Introduction
Hierarchical clustering (HC) is a cornerstone of unsupervised learning, enabling the discovery of nested structures within data, such as taxonomies or evolutionary lineages. However, modern datasets often consist of multiple "views"—different feature sets describing the same underlying objects (e.g., visual features and textual descriptions). Performing HC on multi-view data introduces significant geometric and architectural hurdles.
Traditional approaches often struggle with the "Space Problem": hierarchies grow exponentially, while Euclidean space expands only polynomially. This mismatch leads to severe distortion when embedding deep trees. Furthermore, the "Integration Problem" persists in many existing frameworks that decouple feature learning from clustering, failing to capture the intricate relationships between views.
The MHCN Framework
To address these issues, we present the Multi-view Hierarchical Clustering Network (MHCN). MHCN is a unified, one-stage deep learning framework that maps multi-view data into a shared hyperbolic space to perform hierarchical clustering.
Hyperbolic Representation Learning
The core of MHCN lies in its use of Hyperbolic Autoencoders (HAE). Unlike standard Euclidean autoencoders, HAEs project data onto the Poincaré ball, a model of hyperbolic geometry characterized by negative curvature. Because the volume of hyperbolic space grows exponentially with its radius, it provides a natural mathematical substrate for tree-like structures, allowing for low-distortion hierarchical embeddings.
Optimization Objectives
MHCN employs a triple-loss objective function to refine the latent space:
- Multi-view Alignment Loss: This term ensures that embeddings of the same object across different views are brought closer together in the Poincaré ball, promoting cross-view consistency.
- Reconstruction Loss: By penalizing reconstruction errors, the model preserves the unique, private information inherent to each specific view.
- Hyperbolic Uniformity Loss: This critical innovation prevents "representation collapse," where embeddings cluster into a tiny region of the ball. By encouraging a uniform distribution, it ensures the resulting hierarchy is balanced and interpretable.
Experimental Results
The model was rigorously tested on seven datasets, ranging from regular-scale (e.g., MNIST-USPS, Caltech) to large-scale (e.g., Multi-Fashion, NR-MNIST with 60,000 samples).
Accuracy and Scalability
MHCN achieved state-of-the-art performance across all benchmarks using the Dendrogram Purity (DP) metric. On the MNIST-USPS dataset, MHCN reached a DP of 99.22%, significantly outperforming the previous best deep method, CMHHC (94.49%), and traditional Ward-linkage (approx. 80%).
In terms of scalability, MHCN demonstrated superior efficiency. On the large-scale NR-MNIST dataset, traditional Hierarchical Agglomerative Clustering (HAC) required over 1,800 seconds, whereas MHCN completed the task in 900 seconds without the "Out of Memory" (OOM) issues common in other deep learning baselines.
The Impact of Uniformity
Ablation studies highlighted the necessity of the Hyperbolic Uniformity Loss. Without this term, embeddings tend to bunch together, leading to skewed and less meaningful dendrograms. When applied, the loss forces the data to utilize the vast periphery of the Poincaré ball, resulting in well-separated clusters and a balanced tree structure.
Conclusion
MHCN represents a significant advancement in multi-view hierarchical clustering. By leveraging the geometric advantages of the Poincaré ball and a unified one-stage learning process, it overcomes the limitations of Euclidean-based methods. Its inductive nature further allows it to generalize to new data points without retraining, making it a robust and scalable solution for complex data analysis tasks.
Summary updated