view data where some samples are missing in certain views; existing methods impute first and then select features, ignoring their interaction, and graph-based methods are computationally expensive.
Approach
Jointly optimize missing-view imputation, matrix-factorization feature selection (l2,1 sparsity), and a CP-decomposed tensor of view representations that yields a consistent anchor graph and view-preference weights, which guide both imputation and selection with low time complexity.
Key results
- On ORL mtv, COIL20 and Mfeat, TIME-FS gains an average of over 10% in both ACC and NMI over the best SOTA method (missing ratio 40%).
- On Caltech101 and Aloi, TIME-FS improves average ACC and NMI by more than 9% and 6% over the second-best methods.
- TIME-FS runs much faster than baselines; e.g., on Aloi it takes 64.54 s versus 73.78 s for HSL and 108.73 s for CDNMF, with time complexity O(dnc + c^2 nV).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add adaptive imputation, jointly learned with feature selection (ablation) | TIME-FS · WebKB (40% missing, 40% features) | ACC: 55.40 -> 60.85 | Helped |
| Add adaptive imputation, jointly learned with feature selection (ablation) | TIME-FS · ORL mtv (40% missing, 40% features) | ACC: 34.81 -> 45.55 | Helped |
| Add adaptive imputation, jointly learned with feature selection (ablation) | TIME-FS · COIL20 (40% missing, 40% features) | ACC: 55.39 -> 80.65 | Helped |
| Add adaptive imputation, jointly learned with feature selection (ablation) | TIME-FS · Mfeat (40% missing, 40% features) | ACC: 45.95 -> 72.18 | Helped |
| Add adaptive imputation, jointly learned with feature selection (ablation) | TIME-FS · Caltech101 (40% missing, 40% features) | ACC: 34.81 -> 62.73 | Helped |
| Add adaptive imputation, jointly learned with feature selection (ablation) | TIME-FS · Aloi (40% missing, 40% features) | ACC: 26.44 -> 37.10 | Helped |
| Add tensor-based anchor graph learning (ablation) | TIME-FS · WebKB (40% missing, 40% features) | ACC: 55.21 -> 60.85 | Helped |
| Add tensor-based anchor graph learning (ablation) | TIME-FS · ORL mtv (40% missing, 40% features) | ACC: 33.98 -> 45.55 | Helped |
| Add tensor-based anchor graph learning (ablation) | TIME-FS · COIL20 (40% missing, 40% features) | ACC: 59.80 -> 80.65 | Helped |
| Add tensor-based anchor graph learning (ablation) | TIME-FS · Mfeat (40% missing, 40% features) | ACC: 45.62 -> 72.18 | Helped |
| Add tensor-based anchor graph learning (ablation) | TIME-FS · Caltech101 (40% missing, 40% features) | ACC: 32.08 -> 62.73 | Helped |
| Add tensor-based anchor graph learning (ablation) | TIME-FS · Aloi (40% missing, 40% features) | ACC: 25.82 -> 37.10 | Helped |
Surprising findings
- HSL is among the fastest on some datasets (e.g., Aloi, Mfeat), but far slower on ORL mtv (499 s) and WebKB, indicating runtime is dataset-dependent.
Limitations
- Performance is relatively sensitive to the selected feature ratio.
- Baselines are applied after mean imputation, so comparison is with non-native incomplete-data handling.
- Evaluated only on six datasets with K-means clustering as the downstream proxy.
Experimental setup
Datasets: WebKB ORL mtv COIL20 Mfeat Caltech101 Aloi
Benchmarks: WebKB ORL mtv COIL20 Mfeat Caltech101 Aloi
Metrics: ACC NMI running time
Baselines: TLR-MUFS TRCA-CGL SDFS C2IMUFS CDMvFS JMVFG SCMvFS CDNMF HSL CFSMO AllFea
Builds on: JELSR TLR-MUFS anchor graph learning CP decomposition
Closest papers by meaning
- Low-Rank Kernel Tensor Learning for Incomplete Multi-View Clustering (AAAI 2024 · similarity 0.87)
- Adaptive Feature Imputation with Latent Graph for Deep Incomplete Multi-View Clustering (AAAI 2024 · similarity 0.87)
- KOALA: Kernel Coupling and Element Imputation Induced Multi-View Clustering (AAAI 2025 · similarity 0.86)
- EDISON: Enhanced Dictionary-Induced Tensorized Incomplete Multi-View Clustering with Gaussian Error Rank Minimization (ICML 2024 · similarity 0.86)
- Unsupervised Kernel-based Multi-view Feature Selection with Robust Self-representation and Binary Hashing (AAAI 2025 · similarity 0.85)
- Attention-Induced Embedding Imputation for Incomplete Multi-View Partial Multi-Label Classification (AAAI 2024 · similarity 0.84)
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.