Abstract
Multi-view clustering aims to integrate information from diverse sources to improve data partitioning. Traditional Multi-view K-means (MVKM) often neglects cluster separability and the local geometric structure of data. We present DFMKLS (Discriminatively Fuzzy Multi-View K-means Clustering with Local Structure Preserving), a novel framework that optimizes for cluster compactness, separability, and local data geometry simultaneously. By incorporating fuzzy logic and manifold learning, DFMKLS achieves superior performance on benchmark datasets compared to state-of-the-art methods.
Introduction
Clustering multi-view data—where objects are described by multiple feature sets (e.g., images and their textual tags)—is a critical task in unsupervised learning. Standard MVKM approaches typically focus on global compactness, minimizing within-cluster variance. However, they often fail to ensure cluster separability, leading to overlapping clusters, and ignore local relationships where neighboring points in a specific view should likely share cluster assignments.
Methodology
The DFMKLS framework addresses these limitations through three core components:
- Fuzzy Multi-View K-means: Unlike hard assignments, fuzzy clustering provides a probabilistic membership for each data point, effectively handling data ambiguity.
- Discriminative Property: The objective function explicitly maximizes the distance between pairwise cluster centers. This "push-pull" mechanism pulls points toward their centers while pushing centers apart to enhance separability.
- Local Structure Preservation: By constructing a K-nearest-neighbor (KNN) graph for each view, the model incorporates manifold learning. A penalty is applied if neighboring points in the feature space are assigned to different clusters.
The resulting non-convex optimization problem is solved using an efficient iterative update algorithm with a multiplicative update rule.
Experimental Evaluation
The model was tested on four benchmark datasets: 3Sources (news stories), BBC (segmented documents), WebKB (university webpages), and NUS_WIDE (web images).
Parameter Sensitivity
The performance of DFMKLS is influenced by the number of neighbors $K$ and the trade-off parameter $\alpha$. As shown in Figure 1, the algorithm demonstrates relative stability across a range of $K$ values for various datasets, though optimal performance typically occurs around $K=10$.
Convergence Analysis
Efficiency is a key advantage of the proposed iterative solver. Empirical results indicate that the objective value stabilizes quickly, typically within 100 iterations, making it computationally viable for medium-to-large datasets.
Key Results and Discussion
DFMKLS consistently outperformed 12 state-of-the-art methods (including RMKMC, GMC, and standard Multi-view K-means). Notable highlights include:
- BBC Dataset: Achieved an accuracy of 0.8672, significantly higher than the second-best method (0.8068) and standard multi-view K-means (0.5547).
- Scalability: On the large-scale NUS_WIDE dataset, it maintained superior accuracy (0.2453) and NMI, outperforming graph-based alternatives.
- Robustness: The integration of fuzzy logic and discriminative terms allows for clearer cluster boundaries in high-dimensional spaces.
Conclusion and Future Work
DFMKLS provides a robust framework for multi-view clustering by balancing global and local data structures. While the $O(Cn^2)$ complexity is manageable for medium datasets, future work will explore anchor-based methods to scale to "Big Data." Additionally, automated hyper-parameter tuning and handling incomplete multi-view data remain open areas for investigation.
Summary updated