Multi-view Feature Learning with Discriminative Regularization

Jinglin Xu, Junwei Han, Feiping NieIJCAI 2017 · DOI

Abstract

Multi-View Feature Learning (MVFL) is a supervised framework designed to integrate heterogeneous data sources into a low-dimensional, discriminative space. By utilizing label information as a regularizer and formulating the objective as a trace optimization problem, MVFL provides a closed-form solution via eigenvalue decomposition. This approach ensures global optimality and avoids the local minima common in iterative multi-view clustering methods. Experimental results on benchmark datasets demonstrate that MVFL achieves state-of-the-art performance in clustering accuracy and dimensionality reduction efficiency.

Introduction

In real-world applications, data is often multi-modal. For instance, an image can be characterized by color, texture, and shape features. Effectively integrating these "multi-view" features while managing high dimensionality is a significant challenge. Existing methods are often unsupervised, ignoring label information, or iterative, which can lead to slow convergence and local optima. MVFL addresses these issues by using supervision as a "glue" during feature learning, ensuring that dimensionality reduction preserves class-discriminative information while maintaining separate projection matrices for each view to respect their unique properties.

Methodology

The MVFL framework balances two primary objectives: view-specific learning and discriminative regularization. View-specific learning ensures that the learned low-dimensional features for each view accurately represent the original data through a regression-like objective. Simultaneously, discriminative regularization uses label information to enforce consistency across views, forcing different feature types to "agree" on the underlying data structure based on known classes.

A key technical contribution of MVFL is its closed-form solution. By transforming the objective function into a trace optimization problem, the authors enable the use of eigenvalue decomposition. This guarantees a global optimum and significantly improves computational efficiency compared to iterative solvers that are prone to getting stuck in local extrema.

Experimental Evaluation

The performance of MVFL was evaluated on four benchmark datasets:

  • MSRCv1: 210 images across 7 classes, using 6 distinct feature views (e.g., SIFT, HOG, GIST).
  • Caltech101-7: 441 images across 7 classes, also using 6 feature views.
  • Handwritten: 2,000 samples of handwritten digits (0-9) with 6 views representing different mathematical descriptors.
  • Yale: 165 face images across 15 classes, using 3 feature views.

Results and Analysis

MVFL was compared against state-of-the-art methods such as Robust Multi-view K-Means (RMVKM) and Multi-view Discriminant Analysis (MvDA). It consistently achieved superior accuracy; for example, reaching 96.19% on MSRCv1, significantly outperforming Naive Multi-view K-Means (78.10%).

Comparison of learned feature dimensions across different views and datasets.

As illustrated in the figure above, MVFL maintains high performance while reducing data to a much lower dimensionality than its competitors across various views and datasets. This efficiency is particularly evident in the Handwritten and Yale datasets, where MVFL achieves significant compression without sacrificing discriminative power. The model's ability to achieve these results using a very small number of discriminative features highlights its effectiveness in addressing the "curse of dimensionality."

Conclusion

MVFL provides a robust and efficient framework for multi-view feature learning. By achieving a closed-form solution, it bypasses the limitations of iterative algorithms and ensures global optimality. While its reliance on labeled data and linear transformations are potential areas for future exploration—such as extending to semi-supervised or non-linear settings—its current performance sets a high standard for discriminative multi-view integration in clustering and dimensionality reduction tasks.

Summary updated

BibTeX
@inproceedings{xu2017multi,
  author = {Jinglin Xu and Junwei Han and Feiping Nie},
  title = {{Multi-view Feature Learning with Discriminative Regularization}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2017},
  doi = {10.24963/ijcai.2017/441},
}
Figure