Source-linked AI summary
GCFAgg: Global and Cross-view Feature Aggregation for Multi-view Clustering
Weiqing Yan, Yuanyang Zhang, Chenlei Lv, Chang Tang, Guanghui Yue, Liang Liao, Weisi Lin
TL;DR
Multi-view clustering needs consensus representations that account for relationships across all samples rather than relying only on view-wise aggregation, especially amid noisy or missing views. GCFAggMVC combines global and cross-view feature aggregation with structure-guided contrastive learning, and experiments report strong performance on both complete and incomplete clustering tasks.
Problem
Existing deep multi-view clustering methods commonly use view-wise aggregation and sample-level contrastive pairs, overlooking global sample relationships and misclassifying same-cluster samples as negatives.
Method
GCFAggMVC learns view-specific features with autoencoders, aggregates features across samples and views using global relationships, and aligns consensus and view-specific representations with structure-guided contrastive learning.
Results
GCFAggMVC achieves state-of-the-art performance in both complete and incomplete multi-view clustering tasks.
Takeaways & Limitations
The representation module is flexible enough to plug into other frameworks for incomplete multi-view clustering.
Abstract
from arXiv · showhide
Multi-view clustering can partition data samples into their categories by learning a consensus representation in unsupervised way and has received more and more attention in recent years. However, most existing deep clustering methods learn consensus representation or view-specific representations from multiple views via view-wise aggregation way, where they ignore structure relationship of all samples. In this paper, we propose a novel multi-view clustering network to address these problems, called Global and Cross-view Feature Aggregation for Multi-View Clustering (GCFAggMVC). Specifically, the consensus data presentation from multiple views is obtained via cross-sample and cross-view feature aggregation, which fully explores the complementary ofsimilar samples. Moreover, we align the consensus representation and the view-specific representation by the structure-guided contrastive learning module, which makes the view-specific representations from different samples with high structure relationship similar. The proposed module is a flexible multi-view data representation module, which can be also embedded to the incomplete multi-view data clustering task via plugging our module into other frameworks. Extensive experiments show that the proposed method achieves excellent performance in both complete multi-view data clustering tasks and incomplete multi-view data clustering tasks.
1. Introduction
Multi-view clustering must fuse heterogeneous views despite noise, redundancy, and missing information, while existing contrastive approaches can treat same-cluster samples as negatives. GCFAggMVC addresses these issues through global and cross-view aggregation and structure-guided contrastive learning.
- Multi-view clustering unsupervisedly fuses multiple views to differentiate crucial groupings, but remains challenging.
- Existing traditional methods often operate on noisy and redundant original or kernel features, harming clustering.
- Deep clustering methods learn view-specific representations and fuse them into a consensus representation, while alignment models reduce view-private information.
- Sample-level contrastive learning can incorrectly treat representations from different samples in the same cluster as negative pairs.
- GCFAggMVC learns global similarity relationships and enhances consensus representations using structurally related samples across views.
- The module also applies to incomplete multi-view clustering, where structurally related samples can enhance representations for samples with missing views.
2. RETATED WORK
Representation-learning-based multi-view clustering includes shallow graph and subspace methods and newer deep methods using nonlinear transformations, adversarial alignment, attention, or weighted fusion. These approaches construct consensus representations through graph fusion, self-representation, or view-wise aggregation.
- Shallow multi-view clustering methods comprise multi-view graph clustering and multi-view subspace clustering.
- Multi-view graph clustering: Graph-based methods construct view-specific graphs, fuse them with regularization terms, and cluster the resulting graph.
- Multi-view subspace clustering: Subspace-based methods learn a consensus self-representation matrix and partition data into different subspaces using regularization.
- Deep representation learning-based methods: Deep representation-learning methods use neural networks to model nonlinear mappings and learn embedded feature representations.
- Deep representation learning-based methods: Some deep methods align hidden-representation distributions adversarially, weight views, or combine view representations with norm constraints.
3. PROPOSED METHOD
GCFAggMVC reconstructs view-specific features, aggregates cross-view information using global sample relationships, and aligns consensus with view-specific representations through structure-guided contrastive learning.
- 3.1. Multi-view Data Reconstruction: Autoencoders project each view into low-dimensional representations and reconstruct the inputs to reduce redundancy and random noise.The encoder computes z_i^v from x_i^v, while the decoder reconstructs x̂_i^v; reconstruction loss helps avoid model collapse.
- 3.2. Global and Cross-view Feature Aggregation: GCFAggMVC concatenates view-specific representations and transforms them into a shared space for cross-view feature fusion.The concatenated matrix Z contains all V views, with d = d_v ∗ V, and W_R performs the cross-view transformation.
- 3.2. Global and Cross-view Feature Aggregation: The module computes sample relationships with S = softmax(Q_1Q_2^T/√d) and uses them to enhance each sample representation with correlated samples.Each enhanced representation is a weighted sum of rows of R, allowing samples with high relationships to contribute to one another.
- 3.2. Global and Cross-view Feature Aggregation: A nonlinear and linear transformation processes the original and structure-enhanced representations to produce the consensus representation Ĥ while reducing redundancy.The transformation applies learned weights, biases, and a ReLU operation to Z + Ẑ.
- 3.3. Structure-guided Contrastive Learning: Structure-guided contrastive learning aligns consensus and view-specific representations while weighting negative samples according to their global structure relationships.The method minimizes similarity for different samples with low structure relationships and aligns positive pairs with high relationships, unlike sample-level contrastive learning.
4. Experiments
Experiments evaluate GCFAggMVC on complete and incomplete multi-view clustering benchmarks, using ACC, NMI, and PUR against traditional and deep baselines. Results report stronger clustering performance, convergence, parameter robustness, and component effectiveness, with clearer separated consensus-feature clusters.
- Experimental settings: Experiments use 13 public multi-view datasets and evaluate clustering with ACC, NMI, and PUR.Comparisons include traditional and deep methods for complete and incomplete multi-view datasets.
- Comparative results: GCFAggMVC achieves better results than nine compared methods across eight benchmark datasets using ACC, NMI, and PUR.The compared methods include four traditional and five deep multi-view clustering methods.
- Comparative results: On Caltech, GCFAggMVC outperforms other deep clustering methods when the number of views increases.Table 4 reports comparisons for different numbers of views, with “XV” denoting the number of views.
- Comparative results: GCFAggMVC performs better than other methods on incomplete datasets and obtains the best results when the missing view rate is 0.7.Incomplete samples are generated by randomly removing views while retaining at least one view per sample, with ratios from 0.1 to 0.7.
- Model analysis: The objective value decreases monotonically until convergence, while ACC, NMI, and PUR increase before fluctuating within a narrow range.The reported behavior confirms convergence of GCFAggMVC.
- Model analysis: Removing GCFAgg or SgCL lowers ACC by 18.22, 6.93, and 2.03 percent, or by 14.21, 15.23, and 30.07 percent, respectively.The ablation results attribute the differences to view-private information, noise and redundancy, and structure-guided contrastive alignment.
5. Conclusion
GCFAggMVC combines global similarity learning, consensus enhancement, and structure-guided contrastive learning for multi-view clustering. Experiments report SOTA performance on both complete and incomplete MVC tasks.
- GCFAggMVC learns global sample similarity, enhances consensus representations using structurally related samples, and applies SgCL to reduce same-cluster representation inconsistency.The framework uses view-specific autoencoder features, a GCFAgg module, and a structure-guided contrastive learning module.
- Extensive experiments show SOTA performance in both complete and incomplete multi-view clustering tasks.
- The network reconstructs original data through view-specific autoencoder features before learning consensus representations.