Source-linked AI summary
DealMVC: Dual Contrastive Calibration for Multi-view Clustering
Xihong Yang, Jiaqi Jin, Siwei Wang, Ke Liang, Yue Liu, Yi Wen, Suyuan Liu, Sihang Zhou, Xinwang Liu, En Zhu
TL;DR
Existing multi-view clustering methods emphasize identical-sample consistency across views while overlooking similar but distinct cross-view samples. DealMVC combines adaptive global fusion with global graph alignment and local pairwise-view calibration, and experiments across eight datasets report favorable clustering outcomes, including gains on STL10.
Problem
Existing multi-view clustering methods mainly preserve consistency for identical samples across views while overlooking similar but different cross-view samples.
Method
DealMVC uses adaptive fusion to obtain global cross-view features, aligns feature-similarity and high-confidence pseudo-label graphs globally, and applies local calibration to pairwise-view features.
Results
DealMVC achieves the most favorable clustering outcomes across most datasets and surpasses its closest STL10 competitors by 5.30% ACC, 3.74% NMI, and 5.70% PUR.
Takeaways & Limitations
Dual contrastive calibration keeps similar but different cross-view samples consistent by constraining feature structure with reliable class information.
Takeaways & Limitations
The method assumes label information remains consistent across views when constructing local and global pseudo-label consistency losses.
Abstract
from arXiv · showhide
Benefiting from the strong view-consistent information mining capacity, multi-view contrastive clustering has attracted plenty of attention in recent years. However, we observe the following drawback, which limits the clustering performance from further improvement. The existing multi-view models mainly focus on the consistency of the same samples in different views while ignoring the circumstance of similar but different samples in cross-view scenarios. To solve this problem, we propose a novel Dual contrastive calibration network for Multi-View Clustering (DealMVC). Specifically, we first design a fusion mechanism to obtain a global cross-view feature. Then, a global contrastive calibration loss is proposed by aligning the view feature similarity graph and the high-confidence pseudo-label graph. Moreover, to utilize the diversity of multi-view information, we propose a local contrastive calibration loss to constrain the consistency of pair-wise view features. The feature structure is regularized by reliable class information, thus guaranteeing similar samples have similar features in different views. During the training procedure, the interacted cross-view feature is jointly optimized at both local and global levels. In comparison with other state-of-the-art approaches, the comprehensive experimental results obtained from eight benchmark datasets provide substantial validation of the effectiveness and superiority of our algorithm. We release the code of DealMVC at https://github.com/xihongyang1999/DealMVC on GitHub.
1 INTRODUCTION
Multi-view clustering methods commonly preserve consistency between identical samples across views but overlook similar, distinct cross-view samples. DealMVC addresses this gap with dual contrastive calibration using global fusion, graph alignment, and local pairwise-view constraints, supported by experiments across eight datasets.
- Multi-view clustering reveals semantic information by dividing data into disjoint groups, using conventional or deep algorithms.
- Existing methods often align identical samples across views while neglecting consistency between similar but different cross-view samples.
- DealMVC obtains a global cross-view feature through adaptive fusion using attention and learnable view sampling mechanisms.
- Its global contrastive calibration aligns view-feature similarity graphs with high-confidence pseudo-label graphs, while local calibration constrains pairwise-view feature consistency.
- Experiments on eight benchmark datasets, including ablation and visualization studies, support DealMVC’s effectiveness and superiority.
2 RELATED WORK
Related work spans traditional and deep multi-view clustering, including matrix, kernel, graph, subspace, and contrastive approaches. DealMVC is positioned within deep contrastive learning while combining adaptive fusion with dual calibration of global and local cross-view structure.
- Traditional multi-view clustering includes matrix factorization, kernel learning, graph-based, and subspace clustering methods.
- Deep multi-view clustering methods use learned representations to uncover latent cluster patterns and include one-stage and two-stage algorithms.
- Contrastive learning maximizes similarity between positive samples and minimizes similarity between negative samples in latent space.
- DealMVC extracts view features, adaptively fuses them into a global feature, and aligns feature similarity with high-confidence pseudo-label graphs.
- Prior multi-view contrastive work includes frameworks that learn consistent representations or contrastively optimize feature and pseudo-label objectives.
3 METHODOLOGY
DealMVC builds global and local cross-view representations, then calibrates their feature graphs against high-confidence pseudo-label graphs. Its adaptive fusion combines attention and learnable view sampling, while dual losses preserve consistency for identical and similar-but-different samples.
- 3.1 AutoEncoder Module: DealMVC uses an autoencoder to project each view into a customizable feature space and reconstruct the input.For view v, the encoder produces Z_v and the decoder produces X̃_v; reconstruction loss is applied across views.
- 3.2 Adaptive Global Fusion Module: Adaptive global fusion combines multi-head attention with a learnable view-sampling probability to produce a global cross-view feature.The two distributions are combined through a regulatory factor and adaptive weight vector w.
- 3.2 Adaptive Global Fusion Module: The regulatory factor gives stronger adjustment when attention and view-sampling distributions are more similar, supporting important-view extraction.The fused global feature is obtained from the resulting adaptive view weights.
