Research map / Label-efficient and robust learning / Recognition and long-tail
Interactive Deep Clustering via Value Mining
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 Interactive Deep Clustering (IDC), a plug-in that selects high-value hard, representative, diverse samples, asks users for their cluster affiliations among nearest centers, and finetunes pre-trained deep clustering models with positive, negative and regularization losses.
Problem
Deep clustering methods without class priors struggle to correctly assign hard samples at cluster boundaries, since their pseudo-label-based internal supervision is unreliable there, creating a performance ceiling.
Approach
Instead of mining more supervision internally, select a small budget (M=500) of samples by a value score combining hardness, representativeness (kNN density) and diversity (greedy selection), query users for cluster affiliation among top-5 nearest centers, and finetune the pre-trained model with positive, negative and confidence-based regularization losses.
Key results
- IDC boosts TCL accuracy by 16.3% on CIFAR-20 (to 69.4%) and ProPos by 19.2% on CIFAR-20 (to 78.3%) using only 500 queried samples.
- IDC on ImageNet-Dogs improves TCL and ProPos accuracy by 14.4% and 9.2% respectively.
- Ablation shows hardness alone lowers NMI on CIFAR-20 (51.3 vs 52.2) and hardness+representativeness drops ACC to 36.3, while adding diversity yields 69.4 ACC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add IDC value mining and user-feedback finetuning to TCL | TCL · CIFAR-20 | ACC: 53.1 -> 69.4 | Helped |
| Add IDC value mining and user-feedback finetuning to ProPos | ProPos · ImageNet-Dogs | ACC: 76.9 -> 86.1 | Helped |
| Select valuable samples using hardness only (ablation) | TCL with IDC · CIFAR-20 | ACC: 52.6 -> 57.7 | Helped |
| Add representativeness to hardness-based selection (ablation) | TCL with IDC · CIFAR-20 | ACC: 57.7 -> 36.3 | Hurt |
| Add diversity to hardness-and-representativeness selection (ablation) | TCL with IDC · CIFAR-20 | ACC: 36.3 -> 69.4 | Helped |
| Use only the positive-feedback loss (ablation) | TCL with IDC · CIFAR-20 | ACC: 52.6 -> 65.8 | Helped |
| Use only the negative-feedback loss (ablation) | TCL with IDC · CIFAR-20 | ACC: 52.6 -> 9.6 | Hurt |
| Use only the regularization loss (ablation) | TCL with IDC · CIFAR-20 | ACC: 52.6 -> 52.9 | No clear effect |
| Add negative-feedback loss to positive-feedback loss (ablation) | TCL with IDC · CIFAR-20 | ACC: 65.8 -> 66.2 | Helped |
| Add regularization loss to positive-feedback loss (ablation) | TCL with IDC · CIFAR-20 | ACC: 65.8 -> 67.6 | Helped |
| Add all three loss terms (ablation) | TCL with IDC · CIFAR-20 | ACC: 67.6 -> 69.4 | Helped |
| Use all three loss terms (ablation) | TCL with IDC · ImageNet-Dogs | ACC: 77.5 -> 78.8 | Helped |
Surprising findings
- Selecting only hard and representative samples without diversity collapses the selection into dense subsets and severely degrades clustering (CIFAR-20 ACC 36.3, below the pre-trained model's 52.6).
- Using only the negative loss causes model collapse, as it denies all top-5 predictions.
Limitations
- Assumes user feedback is correct; robustness to user mistakes is not studied.
- ProPos results on STL-10 unavailable because its code was not released.
- Compared against interaction-free baselines and a FixMatch/Cop-Kmeans setup rather than other interactive clustering methods with the same budget.
Experimental setup
Datasets: CIFAR-10 CIFAR-20 STL-10 ImageNet-10 ImageNet-Dogs
Benchmarks: CIFAR-10 CIFAR-20 STL-10 ImageNet-10 ImageNet-Dogs
Metrics: NMI ACC ARI
Baselines: CC SCAN NMM MiCE BYOL GCC SPICE IDFD TCC DivClust SeCu CoNR
Builds on: TCL ProPos contrastive clustering
Closest papers by meaning
- Samples Are Not Equal: A Sample Selection Approach for Deep Clustering (ICLR 2026 · similarity 0.89)
- DeepDPM: Deep Clustering With an Unknown Number of Clusters (CVPR 2022 · similarity 0.86)
- From Logits to Hierarchies: Hierarchical Clustering made Simple (ICML 2025 · similarity 0.85)
- You Can Trust Your Clustering Model: A Parameter-free Self-Boosting Plug-in for Deep Clustering (NeurIPS 2025 · similarity 0.85)
- Hard Regularization to Prevent Deep Online Clustering Collapse without Data Augmentation (AAAI 2024 · similarity 0.84)
- Learning from Sample Stability for Deep Clustering (ICML 2025 · similarity 0.84)
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.