Source-linked AI summary
Incomplete Contrastive Multi-View Clustering with High-Confidence Guiding
Guoqing Chao, Yi Jiang, Dianhui Chu
TL;DR
Incomplete multi-view clustering must handle missing values while using both consistent and complementary information, yet prior methods often separate representation learning from clustering. ICMVC addresses these gaps with relation transfer, GCN encoding, attention and contrastive learning, high-confidence guidance, and joint optimization; experiments on five datasets verified its effectiveness and superiority, while broader view counts increase computational complexity.
Problem
Incomplete multi-view clustering must address missing values, underused complementary information, and the separation of representation learning from clustering.
Method
ICMVC uses consistency relation transfer with GCN, attention fusion, contrastive learning, high-confidence guidance, and an end-to-end joint framework.
Results
Experiments on five datasets verified the effectiveness and superiority of ICMVC.
Takeaways & Limitations
The unified framework obtains clustering results without pre-training or post-processing.
Takeaways & Limitations
Extending the method to more views increases computation complexity dramatically, and the current work weights its three losses equally.
Abstract
from arXiv · showhide
Incomplete multi-view clustering becomes an important research problem, since multi-view data with missing values are ubiquitous in real-world applications. Although great efforts have been made for incomplete multi-view clustering, there are still some challenges: 1) most existing methods didn't make full use of multi-view information to deal with missing values; 2) most methods just employ the consistent information within multi-view data but ignore the complementary information; 3) For the existing incomplete multi-view clustering methods, incomplete multi-view representation learning and clustering are treated as independent processes, which leads to performance gap. In this work, we proposed a novel Incomplete Contrastive Multi-View Clustering method with high-confidence guiding (ICMVC). Firstly, we proposed a multi-view consistency relation transfer plus graph convolutional network to tackle missing values problem. Secondly, instance-level attention fusion and high-confidence guiding are proposed to exploit the complementary information while instance-level contrastive learning for latent representation is designed to employ the consistent information. Thirdly, an end-to-end framework is proposed to integrate multi-view missing values handling, multi-view representation learning and clustering assignment for joint optimization. Experiments compared with state-of-the-art approaches demonstrated the effectiveness and superiority of our method. Our code is publicly available at https://github.com/liunian-Jay/ICMVC.
1 Introduction
Incomplete multi-view clustering addresses data with missing values, but existing methods often underuse complementary information and separate missing-data handling from clustering. ICMVC combines relation transfer, graph encoding, attention, contrastive learning, high-confidence guidance, and joint optimization in an end-to-end framework.
- Incomplete multi-view clustering seeks a final clustering result from multi-view data containing missing values.
- Existing methods often ignore complementary information while exploiting consistent information across views.
- Deep incomplete multi-view methods commonly separate missing-data handling from clustering, limiting their reported performance.
- ICMVC transfers multi-view consistency relations and uses GCN to encode incomplete data and learn representations for missing values.
- The method combines attention-based fusion, view-specific contrastive learning, high-confidence guidance, and joint optimization of missing-data handling, representation learning, and clustering.
2 The proposed method
The method constructs view graphs from similarities and transfers available adjacency relations to missing views under a consistency assumption. GCN encoders then use graph structure and observed features to learn latent representations, with missing representations estimated through neighbor message passing.
- 2.1 Notations: Each view is represented by a feature matrix, while its graph structure is represented by an adjacency matrix.
- 2.1 Notations: The method discusses two views for convenience and states that the model can extend to more than two views.
- 2.2 Multi-View Missing Handling: View-specific similarity matrices are computed and K-nearest neighbors are used to construct the graph structure.
- Multi-View Consistency Relation Transfer: When a view is missing, its adjacency relation can be transferred from an existing view of the same instance under the cross-view consistency assumption.
- Graph Encoder Module: GCN encoders process each view's adjacency matrix and features, estimating missing latent representations from aggregated information of neighboring instances.
- Multi-View Consistency Relation Transfer: For more than two views, the union or intersection of graph structures from existing views can be transferred to a missing view.
2.3 Network Architecture
The network combines view-specific graph encoders, instance-level attention and contrastive alignment, and a shared pseudo-classifier with high-confidence guidance for clustering.
- Network components: The architecture contains graph encoder, multi-view fusion, and clustering modules, with high-confidence guidance designed to exploit complementary information.
- Graph Encoder Module: Each view-specific stacked GCN encoder maps its adjacency matrix and incomplete feature matrix to a latent representation while capturing node attributes and graph structure.The encoder uses skip connections to avoid trivial solutions.
- Multi-View Fusion Module: Instance-level attention learns per-instance fusion weights for complementary view information, while contrastive learning aligns same-instance representations across views.Positive pairs use different views of the same instance; different instances form negative pairs.
- Clustering Module: A shared pseudo-classifier produces soft cluster assignments for each view and the fused representation, whose columns form cluster-wise Assignment Statistics Vectors.The cluster-level contrastive objective encourages different-cluster exclusivity and same-cluster cross-view consistency.
