Source-linked AI summary
Deep Embedded Multi-view Clustering with Collaborative Training
Jie Xu, Yazhou Ren, Guofeng Li, Lili Pan, Ce Zhu, Zenglin Xu
TL;DR
Existing multi-view clustering methods face limited representation capability or high computation and space complexity. DEMVC uses deep autoencoders and collaborative training to learn view representations and assignments while exploiting consensus and complementary information. Experiments on popular datasets report significant improvements and state-of-the-art performance, while the paper notes visual overlap among similar Fashion-10K classes.
Problem
Existing multi-view clustering methods have limited representation capability or high computation and space complexities, restricting large-scale data clustering.
Method
DEMVC learns each view’s embedded representation with a deep autoencoder, then collaboratively trains all views using shared auxiliary distributions and consistent cluster-center initialization.
Results
Experiments on several popular datasets report that DEMVC achieves state-of-the-art performance, with better performance than single-view methods in two- and three-view settings.
Takeaways & Limitations
DEMVC provides an O(N)-complexity multi-view clustering framework that uses common and complementary information and can be applied to large-scale data clustering.
Takeaways & Limitations
On Fashion-10K, visually similar product classes remain difficult to separate, and the authors suggest additional information may be needed.
Abstract
from arXiv · showhide
Multi-view clustering has attracted increasing attentions recently by utilizing information from multiple views. However, existing multi-view clustering methods are either with high computation and space complexities, or lack of representation capability. To address these issues, we propose deep embedded multi-view clustering with collaborative training (DEMVC) in this paper. Firstly, the embedded representations of multiple views are learned individually by deep autoencoders. Then, both consensus and complementary of multiple views are taken into account and a novel collaborative training scheme is proposed. Concretely, the feature representations and cluster assignments of all views are learned collaboratively. A new consistency strategy for cluster centers initialization is further developed to improve the multi-view clustering performance with collaborative training. Experimental results on several popular multi-view datasets show that DEMVC achieves significant improvements over state-of-the-art methods.
1 Introduction
The paper motivates DEMVC by limitations in existing multi-view clustering and proposes collaborative deep autoencoder training to exploit common and complementary information across views.
- Motivation: Multi-view clustering combines common information shared across views with complementary information that completes each view’s partial description.Examples include webpage content and linkage, or different views of the same object.
- Limitations of Existing Methods: Existing shallow MVC methods have limited representation capability, while spectral-clustering and matrix-factorization approaches have high computation and space complexities.These limitations prevent such methods from handling large-scale data clustering tasks.
- Proposed Approach: DEMVC trains multiple deep neural networks collaboratively so one view can guide the others toward useful common and complementary information.The approach uses a referred view whose objective guides its own and the other views’ training.
- Proposed Approach: DEMVC first learns embedded representations with deep autoencoders, then collaboratively refines autoencoders and clustering assignments across views.A referred view supplies an auxiliary target distribution for refining all views.
- Proposed Approach: A shared auxiliary-distribution scheme and consistency strategy for cluster-center initialization are developed to improve multi-view clustering performance.These mechanisms are presented as contributions alongside collaborative training.
- Reported Outcomes: Experiments on several popular datasets report state-of-the-art performance, with good representation capability and efficient applicability to large-scale datasets.The paper presents these properties as outcomes of the proposed model.
2 Related Work
Prior work spans deep embedded clustering and several multi-view clustering paradigms; DEMVC extends deep embedded clustering with collaborative training that shares auxiliary distributions across views.
- Deep Clustering: Deep embedded clustering jointly learns cluster assignments and deep autoencoders in a low-dimensional embedded space.Related variants improve embedded-space preservation or use convolutional autoencoders for image data.
- DEMVC Positioning: DEMVC applies deep embedded clustering to multi-view learning through a multi-view collaborative training strategy.The paper uses deep autoencoders to learn representations and clustering assignments for each view.
- Multi-view Clustering: Multi-view clustering research includes CCA-based, subspace, matrix-factorization, graph-based, and autoencoder-based approaches.These methods learn correlated projections, shared or potential subspaces, low-rank factors, graphs, or deep embedded features.
- DEMVC Positioning: Unlike the cited MVC methods, DEMVC alternately shares an auxiliary distribution to enforce consistency across multi-view predictions.The authors report that this strategy mines both common and complementary information and significantly improves clustering performance.
3 Proposed Method
DEMVC combines deep autoencoder representations with collaborative multi-view clustering, using a referred view to guide all views through a shared auxiliary target distribution. It further initializes cluster centers consistently, preserves reconstruction capability during fine-tuning, and averages final soft assignments across views.
- Deep embedded representations: DEMVC encodes and decodes each view with deep autoencoders, producing low-dimensional embedded features and reconstructed samples.The model jointly uses reconstruction and clustering objectives.
