Tensorial Multi-view Clustering with Deep Anchor Graph Projection

Wei Feng, Dongyuvan Wei, Qianqian Wang, Bo DongIJCAI 2025 · DOI

TL;DR

This paper introduces TMVC-DAGP, a novel multi-view clustering method that leverages deep learning for feature extraction and tensor-based fusion to capture complex inter-view relationships. By projecting anchor graphs through multiple layers and incorporating sparsity constraints, TMVC-DAGP achieves state-of-the-art performance and robustness without relying on post-processing.

What Problem Does This Paper Address?

Multi-view clustering (MVC) aims to group data points based on information from multiple sources (views). While existing MVC methods have made progress, they often struggle with:

  • Deep Structure Extraction: Traditional methods tend to overlook the hierarchical, deep features within each view, relying on shallow representations.
  • Information Loss and Degradation: Many approaches require post-processing to derive final clustering results, which can introduce errors and reduce performance.
  • Inefficient View Fusion: Effectively leveraging the complementary information across different views remains a challenge, with some methods failing to fully exploit these relationships.
  • Scalability and Robustness: Handling high-dimensional data and noise can degrade performance.

This paper addresses these limitations by proposing a method that integrates deep learning for feature extraction, tensor-based fusion for inter-view complementarity, and anchor graph projection for efficiency and robustness.

Methods

The proposed method, Tensorial Multi-view Clustering with Deep Anchor Graph Projection (TMVC-DAGP), introduces several key innovations:

  • Deep Anchor Graph Projection: Instead of a single projection, TMVC-DAGP employs a multi-layer projection strategy. This allows the model to progressively extract deeper, hierarchical latent features from anchor graphs, which represent a subset of the data for efficiency.
  • Sparsity Constraint (l2,p-norm): A sparsity constraint is applied to the projection matrices. This helps to filter out noise and redundancy, leading to more robust feature selection and eliminating the need for post-processing to obtain clustering results.
  • Tensorial Fusion (Weighted Tensor Schatten p-norm): To effectively capture the consistent and complementary information across different views, TMVC-DAGP utilizes a weighted tensor Schatten p-norm regularization. This tensor-based approach is more powerful than traditional matrix-based fusion for modeling higher-order correlations.
  • Anchor Graph Construction: The method uses anchor graphs, which are constructed from a subset of representative samples, to reduce computational complexity and improve scalability, especially for large datasets.

The optimization of TMVC-DAGP is handled using the Augmented Lagrange Multiplier (ALM) method, with an alternating direction minimization strategy for updating various variables.

Datasets

The proposed method was evaluated on eight widely-used multi-view clustering benchmark datasets:

  • Yale
  • BBCSport
  • Sonar
  • Vehicle Sensor
  • NGs
  • WebKB
  • MSRC
  • RGB-D

No new datasets were introduced in this paper.

Key Insights and Ideas

  • Deep Projection is Crucial: The multi-layer deep projection strategy is key to TMVC-DAGP's success, enabling the extraction of richer, hierarchical features that shallow methods miss. This leads to better representation of complex data structures.
  • Sparsity for Robustness and Efficiency: The l2,p-norm sparsity constraint is not just for noise reduction; it's instrumental in achieving one-step clustering, bypassing the error-prone post-processing step common in other methods.
  • Tensorization for Inter-View Synergy: Moving from matrix-based to tensor-based fusion allows for a more comprehensive understanding of how different views relate and complement each other, leading to improved overall clustering.
  • Anchor Graphs Balance Efficiency and Performance: The use of anchor graphs effectively addresses the scalability challenge of high-dimensional data without significantly sacrificing clustering accuracy, making the method practical for real-world applications.

Key Results

  • State-of-the-Art Performance: TMVC-DAGP consistently outperforms several state-of-the-art multi-view clustering methods across multiple datasets, achieving superior Accuracy (ACC), Normalized Mutual Information (NMI), and Purity (PUR). For example, on the RGB-D dataset, it achieved an ACC of 72.25%, surpassing baselines. On the Yale dataset, it achieved an ACC of 98.78%.
  • Robustness: The method demonstrates robustness to noise and parameter variations, maintaining effective performance even when parameters are slightly adjusted.
  • Efficiency: By using anchor graphs and avoiding post-processing, TMVC-DAGP offers a computationally efficient solution. The convergence analysis shows rapid convergence, typically within 80-100 iterations.
  • Ablation Studies: Experiments confirm that removing the Schatten p-norm or l2,p-norm constraints leads to a significant drop in performance, highlighting the importance of these components. Increasing the number of layers in the deep projection also generally improves performance.

Limitations and Future Work

  • Parameter Sensitivity: While the paper demonstrates robustness, the performance is still influenced by parameter choices (e.g., anchor rate, regularization parameters). Further research could explore more adaptive or automated parameter tuning strategies.
  • Anchor Selection: The paper relies on existing anchor selection methods. Investigating novel or adaptive anchor selection techniques could further enhance the method's performance and efficiency.
  • Deep Layer Exploration: While the paper explores deep projection, future work could investigate optimal layer architectures and activation functions for different types of multi-view data.
  • Theoretical Guarantees: While empirical results are strong, further theoretical analysis on the convergence and optimality of the proposed deep projection and tensor fusion could be beneficial.

Summary updated

BibTeX
@inproceedings{feng2025tensorial,
  author = {Wei Feng and Dongyuvan Wei and Qianqian Wang and Bo Dong},
  title = {{Tensorial Multi-view Clustering with Deep Anchor Graph Projection}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/571},
}
Figure