TL;DR
LargeMvC-Net bridges the gap between scalable shallow anchor-based clustering and expressive deep neural networks by unfolding a robust multi-view optimization problem into an interpretable, layer-wise architecture. It achieves state-of-the-art multi-view clustering performance on large-scale datasets while maintaining linear computational complexity and high robustness to noise and missing data.
What Problem Does This Paper Address?
Multi-view clustering (MVC) aims to group data points by integrating information across multiple sources or modalities (such as text, images, and audio). As datasets grow exponentially, scalability becomes a major bottleneck.
To tackle this, researchers typically use one of two approaches:
- Shallow anchor-based methods: These use a small set of representative "anchors" to approximate massive sample-to-sample relationships, reducing computational overhead. However, their shallow, linear nature prevents them from capturing complex, non-linear semantic correlations.
- Deep multi-view clustering: These leverage deep neural networks to learn highly expressive latent features but typically require expensive global similarity computations, making them computationally prohibitive for large-scale datasets.
While recent hybrid methods attempt to combine deep networks with anchor structures, they do so heuristically (e.g., via post-hoc graph construction or message passing). This detaches the network's architecture from the underlying clustering principles, leading to a lack of interpretability and suboptimal optimization.
Methods
The authors introduce LargeMvC-Net, a Deep Unfolding Network (DUN) designed specifically for large-scale multi-view clustering. Instead of treating the neural network as a black box, they mathematically formulate a robust anchor-based optimization problem and map its iterative, step-by-step solution directly onto a feed-forward network architecture.
To handle real-world corruptions, the optimization problem incorporates an $\ell_1$-norm for representation sparsity and an $\ell_{2,1}$-norm for view-specific noise suppression. Each stage (or layer) of LargeMvC-Net directly mirrors one iteration of this optimization routine, split into three interpretable modules:
- RepresentModule: Aggregates multi-view residuals and updates the shared, consistent clustering representation across views using a learnable soft-thresholding operator.
- NoiseModule: Isolates sample-level noise and corruptions unique to each view using a learnable row-wise shrinkage operator.
- AnchorModule: Aligns each view's anchor structural pattern to the latent space using Singular Value Decomposition (SVD) under strict orthogonality constraints.
The entire network is trained end-to-end using an unsupervised reconstruction loss that aligns each view with the anchor-induced latent space.
Datasets
The model was extensively evaluated on 11 benchmark multi-view datasets, spanning both complete and incomplete (missing view) settings:
- Feature-level datasets: Animals, Caltech102, Cifar10, MNIST, NUSWIDEOBJ, YoutubeFace, YTF-50, and YTF-100 (the largest, containing 195,537 samples).
- Modality-level (vision and language) datasets: ESP-Game, Flickr, and IAPR.
Key Insights and Ideas
- Optimization-Driven Architecture: By explicitly translating mathematical optimization steps into network modules, LargeMvC-Net provides structural clarity and optimization traceability that generic autoencoders or Graph Convolutional Networks (GCNs) lack.
- Linear Scalability ($O(n)$): By routing global similarity through a compact set of anchors ($m \ll n$), the network completely avoids expensive pairwise sample computations. This allows deep multi-view clustering to scale linearly with the number of samples.
- Principled Denoising: Rather than relying on auxiliary contrastive losses, embedding the noise suppression matrix directly into the unfolded architecture allows the network to naturally tolerate view-specific corruptions and missing data.
Key Results
- State-of-the-Art Accuracy: LargeMvC-Net consistently outperformed both shallow and deep baselines across all datasets in Clustering Accuracy (ACC), Normalized Mutual Information (NMI), and Adjusted Rand Index (ARI). On the massive YTF-100 dataset, it achieved a top ACC of 76.52%.
- Robustness to Missing Views: In incomplete multi-view scenarios (where views were randomly dropped at rates ranging from 10% to 90%), LargeMvC-Net maintained highly stable performance, significantly outperforming specialized incomplete clustering methods.
Limitations and Future Work
The current iteration of LargeMvC-Net relies on a fixed number of anchors initialized via k-means. The authors note that incorporating dynamic anchor structures that can adaptively update and evolve during the network's training process could further enhance model flexibility. Additionally, they plan to explore multi-task learning frameworks to extend the model's utility.
Summary updated