- Multi-view collaborative training: Each view becomes the referred view in turn, sharing its auxiliary target distribution to collaboratively train all views and align their clustering assignments.This design uses both common and complementary information while encouraging consistent multi-view predictions.
- Training procedure: DEMVC first pre-trains all autoencoders, then jointly fine-tunes autoencoders, cluster assignments, and cluster centers during collaborative training.Pre-training prevents a randomly initialized referred view from guiding the other views blindly.
- Final prediction: Final predictions average the views’ soft labels, preserving consistent assignments and potentially correcting misaligned predictions when views disagree.The averaging step integrates information from all views rather than selecting one view’s prediction.
- Consistency strategy for cluster centers initialization: Cluster centers are initialized consistently across views by applying k-means to the first referred view and sharing those centers during initialization.The centers are shared only initially; later, each view learns its own centers while the referred view’s auxiliary distribution remains shared.
- Complexity and algorithm: DEMVC has time and space complexities at the same level as IDEC and linear in N, supporting large-scale data clustering.The paper summarizes the complete procedure in Algorithm 1.
4 Experimental Setup
The experiments evaluate DEMVC on several constructed multi-view digit and fashion datasets against multi-view and single-view clustering baselines. Performance is measured with ARI, ACC, and NMI, using convolutional autoencoders and repeated training runs.
- Datasets: The evaluation uses NoisyMnist-RotatingMnist, Mnist-USPS, Fashion-10K, and Mnist-10K as multi-view clustering datasets.NoisyMnist-RotatingMnist contains noisy and rotated digit views; Fashion-10K and Mnist-10K construct additional views from same-label samples.
- Compared methods: DEMVC is compared with multi-view methods including DCCA, DCCAE, DiMSC, LMSC, BMVC, and COMIC.These comparisons are performed on NoisyMnist-RotatingMnist and Mnist-USPS.
- Network settings: All autoencoders use a symmetric convolutional architecture whose embedded layer has 10 neurons.The encoder uses convolutional layers with channels 32, 64, and 128, followed by Fc10; ReLU is used except in specified layers.
- Training settings: DEMVC autoencoders are pre-trained for 500 epochs and fine-tuned for 20,000 iterations with Adam and batch size 256.When each view is the referred view, it guides end-to-end training for 200 batches.
- Evaluation metrics: Results are reported with ARI, ACC, and NMI, where larger values indicate better clustering, and most values average five independent runs.Values excerpted from prior papers are excluded from the five-run averaging statement.
5 Results and Analysis
DEMVC achieves strong multi-view clustering results and benefits from collaborative training and consistent cluster-center initialization. Its decoders and embeddings also reconstruct and organize noisy, transformed, and visually similar samples effectively, although some Fashion-10K classes remain close in feature space.
- Clustering comparisons: DEMVC achieves the best performance in the comparisons on NoisyMnist-RotatingMnist and Mnist-USPS.The comparison includes methods whose high complexity prevents them from solving the 70,000-sample task.
- Clustering comparisons: DEMVC outperforms all single-view methods with two views, while three-view DEMVC generally performs better than its two-view version.This supports extracting useful features from multiple views.
- Reconstruction: On digit images, the autoencoder captures common structure such as the hollow region of handwritten 6s; on Fashion-10K, it emphasizes object contours.The paper relates this emphasis to clothing similarity being based mainly on appearance and shape rather than logos or patterns.
- Reconstruction: DEMVC decoders reconstruct noisy, rotated, and fashionable images from low-dimensional embedded features while removing noise, rotation, and redundant information.The reconstructions can patch missing parts and make samples look more standard.
- Embedded features: t-SNE visualizations show highly concentrated same-class points and well-separated clusters across four datasets.Fashion-10K still contains nearby pullover, coat, and shirt samples because their 28×28 images are visually similar.
- Module analysis: Collaborative training lets views teach each other, increasing clustering performance alternately and outperforming independent IDEC training for all views.The referred view is switched during training.
- Module analysis: Coo+SetC produces much better final clustering performance than Coo, supporting the consistency strategy for initial cluster-center sharing.Other views may begin with low NMI when the referred view’s centers poorly represent them, but later improve cooperatively.
- Module analysis: The ablation study indicates that collaborative training and consistent initial centers are useful, while averaging multiple views’ soft labels improves prediction accuracy.The final prediction uses the averaged soft labels of multiple views.
6 Conclusion
DEMVC uses collaborative training to learn embedded features across views, incorporates consensus and complementary information, and initializes cluster centers consistently. The framework is designed for effective large-scale clustering with O(N) complexity.
- DEMVC collaboratively trains views so each view guides the others in learning embedded features.
- The framework incorporates both common and complementary multi-view information to enhance clustering performance.
- A consistency strategy for cluster-center initialization is introduced for collaborative multi-view clustering.
- DEMVC has O(N) complexity and can be used for large-scale data clustering.