Introduction
In modern machine learning, datasets often consist of multiple "views"—different sets of features describing the same instance. For example, an image can be represented by its color histogram, texture descriptors, and shape features. Effectively integrating these diverse information sources is a central challenge. Traditional methods often rely on naive concatenation or shared subspace learning. However, these approaches frequently overlook "private" discriminative information unique to specific views, often misclassifying it as noise.
This paper introduces a novel Robust Multi-View Low-Rank and Sparse Matrix Decomposition framework. By explicitly modeling shared correlations, view-specific discriminative features, and random noise, the proposed method achieves superior performance in complex image classification tasks.
The Proposed Framework
The core of the approach lies in decomposing a concatenated multi-view matrix $X$ into three distinct components:
- Low-Rank Part ($L$): Captures the global structure and shared correlations across all views using trace norm regularization.
- Sparse Part ($S$): Captures the "private" discriminative information unique to individual views.
- Noise Part ($E$): Accounts for random errors and outliers.
To ensure the private component $S$ is informative, the authors apply a group-structured sparsity constraint. This encourages the selection of relevant discriminative elements within each view while maintaining structural consistency. The optimization problem is solved using an Augmented Lagrange Multiplier (ALM) method, which iteratively updates the variables to reach a robust solution. The final classification is performed using an augmented representation, combining a compressed version of the shared component (via PCA) with the significant rows of the private component.
Experimental Evaluation
The method was evaluated on six benchmark datasets: Caltech101 (7 and 20 classes), NUS-WIDE, Handwritten Numerals (HW), Animal with Attributes (AwA), and Scene-15. For each dataset, six distinct visual features (e.g., GIST, HOG, LBP, SIFT) were extracted to represent multiple views.
Results and Discussion
The proposed method consistently outperformed state-of-the-art baselines, including SimpleMKL and lrMMC.
As shown in the results table, the "Full" version of the proposed method achieved a remarkable 98.00% accuracy on the Scene-15 dataset, significantly higher than the 88.37% achieved by the best baseline, $I^2SCA$. The inclusion of group-sparsity proved critical, as the full model outperformed versions using standard $\ell_1$ or $\ell_{2,1}$ norms.
The confusion matrices further illustrate the model's robustness. For the Handwritten Numerals dataset, the model achieved an average accuracy of 98%, with a nearly perfect diagonal, indicating high precision in resolving inter-class ambiguities.
Conclusion
By treating view-specific residuals as valuable "private" information rather than mere noise, this multi-view matrix decomposition scheme offers a powerful new way to explore discriminative data. Future work will focus on reducing computational complexity for massive datasets and extending the framework to semi-supervised learning environments.
Summary updated