High-Confident Local Structure Guided Consensus Graph Learning For Incomplete Multi-view Clustering

Shuping Zhao, Lunke Fei, Qi Lai, Jie Wen, Jinrong Cui, Tingting ChaiIJCAI 2025 · DOI

Abstract

Multi-view clustering (MVC) is often hampered by incomplete data across different views and the presence of "edge samples"—noisy data points at cluster boundaries. This paper introduces HLSCG_IMC, a novel framework that integrates high-confidence local structure guidance with a strict block diagonal regularizer. By filtering out low-confidence relationships and leveraging global structural properties, the proposed method learns a robust consensus graph that effectively handles missing information. Experimental results on benchmark datasets demonstrate that HLSCG_IMC consistently outperforms state-of-the-art incomplete multi-view clustering methods.

Introduction

In real-world applications, data is frequently collected from multiple sources or "views" (e.g., text, images, and metadata). While multi-view clustering (MVC) aims to exploit the complementary information between these views, it often encounters the challenge of incomplete data, where some views are missing for certain samples.

Existing Incomplete Multi-view Clustering (IMC) methods struggle with three primary issues:

  1. Information Imbalance: Disparate rates of missing data across views make integration difficult.
  2. Latent Information Loss: Relationships within missing data segments are often ignored.
  3. The "Edge Sample" Problem: Samples at cluster boundaries typically have weak features, leading to blurred boundaries when using standard pairwise affinity graphs.

Methods

The proposed High-Confident Local Structure Guided Consensus Graph Learning (HLSCG_IMC) framework addresses these challenges through a unified optimization approach.

Overview of the HLSCG_IMC framework, illustrating the transition from incomplete views to a high-confidence local structure and finally a consensus graph.

The core components include:

  • Block Diagonal Structure (BDS) Regularizer: Instead of a simple similarity graph, the model learns a "strict block diagonal" representation. In this matrix, entries are non-zero only for samples within the same cluster, forcing the model to differentiate categories more strictly.
  • High-Confident Local Structure Guidance: To solve the edge sample problem, the authors define a high-confidence graph based on Gaussian kernels and $k$-block neighbors. This specifically filters out low-confidence relationships, ensuring that only the most reliable local structures guide the final consensus graph.
  • Unified Joint Optimization: The framework integrates local structure learning, graph completion, and consensus graph learning into a single objective function, solved using the Alternating Direction Method of Multipliers (ADMM).

Experimental Results

The model was evaluated on four benchmark datasets: BBCSport, COIL-20, Caltech-7, and BUAA-visnir, with simulated missing data rates ranging from 10% to 50%.

Accuracy (ACC) performance of HLSCG_IMC compared to a pairwise graph baseline across different missing rates.

Key Findings

  • Superior Accuracy: HLSCG_IMC consistently outperformed ten state-of-the-art IMC methods. For instance, on the BBCSport dataset with 50% missing data, it achieved an accuracy of 78.48%, significantly higher than the runner-up's 70.86%.
  • Robustness to Missing Data: As shown in the performance plots, the gap between HLSCG_IMC and competing methods generally widens as the missing rate increases, suggesting that the high-confidence guidance is particularly effective when data is scarce.
  • Adaptive View Weighting: The model automatically learns a weight for each view, allowing it to prioritize views that are more complete or contain more discriminative information.

Conclusion

HLSCG_IMC provides a robust solution for incomplete multi-view clustering by prioritizing global block-diagonal structures over simple pairwise affinities and utilizing high-confidence local guidance. While the method shows exceptional performance, future work will focus on reducing its $O(n^3)$ computational complexity to improve scalability for large-scale datasets and exploring its application in multi-view classification tasks.

Summary updated

BibTeX
@inproceedings{zhao2025high,
  author = {Shuping Zhao and Lunke Fei and Qi Lai and Jie Wen and Jinrong Cui and Tingting Chai},
  title = {{High-Confident Local Structure Guided Consensus Graph Learning For Incomplete Multi-view Clustering}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/792},
}
Figure