Research map / Label-efficient and robust learning / Recognition and long-tail
Partial Multi-View Multi-Label Classification via Semantic Invariance Learning and Prototype Modeling
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 SIP, an information-bottleneck VAE with product-of-experts fusion and data-driven label prototypes for partial multi-view multi-label classification, evaluated on five multi-view multi-label image datasets.
Problem
Partial multi-view multi-label classification, where both views and labels are missing, requiring cross-view shared representations that remain task-relevant for predicting multiple labels while modeling label correlations.
Approach
Learns a cross-view representation by an information-bottleneck objective that compresses non-shared view information while keeping reconstruction validity, fuses views with product-of-experts, and models label prototypes as latent distributions aligned with sample embeddings to capture label correlations in a data-driven way.
Key results
- On Corel5k with 50% missing views and labels, SIP reaches AP 0.418 vs 0.381 for the next best (DICNet), ranking first on all datasets in average rank.
- On Pascal07, SIP achieves AP 0.555 vs 0.505 for DICNet.
- Ablation on Corel5k shows removing L_rec drops AP from 0.418 to 0.257, and removing L_sha alone yields 0.383.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SIP instead of DICNet | PMvMLC · Corel5k | AP: 0.3810.004 -> 0.4180.009 | Helped |
| Use SIP instead of CDMM | PMvMLC · Pascal07 | AP: 0.5080.005 -> 0.5550.010 | Helped |
| Use SIP instead of DICNet | PMvMLC · ESPGame | AP: 0.2970.002 -> 0.3110.004 | Helped |
| Use SIP instead of DICNet | PMvMLC · IAPRTC12 | AP: 0.3230.001 -> 0.3310.006 | Helped |
| Use SIP instead of DICNet | PMvMLC · MIRFLICKR | AP: 0.5890.005 -> 0.6140.004 | Helped |
| Add Lsha, Lrec, and LPA (ablation) | SIP · Corel5k | AP: 0.345 -> 0.418 | Helped |
| Add Lsha and LPA (ablation) | SIP · Pascal07 | AP: 0.547 -> 0.555 | Helped |
| Add Lrec and LPA (ablation) | SIP · Corel5k | AP: 0.387 -> 0.418 | Helped |
| Add LPA (ablation) | SIP · Corel5k | AP: 0.415 -> 0.418 | Helped |
| Add LPA (ablation) | SIP · Pascal07 | AP: 0.549 -> 0.555 | Helped |
| Add Lrec (ablation) | SIP · Corel5k | AP: 0.257 -> 0.418 | Helped |
| Add Lsha (ablation) | SIP · Pascal07 | AP: 0.550 -> 0.555 | Helped |
Surprising findings
- Adding the shared-information compression term L_sha without the reconstruction term L_rec significantly degrades performance (AP 0.257 on Corel5k), showing compression must be balanced by information retention.
Limitations
- Missing views and labels are filled with zeros and baselines required ad hoc adaptations, which may bias comparisons.
- Results are reported only on image-based multi-view datasets with hand-crafted features.
- Single fixed missing rate of 50% evaluated in main experiments.
Experimental setup
Datasets: Corel5k Pascal07 ESPGame IAPRTC12 MIRFLICKR
Benchmarks: Corel5k Pascal07 ESPGame IAPRTC12 MIRFLICKR
Metrics: AP 1-HL 1-RL AUC 1-OE 1-Cov average ranking
Baselines: C2AE GLOCAL CDMM DM2L LVSL iMVWL NAIM3L DICNet
Builds on: information bottleneck product of experts DICNet variational autoencoder
Closest papers by meaning
- Incomplete Multi-View Multi-label Learning via Disentangled Representation and Label Semantic Embedding (CVPR 2025 · similarity 0.91)
- Learning Compact Semantic Information for Incomplete Multi-View Missing Multi-Label Classification (ICML 2025 · similarity 0.90)
- Multi-View Multi-Label Classification via View-Label Matching Selection (AAAI 2025 · similarity 0.90)
- Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels (CVPR 2025 · similarity 0.89)
- CFDM: Contrastive Fusion and Disambiguation for Multi-View Partial-Label Learning (AAAI 2025 · similarity 0.89)
- A Two-Stage Information Extraction Network for Incomplete Multi-View Multi-Label Classification (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.