Abstract
Real data are usually complex and contain various components. For example, face images have expressions and genders. Each component mainly reflects one aspect of data and provides information others do not have. Therefore, exploring the semantic information of multiple components as well as the diversity among them is of great benefit to understand data comprehensively and in-depth. However, this cannot be achieved by current nonnegative matrix factorization (NMF)-based methods, despite that NMF has shown remarkable competitiveness in learning parts-based representation of data. To overcome this limitation, we propose a novel multi-component nonnegative matrix factorization (MCNMF). Instead of seeking for only one representation of data, MCNMF learns multiple representations simultaneously, with the help of the Hilbert Schmidt Independence Criterion (HSIC) as a diversity term. HSIC explores the diverse information among the representations, where each representation corresponds to a component. By integrating the multiple representations, a more comprehensive representation is then established. A new iterative updating optimization scheme is derived to solve the objective function of MCNMF, along with its correctness and convergence guarantees. Extensive experimental results on real-world datasets have shown that MCNMF not only achieves more accurate performance over the state-of-the-arts using the aggregated representation, but also interprets data from different aspects with the multiple representations, which is beyond what current NMFs can offer.
Introduction
Nonnegative Matrix Factorization (NMF) has long been a staple in machine learning for finding "parts-based" representations of data. However, standard NMF and its common variants (like Graph-Regularized NMF) typically treat all features as a single block. This "one-size-fits-all" approach is often insufficient for real-world data, which is inherently multi-faceted. For instance, a single face image contains information about gender, facial expression, ethnicity, and lighting conditions. Standard NMF attempts to squash all these distinct semantic layers into a single feature vector, often failing to distinguish between them and leading to suboptimal performance in tasks like clustering and classification.
To address this, we introduce Multi-Component Nonnegative Matrix Factorization (MCNMF). MCNMF moves beyond learning a single data representation to learning multiple, diverse representations simultaneously. By capturing distinct semantic "components" of data, the model provides a more comprehensive and interpretable understanding of complex datasets.
Methodology
The core innovation of MCNMF is the factorization of the data matrix $X$ into $V$ different pairs of basis and representation matrices $(W^{(i)}, H^{(i)})$. Unlike traditional NMF which seeks a single $X \approx WH$, MCNMF acknowledges that data is composed of multiple semantic layers.
Multi-Component Objective and Diversity Constraint
Instead of a single factorization, the model seeks $V$ optimal representations. To ensure these representations are not redundant, we incorporate the Hilbert Schmidt Independence Criterion (HSIC) as a diversity regularization term. HSIC measures the dependence between variables; by minimizing it, MCNMF forces the learned representations ($H^{(i)}$) to be as independent (diverse) as possible. This diversity constraint pushes the model to discover different "angles" or semantic layers of the data, such as separating lighting effects from facial features.
Aggregated Representation and Optimization
After training, the individual representations are concatenated into a single, comprehensive feature vector $H^*$ for downstream tasks. This aggregated representation benefits from the unique information captured by each component. We derived a new iterative multiplicative update rule to solve the MCNMF objective function. These rules are formally proven to satisfy Karush-Kuhn-Tucker (KKT) optimality conditions and ensure the objective function is non-increasing, guaranteeing convergence.
Experimental Evaluation
MCNMF was evaluated on four benchmark datasets: the Yale Face Dataset (165 images, 15 subjects), the ORL Face Dataset (400 images, 40 subjects), Notting-Hill (4,660 faces from a movie), and COIL20 (1,440 object images).
Key Results
MCNMF consistently outperformed state-of-the-art NMF variants, including RNMF, GNMF, and LANMF, across metrics such as Accuracy (AC), Normalized Mutual Information (NMI), and Purity.
- Accuracy Improvements: On the Yale dataset, MCNMF achieved 46.42% accuracy, a significant improvement over the next best competitor, GNMF (41.58%). On the ORL dataset, it reached 62.95% (vs. 59.60% for GNMF).
- Stability and Efficiency: MCNMF exhibited lower standard deviations across multiple runs, indicating higher robustness. Despite the added complexity, the objective function converges rapidly, typically within 5 iterations.
Semantic Decomposition
A key advantage of MCNMF is its interpretability. By inspecting individual components, we found that different representations capture different semantic aspects. For example, one representation might cluster faces by expression (e.g., neutral vs. smiling), while another might cluster them by lighting direction. Experimental results showed a sharp increase in accuracy when moving from $V=1$ (standard NMF) to $V=3$, suggesting that most real-world datasets possess multiple underlying semantic layers that a single-component model cannot resolve.
Conclusion and Future Work
MCNMF demonstrates that forcing representations to be diverse actually enhances the quality of the final aggregated representation. While the model is stable, selecting the optimal number of components $V$ for unknown datasets remains an area for further study. Future work will explore extending MCNMF into a semi-supervised framework, where label information can be used to guide specific components toward known real-world attributes, further enhancing its utility for targeted data analysis.
Summary updated