Research map / Learning theory and algorithms / Clustering
KOALA: Kernel Coupling and Element Imputation Induced 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 KOALA, a kernel-based multi-view clustering method that jointly recovers cross-view alignment (coupling matrices) and imputes missing kernel elements via a low-rank tensor constraint for uncoupled incomplete multi-view data.
Problem
Clustering multi-view data where samples are both incomplete (missing views/kernel elements) and uncoupled (cross-view sample correspondence unknown).
Approach
Introduces learnable coupling matrices in RKHS to realign uncoupled kernel matrices (cross-kernel alignment), stacks aligned kernels into an optimizable tensor with a t-SVD tensor nuclear norm to impute missing kernel entries, and jointly optimizes imputation, alignment, and k-means clustering via ADMM with curvilinear search.
Key results
- On ORL at missing ratio 0.3, KOALA improves ACC over the second-best method by 4.07%; improvements of 3.64%, 6.70%, 0.49%, 23.09%, 5.92% and 5.92% are reported across the other benchmarks at the same ratio.
- On Caltech101-7 at missing ratio 0.1, KOALA reaches Fscore 88.29 with ACC 33.95 vs. around 8 for baselines on Caltech101-20 (reported in Table 2 for the 0.1 ratio).
- Ablations show both the missing kernel element recovery and cross-kernel coupling learning improve clustering on proteinFold and Caltech101-7; the objective decreases monotonically and converges quickly.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use KOALA instead of the strongest reported baseline, MKKM+MF | Uncoupled incomplete multi-view clustering · ORL, missing ratio 0.3 | ACC (%): 21.52 (0.79) vs 16.40 (0.65) | Helped |
| Use KOALA instead of the strongest reported baseline, MKKM+MF | Uncoupled incomplete multi-view clustering · proteinFold, missing ratio 0.3 | ACC (%): 15.71 (0.49) vs 12.13 (0.64) | Helped |
| Use KOALA instead of the strongest reported baseline, MKKM-IK-MKC | Uncoupled incomplete multi-view clustering · Cora, missing ratio 0.3 | ACC (%): 50.87 (1.47) vs 44.17 (1.17) | Helped |
| Use KOALA instead of the strongest reported baseline, LRKT-IMVC | Uncoupled incomplete multi-view clustering · UCI Digits, missing ratio 0.3 | ACC (%): 13.42 (0.30) vs 12.93 (0.39) | Helped |
| Use KOALA instead of the strongest reported baseline, UIMC | Uncoupled incomplete multi-view clustering · Caltech101-7, missing ratio 0.3 | ACC (%): 50.61 (0.87) vs 19.70 (0.17) | Helped |
| Use KOALA instead of the strongest reported baseline, MKKM-IK-MKC | Uncoupled incomplete multi-view clustering · Caltech101-20, missing ratio 0.3 | ACC (%): 32.16 (0.58) vs 8.82 (0.30) | Helped |
| Use KOALA instead of the strongest reported baseline, MKKM+MF | Uncoupled incomplete multi-view clustering · Caltech101-all, missing ratio 0.3 | ACC (%): 10.02 (0.12) vs 4.10 (0.09) | Helped |
| Add the missing kernel elements recovery strategy (ablation) | KOALA · ORL and Caltech101-7 | clustering performance | Helped |
| Add cross-kernel coupling learning (ablation) | KOALA · proteinFold and Caltech101-7 | clustering performance | Helped |
Limitations
- Kernel-based only; relies on pre-computed kernels and a fixed reliable kernel (q=1 chosen by assumption).
- Convergence guarantee is only to a local optimum; sensitivity to hyperparameters lambda and theta is shown only on two datasets.
- Assumes at least one view available per sample.
Experimental setup
Datasets: ORL proteinFold Caltech101-7 UCI Digits Caltech101-20 Cora Caltech101-all
Benchmarks: ORL proteinFold Caltech101-7 UCI Digits Caltech101-20 Cora Caltech101-all
Metrics: ACC Purity ARI Fscore Precision
Baselines: MKKM+ZF MKKM+MF MKKM-IK+MF MKKM-IK+ZF MKKM-IK+KNN MKKM-IK-MKC LRKT-IMVC UMIC
Builds on: MKKM-IK UMIC multiple kernel k-means t-SVD tensor nuclear norm
Closest papers by meaning
- Low-Rank Kernel Tensor Learning for Incomplete Multi-View Clustering (AAAI 2024 · similarity 0.92)
- Adaptive Feature Imputation with Latent Graph for Deep Incomplete Multi-View Clustering (AAAI 2024 · similarity 0.91)
- Deep Variational Incomplete Multi-View Clustering: Exploring Shared Clustering Structures (AAAI 2024 · similarity 0.90)
- Incomplete Multi-view Clustering via Diffusion Contrastive Generation (AAAI 2025 · similarity 0.90)
- EDISON: Enhanced Dictionary-Induced Tensorized Incomplete Multi-View Clustering with Gaussian Error Rank Minimization (ICML 2024 · similarity 0.89)
- Incomplete Contrastive Multi-View Clustering with High-Confidence Guiding (AAAI 2024 · similarity 0.89)
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.