Deep Incomplete Multi-View Multiple Clusterings

Shaowei Wei, Jun Wang, Guoxian Yu, Carlotta Domeniconi, Xiangliang Zhang

Introduction

In real-world applications, data is inherently multi-view, meaning it can be described by multiple distinct feature sets or sources. For instance, a video consists of visual frames, audio tracks, and textual subtitles. While traditional clustering algorithms aim to find a single optimal grouping, multi-view data often contains multiple valid structures (e.g., grouping songs by genre versus by decade).

Two significant challenges hinder effective multi-view clustering: multiplicity and incompleteness. Most existing algorithms are designed to find only one clustering and assume that all views are available for every data point. In practice, however, sensors often fail or data is lost, leading to incomplete views. This paper introduces DiMVMC (Deep Incomplete Multi-View Multiple Clusterings), the first deep learning framework designed to simultaneously address both challenges.

The DiMVMC Framework

DiMVMC departs from the standard autoencoder paradigm by utilizing an encoder-less architecture. Instead of compressing incomplete input data, the model initializes $M$ shared subspaces ($H^1, \dots, H^M$) that represent different latent structures of the data.

The DiMVMC architecture showing the generation of multiple clusterings from incomplete multi-view data.

As shown in the architecture above, each subspace serves as input to a dedicated group of deep decoder networks. These decoders attempt to reconstruct the original multi-view data. By calculating loss only on the observed data (using an indicator matrix), the model effectively learns to fill in missing information. To ensure that the $M$ resulting clusterings are distinct, the authors incorporate the Hilbert-Schmidt Independence Criterion (HSIC) as a diversity-promoting constraint in the loss function. This penalty prevents the learned subspaces from becoming redundant.

Experimental Results

The model was evaluated on several benchmark datasets, including Caltech20, Handwritten digits, and Reuters documents. A key finding is DiMVMC's exceptional robustness to missing data.

Performance comparison (SC and 1-NMI) across different missing rates on Caltech20, Handwritten, and Reuters datasets.

As illustrated in the performance plots, DiMVMC maintains stable clustering quality (measured by the Silhouette Coefficient, SC) and high diversity (measured by 1-NMI) even as the missing rate increases to 50%. In contrast, shallow matrix factorization methods and other state-of-the-art competitors show significant performance degradation as data becomes more incomplete. The use of deep neural networks allows DiMVMC to capture complex, non-linear correlations between views that simpler models miss.

Conclusion

DiMVMC provides a robust solution for uncovering multiple latent structures in complex, incomplete datasets. By balancing quality and diversity through deep decoders and HSIC constraints, it outperforms traditional methods in both stability and expressive power. Future work will focus on optimizing the HSIC term's quadratic complexity to enable scaling to larger datasets through batch optimization techniques.

Summary updated

BibTeX
@inproceedings{wei2020deep,
  author = {Shaowei Wei and Jun Wang and Guoxian Yu and Carlotta Domeniconi and Xiangliang Zhang},
  title = {{Deep Incomplete Multi-View Multiple Clusterings}},
  booktitle = {ICDM},
  year = {2020},
  doi = {10.1109/icdm50108.2020.00074},
}
Figure