Research map / Label-efficient and robust learning / Recognition and long-tail

Uncertainty-Aware Probabilistic Constrained Clustering from Entangled Pairwise Supervision

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Learning from unlabeled and positive-unlabeled data · 0 citations (OpenAlex)

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.

Formalizes uncertainty-aware probabilistic constrained clustering with a heterogeneous observation model, proposes the ProbPair angular soft-pairwise loss and the Estimator-Corrector-Integrator (ECI-PP) framework that refines noisy real-valued pairwise supervision, and reports gains over DCC baselines on eight benchmarks.

Problem

Constrained clustering usually assumes hard must-link/cannot-link labels, but real pairwise supervision is real-valued and mixes aleatoric ambiguity, expert-specific judgment bias, and random corruption.

Approach

Model probabilistic pairwise labels with an angular sigmoid readout on cosine similarity (ProbPair), and iteratively refine supervision via cross-fitted out-of-fold beliefs, per-expert neural correctors, and reliability-weighted Bayesian-confidence targets (ECI-PP).

Key results

• ECI-PP ranks first in 41 of 48 test entries (2 expert regimes x 8 datasets x 3 metrics) and top-two in 46/48 on the main comparison.
• On single-expert CIFAR10 (lv0.01), ECI-PP reaches 87.8% ACC and 79.0% NMI versus 84.9% and 72.4% for SpherePair.
• On multi-expert Reuters (multi3), ECI-PP reaches 88.3% ACC and 65.4% NMI versus 81.0% and 51.5% for the next best Weighted ProbPair.

Surprising findings

• SpherePair retains the best ACC on the imbalanced RCV1-10 (single-expert 71.2 vs ECI-PP 69.0), though ECI-PP leads on NMI/ARI.
• Under single-expert FMNIST, ProbPair-family methods (Weighted ProbPair and ProbPair) outperform ECI-PP on ACC, so ECI-PP is not uniformly best.

Limitations

• Identifiability is population-level and conditional; finite-sample factor-wise estimation is not established.
• Experiments use controlled simulated experts rather than real annotators.
• Corrector and screening weight are heuristic surrogates rather than estimators of individual latent factors.

Experimental setup

Datasets: CIFAR100-20 CIFAR10 FMNIST ImageNet10 MNIST STL10 Reuters subset RCV1-10

Benchmarks: CIFAR100-20 CIFAR10 FMNIST ImageNet10 MNIST STL10 Reuters subset RCV1-10

Metrics: ACC NMI ARI Brier score AUC AP

Baselines: VanillaDCC VolMaxDCC CIDEC SpherePair ProbPair Weighted ProbPair

Builds on: SpherePair VolMaxDCC CIDEC VanillaDCC DEC

Closest papers by meaning

• Deep Conditional Gaussian Mixture Model for Constrained Clustering (NeurIPS 2021 · similarity 0.89)
• Fairness, Semi-Supervised Learning, and More: A General Framework for Clustering with Stochastic Pairwise Constraints (AAAI 2021 · similarity 0.88)
• Angular Constraint Embedding via SpherePair Loss for Constrained Clustering (NeurIPS 2025 · similarity 0.86)
• Towards Better-than-2 Approximation for Constrained Correlation Clustering (ICML 2025 · similarity 0.84)
• P$^2$OT: Progressive Partial Optimal Transport for Deep Imbalanced Clustering (ICLR 2024 · similarity 0.83)
• Efficient Constrained K-center Clustering with Background Knowledge (AAAI 2024 · similarity 0.83)