Research map / Label-efficient and robust learning / Recognition and long-tail
Personalized Clustering via Targeted Representation Learning
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 PCL, a personalized clustering method that actively queries must-link/cannot-link pairs (uncertainty plus hard-negative scoring) to drive attention-augmented targeted representation learning with a constrained contrastive loss.
Problem
Standard deep clustering finds the default grouping structure, which may not match a user's desired clustering orientation; constrained clustering with random or default-oriented queries fails to steer clustering toward personalized criteria.
Approach
Actively query the most informative pairs (uncertain pairs by cosine similarity near a threshold, plus hard negatives by relative entropy between model and constraint-inferred probabilities) and train with a constrained contrastive loss plus cross-instance attention so that representations align with the user-specified cluster orientation.
Key results
- On personalized orientation of CIFAR10-2, PCL reaches NMI 0.453 and ACC 0.873, versus best baseline DCC at NMI 0.244 and ACC 0.789.
- On personalized CIFAR100-4, PCL achieves NMI 0.234 and ACC 0.597, while GCC gets NMI 0.230 and ACC 0.508.
- Under the default orientation, PCL is best only on CIFAR10-2 (NMI 0.737); GCC leads on CIFAR100-4 and ImageNet10-2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PCL instead of the strongest baseline, DCC | Personalized clustering · CIFAR10-2, personalized orientation | ACC: 0.789 -> 0.873 | Helped |
| Use PCL instead of the strongest baseline, GCC | Personalized clustering · CIFAR100-4, personalized orientation | ACC: 0.508 -> 0.597 | Helped |
| Use PCL instead of the strongest baseline, SPICE | Personalized clustering · ImageNet10-2, personalized orientation | ACC: 0.648 -> 0.857 | Helped |
| Use PCL instead of the strongest baseline, GCC | Clustering · CIFAR10-2, default orientation | ACC: 0.934 -> 0.955 | Helped |
| Use PCL instead of the strongest baseline, GCC | Clustering · CIFAR100-4, default orientation | ACC: 0.604 -> 0.575 | Hurt |
| Use PCL instead of the strongest baseline, GCC | Clustering · ImageNet10-2, default orientation | ACC: 0.988 -> 0.952 | Hurt |
| Add the cross-attention module (ablation) | PCL · CIFAR100-4, personalized orientation | NMI: 0.127 -> 0.234 | Helped |
| Use random queries instead of the full joint query strategy (ablation) | PCL · CIFAR100-4, personalized orientation | NMI: 0.234 -> 0.172 | Hurt |
| Use the uncertainty score only instead of the full joint query strategy (ablation) | PCL · CIFAR100-4, personalized orientation | NMI: 0.234 -> 0.196 | Hurt |
| Use the hard-negative score only instead of the full joint query strategy (ablation) | PCL · CIFAR100-4, personalized orientation | NMI: 0.234 -> 0.160 | Hurt |
Surprising findings
- Under the default orientation, unsupervised methods such as GCC and MiCE match or beat constrained methods.
- Constraint-free methods perform near chance on personalized orientation (e.g., SPICE NMI 0.005 on CIFAR10-2).
Limitations
- Personalized orientations are artificially constructed by regrouping existing class labels, not from real user preferences.
- Default-orientation performance is comparable to rather than uniformly better than baselines (PCL trails GCC on CIFAR100-4 and ImageNet10-2 in default setting).
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet-10
Benchmarks: CIFAR10-2 CIFAR100-4 ImageNet10-2
Metrics: NMI ARI F-measure ACC
Baselines: K-means MiCE CC GCC SPICE DCC ADC SDEC DC-GMM
Builds on: infoNCE SPICE DC-GMM ADC DCC
Closest papers by meaning
- Fairness, Semi-Supervised Learning, and More: A General Framework for Clustering with Stochastic Pairwise Constraints (AAAI 2021 · similarity 0.85)
- Parameter-Free Clustering via Self-Supervised Consensus Maximization (AAAI 2026 · similarity 0.85)
- Interactive Deep Clustering via Value Mining (NeurIPS 2024 · similarity 0.83)
- Clustering Properties of Self-Supervised Learning (ICML 2025 · similarity 0.83)
- THESAURUS: Contrastive Graph Clustering by Swapping Fused Gromov-Wasserstein Couplings (AAAI 2025 · similarity 0.83)
- Partial Multi-View Clustering via Self-Supervised Network (AAAI 2024 · similarity 0.83)
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.