Research map / Label-efficient and robust learning / Recognition and long-tail
Tensorized Multi-View Multi-Label Classification via Laplace Tensor Rank
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 TMvML, a tensorized multi-view multi-label classifier that rotates the stacked per-view mapping matrices into a label-view-feature tensor and penalizes it with a new Laplace Tensor Rank, optimized via ADMM.
Problem
Multi-view multi-label classification must capture cross-view consistent correlations and multi-label semantic relationships jointly, but existing methods treat these with separate components and only capture low-order correlations.
Approach
Reshape per-view mapping matrices into a 3-way tensor, rotate it so label-by-view slices are frontal, and impose a low-rank constraint via a new Laplace Tensor Rank (a non-convex surrogate with f(x)=1-exp(-e^{δx}/δ)) so that view and label correlations are learned together.
Key results
- TMvML ranks first on all five metrics per Bonferroni-Dunn CD diagrams and wins 180-comparison tally on Emotions, Yeast, Plant and Human datasets.
- On Emotions, TMvML reaches AP 0.811 vs 0.806 for the next best (FIMAN) and One Error 0.248.
- On Yeast, TMvML reaches AP 0.771 vs 0.740 for FIMAN and Coverage 0.470 vs 6.538 for IMvMLC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TMvML instead of FIMAN | MVML classification · Emotions | Average Precision: 0.806±0.027 -> 0.811±0.020 | Helped |
| Use TMvML instead of FIMAN | MVML classification · Yeast | Average Precision: 0.740±0.007 -> 0.771±0.008 | Helped |
| Use TMvML instead of FIMAN | MVML classification · Corel5k | Average Precision: 0.430±0.007 -> 0.440±0.008 | Helped |
| Use TMvML instead of ICM2L | MVML classification · Plant | Average Precision: 0.587±0.047 -> 0.608±0.007 | Helped |
| Use TMvML instead of SIMM | MVML classification · Espgame | Average Precision: 0.304±0.002 -> 0.306±0.001 | Helped |
| Use TMvML instead of ICM2L | MVML classification · Human | Average Precision: 0.603±0.034 -> 0.631±0.010 | Helped |
| Replace LTR with ETR (ablation) | TMvML · Emotions and Corel5k | Average Precision | Hurt |
| Replace LTR with TNN (ablation) | TMvML · Emotions and Corel5k | Average Precision | Hurt |
| Vary δ in LTR (ablation) | TMvML · Emotions and Corel5k | Average Precision | Mixed |
Surprising findings
- ICM2L shows Hamming Loss of 0.920 on Plant and Human, indicating very poor calibration in those baselines.
- TMvML-ETR can outperform full TMvML on a few cases, showing sensitivity to the delta parameter.
Limitations
- Parameter delta strongly affects performance and must be tuned per dataset.
- Evaluated only on six relatively small, classic MVML benchmarks with features rather than deep representations.
- Compared against a limited set of MVML baselines; no deep-learning MVML baselines reported.
Experimental setup
Datasets: Emotions Yeast Corel5k Plant Human Espgame
Benchmarks: Emotions Yeast Corel5k Plant Human Espgame
Metrics: Average Precision Hamming Loss One Error Ranking Loss Coverage
Baselines: LrMMC SIMM FIMAN ICM2L ML-BVAE IMvMLC TMvML-ETR TMvML-TNN
Builds on: tensor nuclear norm alternating direction method of multipliers enhanced tensor rank
Closest papers by meaning
- Multi-View Multi-Label Classification via View-Label Matching Selection (AAAI 2025 · similarity 0.90)
- Incomplete Multi-View Multi-label Learning via Disentangled Representation and Label Semantic Embedding (CVPR 2025 · similarity 0.88)
- T-MDML: Triplet-based Multiple Distance Metric Learning for Multi-Instance Multi-Label Classification with Label Correlation (AAAI 2025 · similarity 0.87)
- S2MVTC: a Simple yet Efficient Scalable Multi-View Tensor Clustering (CVPR 2024 · similarity 0.87)
- Large-scale Multi-view Tensor Clustering with Implicit Linear Kernels (CVPR 2025 · similarity 0.87)
- Partial Multi-View Multi-Label Classification via Semantic Invariance Learning and Prototype Modeling (ICML 2024 · similarity 0.87)
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.