- Clustering Module: High-confidence guidance selects the highest-probability assignment as an auxiliary target distribution, emphasizing confident assignments and blurring cluster-boundary instances.The training procedure jointly computes representations, assignments, contrastive losses, and the high-confidence-guidance loss.
2.4 Objective Function
The objective function is optimized in an end-to-end clustering model that jointly trains the full pipeline without separate pre-training or a final k-means stage.
- Objective Function: The model jointly optimizes missing-view handling, representation learning, and clustering assignment through a single total loss.The authors state that equal loss weights achieved good performance.
- Objective Function: The end-to-end procedure requires neither pre-training to complete missing views or initialize cluster centroids nor additional k-means clustering for final results.
3 Experiments
Experiments evaluate ICMVC on five datasets using standard clustering metrics, comparisons with single-view and incomplete multi-view baselines, missing-rate studies, ablations, visualizations, and convergence analysis. ICMVC generally outperforms the compared methods, while performance declines at very high missing rates.
- Experimental settings: Experiments use five datasets, three clustering metrics, and comparisons against two single-view and six state-of-the-art incomplete multi-view methods.The metrics are ACC, NMI, and ARI; higher values indicate better clustering performance.
- Baselines and experimental results: At missing rate 0.3 on MSRC-V1, ICMVC exceeds state-of-the-art baselines by 10.96% ACC, 8.46% NMI, and 13.21% ARI.The reported comparison uses the incomplete setting with missing rate η = 0.3.
- Performance with different missing rates: As the missing rate increases, ICMVC remains strong at low rates but degrades quickly above 0.5.The authors attribute the decline to difficulty obtaining the correct graph structure through multi-view consistent relation transfer.
- Baselines and experimental results: ICMVC has relatively small standard deviation compared with deep-learning baselines, which the authors associate with lower dependence on initialization and training randomness.The error-band padding represents standard deviation.
- Ablation study: Removing modules in the ablation study shows that each module contributes to performance, with clustering and guidance handled jointly when the clustering loss is removed.When Lclu is removed, k means is used because the pseudo-classifier no longer provides valid clustering results.
- Visualization analysis: The learned fused representation produces more structured similarity patterns than individual views, while the latent-space visualization shows fewer overlaps than DAIMC.The fused similarity matrix exhibits high within-cluster and low between-cluster similarity.
- Convergence analysis: During convergence, the loss drops rapidly in the first 100 epochs, while ACC, NMI, and ARI rise and eventually converge.Later training shows slower loss decrease with fluctuations and continued metric improvement.
4 Conclusion
ICMVC unifies missing-value handling, representation learning, and clustering in an end-to-end framework with high-confidence guidance. Experiments on five datasets support its effectiveness, while future work targets scalability and adaptive loss weighting.
- ICMVC combines multi-view consistency relation transfer, GCN-based missing-value handling, instance-level attention, and high-confidence guidance in one framework.
- The unified model obtains clustering results without pre-training or post-processing.
- Experiments on five datasets verified the effectiveness and superiority of ICMVC.
- The method can extend beyond two views, but computational complexity increases dramatically as the number of views grows.
- Future work may explore adaptive weighting instead of treating the three losses equally.
A Datasets
The evaluation uses five widely used public datasets, whose statistics are summarized in Table 3.
- The evaluation includes HandWritten, Scene-15, LandUse-21, MSRC-v1, and Noisy MNIST.
- Table 3 presents the statistics of the five datasets.
B Visualization
Visualization analyses show that ICMVC’s learned representations develop clearer cluster structure during training and compare favorably with other methods.
- As training progresses, the fusion representation becomes more conducive to clustering, with separated clusters and compact within-cluster structure emerging.
- Compared with other methods, ICMVC learns a representation with clearer cluster structure that is more conducive to clustering.
- Figure 8 visualizes latent representations obtained at different training epochs on MSRC-V1 with missing rate 0.3.
C.1 Experimental setup
The experiments standardize settings across methods because competing approaches use different randomness and missing-data protocols. A t-SNE comparison is also included for MSRC-V1 at missing rate 0.3.
- Figure 9 visualizes latent representations obtained by different methods on MSRC-V1 with missing rate 0.3.
- The experiments reset compared methods to a common protocol using differing model initialization and the same batch of missing data.
- Some methods fix model randomness and randomly remove data, whereas others reinitialize models while using the same missing-data batch.
C.2 Experimental results
The experiments report full results across five datasets, including standard deviations from five runs and performance across missing rates on other datasets. ICMVC performs well under low missing rates, while Table 4 identifies the best results and unavailable results caused by non-negative constraint violations.
- Experimental results: Five runs per dataset were used to calculate standard deviations for the full experimental results.The complete results are provided in Table 4.
- Experimental results: ICMVC performs well on other datasets under low missing rates, as shown by performance plots across missing-rate settings.Figure 10 reports clustering performance as the missing rate increases.
- Experimental results: Table 4 marks the first- and second-best clustering results in bold and underlining, respectively.A dash indicates results unavailable because of non-negative constraint violations.