Research map / Learning theory and algorithms / Clustering
Bifurcate then Alienate: Incomplete Multi-view Clustering via Coupled Distribution Learning with Linear Overhead
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 BACDL, an incomplete multi-view clustering method that bifurcates feature clusters into perspective-shared and perspective-specific determinants, alienates them via mutual exclusion, couples view guidance via distribution learning, and bridges views with a linear-complexity multiplicative update scheme.
Problem
Incomplete multi-view clustering (IMC) methods typically exploit either perspective-shared or perspective-specific determinants alone, which limits how well heterogeneous multi-view data is exploited, and many methods are costly on large datasets.
Approach
Simultaneously model shared and view-specific feature-cluster determinants by bifurcating feature clusters (view guidance concatenated with view-specific clusters), enforce their discrimination with an inner-product mutual-exclusion loss, and couple views through shared marginal/conditional distribution learning in a non-negative matrix factorization framework with an overall linear-time update ru
Key results
- On SYNTHREED at missing ratio 0.2, BACDL reaches 42.50 PUR versus the best baseline around 40.45 (HCCGL) and 40.38 (IMVCI).
- BACDL is reported as the most memory-efficient method on SYNTHREED, BGFEA, AWTEN, HDIGTWO and YOUFOURV, and the only one of the compared methods that scales to YOUFOURV among several baselines that run out of resources.
- Ablations show that removing either the shared determinants, the specific determinants, or the alienation term lowers clustering scores across all eight datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add perspective-shared determinants (ablation) | BACDL · FLOEVEN, missing ratio 0.2 | ACC: 9.53 -> 10.80 | Helped |
| Add perspective-specific determinants (ablation) | BACDL · FLOEVEN, missing ratio 0.2 | ACC: 9.07 -> 10.80 | Helped |
| Use the alienating action (ablation) | BACDL · FLOEVEN, missing ratio 0.2 | ACC: 9.57 -> 10.80 | Helped |
| Use BACDL instead of IMVCI | incomplete multi-view clustering · FLOEVEN, missing ratio 0.2 | ACC: 10.86 -> 10.80 | Hurt |
| Use BACDL instead of PIMVC | incomplete multi-view clustering · SYNTHREED, missing ratio 0.2 | ACC: 40.67 -> 42.12 | Helped |
| Use BACDL instead of PIMVC | incomplete multi-view clustering · DEOLOG, missing ratio 0.2 | ACC: 22.04 -> 22.32 | Helped |
| Use BACDL instead of LRTL | incomplete multi-view clustering · YALTHREE, missing ratio 0.2 | ACC: 21.52 -> 23.55 | Helped |
| Use BACDL instead of PIMVC | incomplete multi-view clustering · BGFEA, missing ratio 0.2 | ACC: 20.52 -> 22.08 | Helped |
| Use BACDL instead of GIMC | incomplete multi-view clustering · AWTEN, missing ratio 0.2 | ACC: 10.98 -> 12.24 | Helped |
| Use BACDL instead of HCCGL | incomplete multi-view clustering · HDIGTWO, missing ratio 0.2 | ACC: 11.27 -> 11.51 | Helped |
| Use BACDL instead of LBIMV | incomplete multi-view clustering · YOUFOURV, missing ratio 0.2 | ACC: 10.54 -> 10.91 | Helped |
Surprising findings
- Several baselines (LRTL, TCIMC, AGCIM, LSIMV, IMVCI, HCCGL, UIMC) cannot run on the large YOUFOURV dataset due to complexity limits.
Limitations
- Authors attribute inferior results on FLOEVEN and YALTHREE to generating cluster labels from the spectrum rather than from original data, which may reduce view diversity.
- Several baselines are slower or faster in some cases (LSIMV, PIMVC, HCCGL run faster on some datasets).
Experimental setup
Datasets: FLOEVEN SYNTHREED DEOLOG YALTHREE BGFEA AWTEN HDIGTWO YOUFOURV
Benchmarks: FLOEVEN SYNTHREED DEOLOG YALTHREE BGFEA AWTEN HDIGTWO YOUFOURV
Metrics: PUR ACC FSC running time memory overhead
Baselines: LRTL TCIMC AGCIM LSIMV GIMC IMVCI PIMVC HCCGL USETL LBIMV UIMC
Builds on: non-negative matrix factorization multi-view NMF
Closest papers by meaning
- Towards Resource-friendly, Extensible and Stable Incomplete Multi-view Clustering (ICML 2024 · similarity 0.92)
- 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 0.89)
- Incomplete Contrastive Multi-View Clustering with High-Confidence Guiding (AAAI 2024 · similarity 0.89)
- Adaptive Feature Imputation with Latent Graph for Deep Incomplete Multi-View Clustering (AAAI 2024 · similarity 0.88)
- Sample-Level Cross-View Similarity Learning for Incomplete Multi-View Clustering (AAAI 2024 · 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.