- 3.3 Dual Contrastive Calibration Module: Global contrastive calibration aligns the global cross-view feature similarity graph with a high-confidence pseudo-label graph.Its two terms bring identical samples together and encourage similar-but-different samples to share clusters.
- 3.3 Dual Contrastive Calibration Module: The training procedure constructs pseudo-label and cross-view similarity graphs, aligns them with contrastive losses, and updates the model by minimizing the combined loss.A mean squared error loss also keeps pseudo-label graphs consistent at global and local levels.
- 3.3 Dual Contrastive Calibration Module: Local contrastive calibration aligns pair-wise view feature graphs with pseudo-label graphs across all pair-wise view combinations.The local objective constrains both identical samples and different samples with similar pseudo labels to have similar cross-view features.
4.1 Dataset
Experiments use eight benchmark datasets to evaluate DealMVC through clustering performance, module effects, hyper-parameter sensitivity, and revealed clustering structure.
- DealMVC is evaluated on eight datasets: BBCSport, Reuters, Caltech101_7, Cora, Wiki, Hdigit, Caltech101, and STL10.
- The experiments address effectiveness, module influence, hyper-parameter impact, and clustering structure through four research questions.
- Tables 3 and 4 report clustering performance across the eight multi-view benchmark datasets, marking the best and second-best results.
4.2 Experiment Setup
The study compares DealMVC with deep and contrastive multi-view clustering baselines using standardized implementation settings and clustering metrics.
- Experiments run with an Intel Core i7-7820x CPU, NVIDIA GeForce RTX 3080 GPU, 64GB RAM, PyTorch, and Adam optimization.
- DealMVC is compared with deep methods AE2-Net, CoMIC, DEMVC, SDMVC, SDSNE, DSMVC, and SiMVC, plus contrastive methods CoMVC and MFLVC.
- Baseline results are replicated with their provided source code and original configurations, while DealMVC uses fixed training settings across datasets.DealMVC uses batch size 256, learning rate 0.0003, 300 pre-training epochs, and 100 contrastive-learning epochs.
- Clustering accuracy (ACC), normalized mutual information (NMI), and purity (PUR) measure performance.
4.3 Performance Comparison (RQ1)
DealMVC achieves favorable clustering performance across the benchmark comparison, including gains over competing methods on STL10 and strong results on larger datasets.
- DealMVC consistently achieves the most favorable clustering outcomes across the majority of datasets compared with classical deep multi-view clustering algorithms.
- DealMVC obtains better performance than contrastive deep multi-view clustering algorithms through its local and global contrastive strategies.
- DealMVC achieves promising performance on the larger Hdigit and STL10 datasets, supporting its reported generalization capability.
4.4 Ablation Studies (RQ2)
Ablation studies show that DealMVC benefits from both adaptive fusion components and dual contrastive calibration, while pseudo-label consistency supports multi-view class-information consistency.
- Adaptive global fusion module: Removing the attention mechanism, learnable sampling network, or both reduces clustering performance.The attention mechanism and sampling network are jointly optimized to produce a more representative global cross-view feature.
- Dual contrastive calibration module: Removing either the local or global contrastive calibration loss decreases performance, indicating contributions from both modules.
- Dual contrastive calibration module: Training feature structure with reliable class information improves the discriminative capacity of learned features.
- Pseudo-label consistency loss: Removing the pseudo-label consistency loss limits multi-view clustering performance.The loss keeps class information consistent across local and global cross-view feature circumstances by minimizing pseudo-label-graph mean squared error.
4.5 Hyper-parameter Analysis (RQ3)
On Reuters, DealMVC’s performance is notably sensitive to α, while μ has comparatively little influence.
- The hyper-parameter analysis evaluates DealMVC’s trade-off parameters on the Reuters dataset.
- α has a notable influence on DealMVC’s performance.
- μ has a relatively minor impact on DealMVC’s performance.
4.6 Visualization Analysis (RQ4)
Visualization on Caltech101_7 and Hdigit indicates that DealMVC learns embeddings with clearer intrinsic clustering structure than the raw features.
- DealMVC’s learned embeddings are visualized on Caltech101_7 and Hdigit using t-SNE.
- The visualization experiment compares DealMVC’s learned embeddings with raw features.
- Compared with raw features, DealMVC better reveals the intrinsic clustering structure.
5 CONCLUSION
DealMVC combines global and local contrastive calibration to regularize cross-view features using reliable class information. Experiments across eight datasets demonstrate the method’s effectiveness for multi-view clustering.
- DealMVC obtains a global cross-view feature through a fusion mechanism.
- Global calibration aligns view-feature similarity with a high-confidence pseudo-label graph, while local calibration constrains pair-wise view-feature consistency.
- Reliable class information regularizes cross-view local features so similar but different samples can share a cluster.
- The cross-view feature is optimized jointly at local and global levels during training.
- Experiments across eight datasets demonstrate the effectiveness of DealMVC for multi-view clustering.