Multi-View Utility-Based Clustering: A Mutually Supervised Perspective

Zhibin Jiang, Jie Zhou, Shitong WangSymmetry 2025 · DOI

TL;DR

This paper introduces a novel multi-view clustering method (MUC) that leverages "utility-based partition-level side information" from other views to guide the clustering of a current view. This approach, framed as mutual supervision, significantly improves clustering performance by treating cluster assignments as higher-level information, akin to human intuition, and is efficiently solved using a K-means-like optimization.

What Problem Does This Paper Address?

In today's data-rich world, information often comes in multiple "views" or representations of the same underlying objects (e.g., images of a face from different angles, or text and images describing a product). Multi-view clustering aims to combine these diverse perspectives to achieve more robust and accurate clustering than using any single view alone. The core challenges are: 1) effectively extracting and utilizing complementary information between different views to enhance the clustering of each individual view, and 2) intelligently merging the clustering results from multiple views. Existing methods often struggle with how to best leverage this inter-view information, either by simply concatenating features or by using less intuitive sample- or feature-level information.

Methods

The paper proposes the Multi-View Utility-Based Clustering (MUC) method. Its key innovations are:

  • Utility-Based Partition-Level Side Information: Instead of relying on raw features or individual sample similarities, MUC uses the partition assignments (i.e., which cluster each data point belongs to) from other views as a form of "side information." This partition-level information is considered more "instructional" and aligned with human intuition for clustering. A "categorical utility function" quantifies the agreement between a current view's partition and the aggregated partition from other views.
  • Mutual Supervision: MUC employs a novel mutual supervision mechanism. The utility-based partition-level information from other views acts as auxiliary supervision for the current view's clustering. Conversely, the refined clustering results of the current view then become the auxiliary information for other views. This process is iterated until convergence.
  • K-means-like Optimization: The objective function, incorporating both data fidelity and the mutual supervision term, is solved efficiently using an alternating optimization strategy that resembles the K-means algorithm.
  • Multi-View Weight Learning: To account for the varying importance of different views, MUC incorporates a maximum entropy-based scheme to learn weights for each view, further refining the final clustering.

Datasets

The proposed MUC method was evaluated on a variety of datasets:

  • Synthetic Dataset: A 3D synthetic dataset was generated and mapped into 2D subspaces to create multi-view data.
  • UCI Datasets: Seven multi-view datasets from the UCI Machine Learning Repository were used, covering domains like Dermatology, Forest Type, Image Segmentation, Iris, Multiple Features, SPECTF, and Water Treatment.
  • Real-Life Image Datasets:
    • CMU PIE: A facial image dataset consisting of five poses (views) was used to create five combinational multi-view datasets.
  • AwA (Animals with Attributes): A dataset of animal images with pre-extracted feature representations (views) was utilized.

Key Insights and Ideas

  • Partition-Level Information is Powerful: The paper argues that using cluster assignments (partitions) from other views is more effective than using raw features or sample-level information. This aligns with how humans often group things based on perceived categories.
  • Mutual Supervision Enhances Robustness: The iterative process where views "teach" each other leads to more stable and accurate cluster assignments, as it leverages complementary information across the entire dataset.
  • K-means Adaptability: The ability to solve a complex multi-view problem with a K-means-like framework is a significant practical advantage, making the method more accessible and computationally efficient.
  • View Weighting Matters: Not all views contribute equally. Learning weights for each view allows the model to prioritize more informative perspectives, leading to better overall performance.

Key Results

  • Superior Performance: MUC consistently outperformed both single-view clustering methods (K-means, ComBKM) and several state-of-the-art multi-view clustering methods (Co-FKM, MVKKM, MVSpec, WV-Co-FCM, TW-K-means, MVASM) across various datasets, as measured by Normalized Mutual Information (NMI) and Rand Index (RI).
  • Competitive or Better Speed: While not explicitly detailed with quantitative speedup figures, the paper states that MUC is comparable to existing methods in terms of running speed, likely due to its K-means-like optimization.
  • Statistical Significance: Friedman tests and post hoc hypothesis tests confirmed statistically significant performance improvements of MUC over most other methods.

Limitations and Future Work

  • Parameter Sensitivity: The paper acknowledges that the performance of MUC is sensitive to the trade-off parameters $\eta$ and $\lambda$. While grid search was used for parameter tuning, this can be computationally expensive, especially for large datasets.
  • Extension to Other Methods: The authors suggest extending the utility-based partition-level side information and mutual supervision framework to other classical clustering methods, such as Fuzzy C-Means (FCM).
  • Computational Efficiency: Developing more efficient methods for parameter determination and speeding up the optimization procedure for large-scale datasets is highlighted as an important area for future research.

Summary updated

Abstract

Information in multiple views is typically symmetrical and distinctive. Multi-view clustering generally attempts to address two key issues: (1) how to mine/leverage additional useful information between different views for current view clustering, and (2) how to combine multiple clustering results from multiple views by using compatible and complementary information hidden in multiple views. In order to achieve this, we propose a multi-view clustering method, namely the multi-view utility-based clustering method (MUC), from the novel perspective of utility-based mutual supervision between different views. The proposed method, MUC, has notable merits: (1) It moves multi-view clustering from the level of feature and/or sample side to partition side information. That is, in order to handle the first issue, MUC considers how to use utility-based partition-level side information from all the other views for current view clustering. Such partition-level side information is consistent with human thinking; it is high-caliber and instructional for clustering. (2) The utility-based partition-level side information provided by other views complements current view information in a mutually supervised way. Then, we conduct multi-view clustering via a mutual supervision mode to circumvent the second issue. As a result, using an alternating optimization strategy, the objective function of MUC can be solved in a K-means-like way. Moreover, we leverage multi-view weight learning based on maximum entropy to integrate multi-view clustering results and further improve performance. The extensive experimental results on various multi-view datasets indicate that the proposed method is better than—or at least comparable to—the existing commonly used single- and multi-view clustering methods, in terms of both clustering performance and running speed.
BibTeX
@inproceedings{jiang2025multi,
  author = {Zhibin Jiang and Jie Zhou and Shitong Wang},
  title = {{Multi-View Utility-Based Clustering: A Mutually Supervised Perspective}},
  booktitle = {Symmetry},
  year = {2025},
  doi = {10.3390/sym17060924},
}
Figure