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

GeCC: Generalized Contrastive Clustering with Domain Shifts Modeling

AAAI 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Contrastive representation theory · 1 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.

Proposes GeCC, a contrastive deep clustering framework that synthesizes a cluster-guided reference view via domain shift modeling and reweights instance and cluster contrastive losses with attention weights to reduce augmentation-specific features.

Problem

Contrastive clustering relies on predefined data augmentations, so representations get stuck in domain-specific information, degrading clustering quality and generalization.

Approach

Extends DSU-style feature-statistic perturbation with cluster-guided per-cluster statistics to create a reference view, and adds domain (instance-level) and distribution (cluster-level) attention weights to the contrastive losses so the model learns from more diverse domains and disentangles from augmentation-specific features.

Key results

• On ImageNet-Dogs, GeCC improves over the second-best DivClust by up to 3.35% NMI, 7.38% ACC and 4.53% ARI.
• On ImageNet-10, GeCC reaches 0.9615 ACC and 0.9167 ARI, versus 0.9360 ACC for DivClust.
• On CIFAR-10, GeCC achieves 0.7559 NMI and 0.8440 ACC, outperforming CC (0.7050 NMI, 0.7900 ACC); on CIFAR-100 DeepCluE remains best, with GeCC comparable.

Measured effects reported in the paper

Surprising findings

• Replacing the cluster-guided reference view with plain DSU slightly degrades CIFAR-10 and CIFAR-100 clustering relative to the CC baseline, which the authors attribute to DSU ignoring cluster assignments.
• Applying the cluster distribution attention weight alone (w Wc) deteriorates results substantially, below the CC baseline on most datasets.

Limitations

• On CIFAR-100, the method is not best; DeepCluE achieves higher results.
• Gains on ImageNet datasets from the domain shift module are limited according to ablations.
• Relies on pseudo-labels from the cluster assignment, which are less accurate early in training, requiring a heuristic probability schedule.

Experimental setup

Datasets: CIFAR-10 CIFAR-100 ImageNet-10 ImageNet-Dogs

Benchmarks: CIFAR-10 CIFAR-100 ImageNet-10 ImageNet-Dogs

Metrics: NMI ACC ARI

Baselines: K-means SC AC NMF AE DAE GAN DeCNN VAE JULE DEC DAC

Builds on: Contrastive Clustering (CC) Domain Shifts with Uncertainty (DSU) NT-Xent contrastive loss SimCLR

Closest papers by meaning

• Cross-Domain Contrastive Learning for Time Series Clustering (AAAI 2024 · similarity 0.88)
• Contrastive Mean-Shift Learning for Generalized Category Discovery (CVPR 2024 · similarity 0.87)
• Generalized Category Discovery under Domain Shift: A Frequency Domain Perspective (NeurIPS 2025 · similarity 0.86)
• THESAURUS: Contrastive Graph Clustering by Swapping Fused Gromov-Wasserstein Couplings (AAAI 2025 · similarity 0.86)
• Contrastive Multi-view Subspace Clustering via Tensor Transformers Autoencoder (AAAI 2025 · similarity 0.86)
• Deep Contrastive Graph Learning with Clustering-Oriented Guidance (AAAI 2024 · similarity 0.85)