💻 computer science

Class-Geometry Aware Correlated Joint Subspace Analysis for Multi-View Data

This paper proposes a unified supervised multi-view subspace learning method that integrates class-wise geometric structures and adaptive view weighting to efficiently construct a discriminative joint latent space, outperforming state-of-the-art approaches in both accuracy and computational speed.

Original authors: Sankar Mondal, Pradipta Maji

Published 2026-08-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Sankar Mondal, Pradipta Maji

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the modern era of big data, information rarely comes from a single source. A patient's health record might include genetic sequences, blood test results, and medical imaging scans, each offering a different perspective on the same condition. In the field of machine learning, these distinct sources are called "views." The challenge for scientists is not just to look at these views separately, but to weave them together into a single, coherent understanding. Traditional methods often struggle with this task; some simply mash all the data together, losing the unique character of each source, while others treat every source as equally important, even when some are noisy or irrelevant. Furthermore, many existing techniques fail to use the known categories of the data, such as specific disease types, to guide the learning process, missing an opportunity to make the final result more accurate and easier to interpret.

Researchers at the Indian Statistical Institute have developed a new method to solve these problems, designed to find the hidden connections between different types of data while respecting the known structure of the groups within that data. They call their approach SGR-MCCDA. Instead of forcing all data views to be treated the same, this method learns how important each view is for the specific task at hand. It also uses the known labels of the data, such as cancer subtypes, to ensure that the final combined view keeps similar samples close together and different samples far apart. The result is a unified, low-dimensional space where the data is organized clearly, making it much easier to classify and understand.

The core of this new method lies in balancing two competing needs. First, it seeks to find the common ground shared by all the different views, ensuring that the information they agree on is preserved. Second, it looks for the unique, complementary information within each view that helps distinguish between different categories. Previous approaches often focused on just one of these aspects or ignored the specific geometry of how data points are arranged within their classes. The new technique incorporates a "graph" structure, which acts like a map of relationships. It connects samples that belong to the same class and separates those that belong to different classes. By weaving this map into the learning process, the algorithm ensures that the final representation of the data keeps the local neighborhoods of similar items intact while maintaining the global separation between different groups.

To make the process efficient and realistic, the method also assigns a weight to each view, effectively asking the algorithm to decide which sources of information are most reliable. If one view is full of noise or irrelevant details, the method learns to give it less influence, while boosting the contribution of clearer, more informative views. This dynamic weighting prevents the model from being misled by poor-quality data. The researchers tested this approach on a variety of complex datasets, including four different types of cancer data from The Cancer Genome Atlas and several standard benchmark datasets used in computer science. These datasets ranged from medical records with thousands of features to image collections and text documents.

The results showed that this new method consistently outperformed existing state-of-the-art techniques. On the cancer datasets, it achieved significantly higher accuracy in classifying different tumor types compared to other algorithms, including those based on deep learning. For instance, on a dataset of lower-grade glioma, the method reached an accuracy of nearly 98 percent. On a dataset of plant leaves with one hundred different species, it correctly identified the species with 98.5 percent accuracy. Beyond just being more accurate, the method was also faster. While deep learning models often require powerful graphics processors and long training times, this new approach ran efficiently on standard computer processors, often completing its tasks in a fraction of the time required by its competitors.

The researchers also demonstrated that the method converges quickly, meaning it finds a stable and optimal solution after only a few rounds of calculation. They introduced a new way to select the best settings for the algorithm by measuring how well the data groups were separated geometrically. This allowed them to fine-tune the model without needing to guess or test every possible combination manually. The study confirmed that by combining the shared information across views with the unique discriminative power of each view, and by respecting the known class structure of the data, it is possible to build a much more robust and interpretable model. The work suggests that for many real-world problems where data comes from multiple sources, a method that can intelligently weigh and integrate these sources while preserving their underlying structure offers a superior path forward.

Technical Summary: Class-Geometry Aware Correlated Joint Subspace Analysis for Multi-View Data

Problem Statement
The paper addresses the challenge of multi-view learning, where data is represented by multiple feature sets (views) derived from a common source. While deep learning approaches have shown efficacy, they often suffer from high computational costs, memory inefficiency, and a lack of interpretability regarding the joint latent space. Conversely, traditional multi-view subspace learning techniques, such as Multiset Canonical Correlation Analysis (MCCA), aim to find a low-dimensional joint subspace by maximizing correlations among views. However, existing methods face several limitations:

They often fail to exploit prior knowledge of class labels or the geometric structure of data classes.
They typically ignore variations in view quality, assuming all views contribute equally, which may not hold in real-world scenarios.
Many approaches focus solely on shared correlation (consensus) while neglecting complementary information within individual views, or vice versa, failing to preserve both local and global class structures in the joint subspace.

