Research map / Label-efficient and robust learning / Recognition and long-tail
Enhanced Denesity Peak Clustering for High-Dimensional Data
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 EDPC, a density peak clustering method that selects potential centers using average density and density-related distance thresholds, trains a one-hidden-layer MLP on sub-cluster pseudo-labels for dimensionality reduction, and merges sub-clusters hierarchically via KMD-linkage.
Problem
Density peak clustering (DPC) tends to pick cluster centers only in high-density regions, missing representative peaks in low-density areas, and its performance degrades in high-dimensional, sparse data where distances become less distinct; one-step label assignment also propagates cascading errors.
Approach
Select potential centers with the threshold rule rho >= λ·mean(rho) and delta >= β·mean(delta) so low-density peaks are included; use DPC sub-cluster pseudo-labels to supervise a single-hidden-layer MLP whose hidden layer gives a reduced embedding; then merge sub-clusters hierarchically with KMD-linkage.
Key results
- On 8 synthetic datasets, EDPC achieves average ACC 0.9916 versus 0.9388 for the best baseline SMMP (AMI 0.9814 vs 0.9464).
- Across 18 real high-dimensional datasets, EDPC improves average ACC by 23.0% over DPC, 16.9% over SNN-DPC, and 10.0% over spectral method CDGC.
- EDPC reaches 1.0000 ACC on Alizadeh, versus 0.9839 for CDGC, the strongest baseline on that dataset.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add hierarchical merging of sub-clusters instead of assigning labels directly through cluster centers (ablation) | EDPC-2 · High-dimensional datasets (Alizadeh, Bhatt, Gordon, Golub1) | ACC | Helped |
| Add MLP-based dimensionality reduction guided by sub-cluster pseudo-labels (ablation) | EDPC-2 · High-dimensional datasets (Alizadeh, Bhatt, Gordon, Golub1) | ACC | Helped |
| Use λ and β greater than 0.5 when selecting potential cluster centers | EDPC · Alizadeh, Gordon, Golub1, Bhatt | ACC | Helped |
| Set φ in the range [5, 20] for KMD-linkage | EDPC · Alizadeh, Bhatt, Gordon, Golub1 | ACC | Helped |
Surprising findings
- The paper reports that EDPC-2 (without MLP reduction) outperforms EDPC-1 (without sub-cluster merging), highlighting the value of hierarchical label assignment.
Limitations
- Reduced dimensionality fixed at 50 without a method for choosing it optimally
- Performance sensitive to thresholds lambda and beta, with best results when both exceed 0.5
- Quadratic-order cost of distance computation O(n^2) for large n
Experimental setup
Datasets: Dart Donut Cno3 Cno4 Curet Cuboids D31 Jain Alizadeh Armstrong Bhatt Garber
Benchmarks: Dart Donut Cno3 Cno4 Curet Cuboids D31 Jain Alizadeh Armstrong Bhatt Garber
Metrics: ACC AMI
Baselines: DPC SNN-DPC SMMP KNN-DPC DB-DPC FHC-LDP CDGC DPC-tSNE SMMP-tSNE
Builds on: Density Peak Clustering (DPC) SNN-DPC KMD-linkage t-SNE PCA
Closest papers by meaning
- Resolution-Aware Structural Density Peak Clustering (NeurIPS 2026 · similarity 0.90)
- Samples Are Not Equal: A Sample Selection Approach for Deep Clustering (ICLR 2026 · similarity 0.88)
- DBSCAN++: Towards fast and scalable density clustering (ICML 2019 · similarity 0.85)
- Scalable DBSCAN with Random Projections (NeurIPS 2024 · similarity 0.83)
- From Logits to Hierarchies: Hierarchical Clustering made Simple (ICML 2025 · similarity 0.83)
- Interactive Deep Clustering via Value Mining (NeurIPS 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.