Abstract
Multiview learning aims to integrate diverse feature sets—such as text, images, and video—to improve object representation and classification. However, traditional Nonnegative Matrix Factorization (NMF) methods often prioritize view consistency over class discriminative information and are highly sensitive to outliers. This paper introduces DMultiNMF, a novel framework that combines local and global discriminative constraints with a robust error metric. By incorporating a large-margin principle, local patch alignment, and the Correntropy-Induced Metric (CIM), DMultiNMF ensures that the resulting latent representations are optimized for classification even in noisy environments.
Introduction
In modern machine learning, we frequently encounter "multiview" data, where objects are described by multiple distinct feature sets. For example, a webpage can be characterized by its textual content, embedded images, and associated metadata. While each view provides unique information, they all describe the same underlying entity. The central challenge in multiview learning is fusing these disparate sources into a single, coherent representation that enhances downstream tasks like classification.
The Limitations of Standard NMF
Existing multiview learning techniques, particularly those based on Nonnegative Matrix Factorization (NMF), typically face two significant hurdles:
- Ignoring Discrimination: Most NMF-based methods focus on "view consistency"—ensuring that the latent representations derived from different views are similar. While this captures shared information, it often ignores class labels. A representation can be consistent across views but still fail to separate distinct classes (e.g., "dog" vs. "wolf") effectively in the latent space.
- Sensitivity to Noise: Standard NMF utilizes the $L_2$ norm (squared error) to measure reconstruction quality. Because the $L_2$ norm penalizes large errors quadratically, a few noisy data points or outliers can disproportionately influence the model, leading to poor generalization.
The DMultiNMF Framework
To address these issues, the authors propose Discriminative Multiview NMF (DMultiNMF). This framework introduces three key innovations into the standard NMF objective function to make it more robust and discriminative.
1. Global Discrimination (Large Margin Principle)
Borrowing from Support Vector Machines (SVMs), DMultiNMF incorporates a "hinge loss" term. This constraint pushes the latent representations of different classes as far apart as possible. By maximizing the margin between classes in the latent space, the model facilitates more accurate and reliable classification.
2. Local Discrimination (Patch Alignment)
Beyond global margins, the model considers the local geometry of the data. Using a "patch alignment" framework, DMultiNMF ensures that a data point remains close to its neighbors from the same class while being actively pushed away from neighbors belonging to different classes. This preserves the fine-grained manifold structure of the data.
3. Robustness via Correntropy-Induced Metric (CIM)
Instead of the sensitive $L_2$ norm, DMultiNMF employs the Correntropy-Induced Metric (CIM) to measure reconstruction error. CIM is a non-linear similarity measure that behaves like the $L_2$ norm for small errors but "levels off" for large errors. Consequently, the model essentially ignores massive outliers rather than attempting to overfit to them, significantly improving performance on real-world, "messy" datasets.
Experimental Evaluation
The authors evaluated DMultiNMF on four diverse benchmark datasets:
- ORL Face Dataset: 400 images across 40 subjects, using raw pixels and Local Binary Patterns (LBP) as views.
- UCI Handwritten Digit Dataset: 2,000 samples of digits 0-9, using Fourier coefficients and pixel averages.
- XMedia Dataset: 5,000 text/image pairs across 20 categories.
- Wikipedia Dataset: 2,866 document-image pairs categorized into 10 semantic classes.
Results and Discussion
The full version of the algorithm, DMultiNMF-2 (which includes both global and local constraints), consistently outperformed all baseline methods.
- Superior Accuracy: On the ORL face dataset, the method achieved over 91% accuracy, and it surpassed 95% on the Handwritten Digit dataset.
- Performance Gains: DMultiNMF significantly outperformed "Single View" NMF, simple "Feature Concatenation," and standard MultiNMF. These results validate that adding discriminative constraints and the CIM metric provides a tangible performance boost over traditional consistency-only approaches.
- The Robustness Trade-off: The experiments confirmed that switching to CIM allows the model to maintain high accuracy even when the input data contains significant noise, a common occurrence in practical applications.
Conclusion and Future Work
DMultiNMF represents a significant step forward in multiview learning by explicitly linking the latent representation discovery process with the final classification goal. By balancing global margins, local manifold structure, and robustness to noise, it provides a more reliable framework for complex data fusion.
A current limitation of the model is its sensitivity to hyperparameters ($\lambda, \alpha, \beta, \gamma$), which currently require empirical tuning. Future research will focus on adapting DMultiNMF for semi-supervised learning, allowing the model to leverage large amounts of unlabeled data alongside limited labeled samples, further increasing its utility for massive real-world datasets.
Summary updated