Methodology
The authors propose a unified supervised multi-view subspace learning method termed Supervised Graph Regularized Multi-View Canonical Correlation and Discrimination Analysis (SGR-MCCDA). The method is built upon the Maximum Variance (MAXVAR) formulation of MCCA and integrates three core components:

Consensus and Complementarity: The objective function minimizes the dissimilarity (Euclidean distance) between the linear projections of multiple views and a common latent space (consensus principle). Simultaneously, it maximizes the between-class scatter within each view (complementary principle) to enhance discriminative power.
Class-Geometry Awareness: To preserve the local and global class structures, the method incorporates a graph regularizer. Two graphs are constructed based on class labels: a within-class graph ( $G_w$ ) connecting samples of the same class, and a between-class graph ( $G_b$ ) connecting samples of different classes. These are combined into a modified graph Laplacian ( $L_{G'}$ ) which acts as a regularizer to ensure that samples of the same class remain close and samples of different classes remain far apart in the projected space.
View Relevance Learning: Unlike methods that treat all views equally, SGR-MCCDA learns a relevance coefficient ( $\omega_v$ ) for each view. An entropy term is added to the objective function to prevent trivial solutions and encourage a balanced yet adaptive weighting of views based on their importance to the task.

Optimization
The optimization problem is non-convex with respect to all variables simultaneously but is convex with respect to each variable when others are fixed. The authors employ an alternating optimization technique to solve the problem iteratively:

Projection Matrices ( $W_v$ ): Updated using a closed-form solution involving the inverse of the within-class scatter matrix.
Common Space ( $S$ ): Updated by solving a generalized eigenvalue problem involving the weighted sum of view-specific scatter matrices and the graph Laplacian.
View Coefficients ( $\omega_v$ ): Updated using a softmax-like rule derived from the entropy constraint.

The paper provides a theoretical proof of convergence, demonstrating that the objective function is bounded below and decreases monotonically with each iteration. Additionally, a low-rank approximation technique is employed to handle cases where the within-class scatter matrix is singular (i.e., when feature dimensions exceed sample size).

Key Contributions

Unified Framework: The introduction of SGR-MCCDA, which simultaneously enforces consensus among views and maximizes discriminative information within views while preserving class-wise geometry.
Geometric Regularization: The integration of within-class and between-class graph Laplacians to maintain local class structures in the joint subspace, a feature often missing in standard MCCA extensions.
Adaptive View Weighting: A mechanism to automatically learn the relevance of different views, allowing the model to down-weight irrelevant or redundant views.
Parameter Selection Index: The proposal of a novel Modified Geometric Separability Index (MGSI) to select optimal model parameters ( $\alpha, \beta, \gamma$ ) based on the geometric separability of the extracted features.
Theoretical Analysis: Rigorous analysis of the algorithm's properties, including convergence guarantees and computational complexity analysis, showing efficiency particularly for "large $p$ , small $n$ " problems.

Experimental Results
The authors evaluated SGR-MCCDA on four cancer datasets (BRCA, CRC, LGG, GBM) from The Cancer Genome Atlas (TCGA) and four benchmark datasets (Cora, Reuters, Caltech-20, 100Leaves). The method was compared against state-of-the-art non-deep learning algorithms (e.g., GMA, MvDA, MvCCDA, GMCCA) and deep learning approaches (e.g., SMDDRL, TMC, GEGCN).

Performance: SGR-MCCDA achieved the highest classification accuracy in the majority of cases across both train-test and 10-fold cross-validation setups. For instance, it achieved approximately 98% accuracy on the LGG dataset and outperformed deep learning models in terms of both accuracy and execution time on several datasets.
Statistical Significance: Paired t-tests and Wilcoxon signed-rank tests confirmed that the performance improvements were statistically significant in most comparisons.
Efficiency: The proposed method demonstrated significantly lower execution times compared to many existing algorithms, attributed to the quick convergence of the iterative optimization (typically stabilizing within a few iterations).
Ablation Studies: Experiments confirmed that incorporating graph regularization ( $\beta \neq 0$ ) and learning view-specific weights (proposed weights vs. equal weights) consistently improved performance over baseline variants.

Significance and Claims
The paper claims that SGR-MCCDA offers a superior balance between computational efficiency and classification performance compared to both traditional subspace learning methods and deep learning approaches. By explicitly modeling class geometry and view relevance, the method produces a joint subspace that is not only highly discriminative but also interpretable. The authors emphasize that their approach effectively handles the "large $p$ , small $n$ " scenario common in bioinformatics (e.g., cancer genomics) without the heavy computational burden of deep neural networks. The work concludes that integrating consensus, complementarity, and class-geometry awareness is crucial for robust multi-view learning. Future directions suggested include extending the framework to unaligned or incomplete multi-view data and exploring optimization on generalized Stiefel manifolds.

More like this