Abstract
Multi-view Clustering via Multi-granularity Ensemble (MGE) is a novel framework designed to enhance data grouping from multiple sources by exploiting hierarchical structures. Traditional multi-view clustering (MVC) methods often face challenges with heterogeneous views or information loss from single-granularity labels. MGE overcomes these by extracting "fine-to-coarse" patterns across views and fusing them into a robust co-association matrix. Evaluation on diverse benchmarks shows MGE achieves state-of-the-art accuracy while maintaining high computational efficiency.
Introduction: The Challenge of Multi-view Data
Multi-view clustering (MVC) aims to integrate information from diverse modalities—such as visual features and textual descriptions—to create more robust clusters than single-view methods. However, existing techniques typically fall into two categories, both with significant drawbacks:
- Early Fusion (Representation/Graph-based): These methods attempt to merge features or similarity graphs at the start. They often struggle with heterogeneous views where data distributions differ wildly, requiring complex alignment or manual weight tuning.
- Late Fusion (Ensemble-based): These methods run clustering on each view independently and seek a consensus. This often leads to information loss because they rely on "single-granularity" labels, ignoring the rich hierarchical relationships (fine-grained vs. coarse-grained structures) inherent in the data.
MGE addresses these issues by capturing the "local and global" patterns that single-level labels miss, providing a more comprehensive representation of the underlying data structure.
Methodology: The MGE Framework
The MGE framework operates through a three-stage pipeline designed to preserve and fuse multi-level information:
1. Multi-granularity Label Generation
Instead of selecting a single "best" clustering result per view, MGE employs Constrained Hierarchical Clustering (CHC). CHC is a specialized version of hierarchical clustering that utilizes adjacency constraints and cluster-size priorities. This ensures high-purity merges, generating a spectrum of labels from many small (fine-grained) clusters to a few large (coarse-grained) clusters for every view.
2. Cross-view and Cross-granularity Fusion
The framework constructs a comprehensive "cluster-wise similarity graph" where nodes represent clusters from any view at any granularity level. The Jaccard coefficient measures the overlap between these clusters. To uncover hidden connections and capture indirect relationships between data points, MGE applies a random walk propagation process.
3. Co-association Matrix and Final Clustering
The propagated cluster similarities are mapped back to the object level to form a co-association matrix. This matrix serves as a refined similarity map for the entire dataset. Finally, CHC is applied to this matrix to produce the final, high-quality clustering result.
Experimental Results and Performance
The authors evaluated MGE against ten state-of-the-art competitors (including GMC, CoReg, and FMVEC) on one synthetic and six real-world benchmarks, such as 100Leaves, UCI-digits, and COIL20.
Accuracy and Robustness
MGE consistently outperformed all competitors, achieving an average accuracy (ACC) of 95.57%. Notably, it reached a perfect 100% accuracy on the COIL20 dataset. Compared to the best existing ensemble method (FMVEC), MGE showed a 17.62% improvement in average ACC, and a 10.07% lead over the top graph-fusion method (CoReg).
Computational Efficiency
Despite the complexity of handling multiple granularity levels, MGE is remarkably efficient. By avoiding iterative optimization and complex graph alignments, its runtime remains significantly lower than that of competitors like CoReg or V3H as the dataset size increases.
Discussion and Future Directions
MGE's success stems from its "fine-to-coarse" approach, allowing strong levels in one view to compensate for noise in another. It also eliminates the need for manual weight tuning by naturally identifying consistent structures across views.
However, some limitations remain. The framework relies on a hyperparameter, $\lambda$, to control the number of granularity levels. While the paper demonstrates a stable range (around 0.5), full automation is a goal for future work. Additionally, the current model assumes complete views for all samples; extending MGE to handle "missing views" and improving scalability for extremely large-scale datasets are key areas for further research.
Summary updated