🤖 machine learning

Intuitionistic Fuzzy Graph Embedded Random Vector Functional Link with Multiview Learning

This paper proposes the IFGRVFL-MV model, which integrates intuitionistic fuzzy sets, graph embedding, and multiview learning into Random Vector Functional Link networks to effectively handle uncertainty, preserve geometric structures, and leverage complementary feature information, thereby achieving superior classification accuracy on benchmark datasets compared to existing models.

Original authors: Vrushank Ahire, Yogesh Kumar, M. A. Ganaie

Published 2026-07-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Vrushank Ahire, Yogesh Kumar, M. A. Ganaie

Original paper licensed under CC BY 4.0 (http://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

The Big Picture: Building a Smarter "Guessing Machine"

Imagine you are trying to teach a computer to sort a pile of mixed-up mail into "Keep" and "Throw Away" bins. This is a classification problem. For a long time, computers have used a tool called RVFL (Random Vector Functional Link) to do this.

Think of RVFL as a very fast, efficient mail sorter. Unlike other sorters that have to learn slowly by trial and error (like a student studying for years), RVFL gets a "head start." It randomly assigns rules to its internal gears and then just calculates the final answer instantly. It's fast and usually good, but it has two main weaknesses:

It gets confused by "bad data": If a piece of mail is torn, stained, or looks like both "Keep" and "Throw Away," the sorter gets stuck.
It ignores the "shape" of the data: It treats every piece of mail as an isolated item, not realizing that similar-looking letters often belong in the same bin.
It only looks at one angle: It tries to sort based on just one description of the mail (e.g., color), ignoring other helpful details (e.g., texture or weight).

The authors of this paper, Vrushank Ahire, Yogesh Kumar, and M.A. Ganaie, built a new, upgraded sorter called IFGRVFL-MV. They combined three powerful tools to fix the weaknesses of the old sorter.

The Three Upgrades (The "Secret Sauce")

1. The "Fuzzy Judge" (Intuitionistic Fuzzy Sets)

The Problem: Sometimes, a piece of mail is so damaged or weird that the sorter doesn't know if it belongs in the "Keep" pile or the "Throw Away" pile. It's a gray area.
The Solution: The authors added a Fuzzy Judge. Instead of forcing a hard "Yes" or "No," this judge gives a score of "How much do I think this belongs?" and "How much do I think it doesn't belong?"

Analogy: Imagine a security guard at a club. A normal guard says, "You're in" or "You're out." The Fuzzy Judge says, "You look 80% like a VIP, but 20% like a troublemaker." If the person is clearly a troublemaker (an outlier or noise), the judge gives them a low score, telling the sorter to ignore them or treat them carefully. This makes the system robust against messy, noisy data.

2. The "Map Maker" (Graph Embedding)

The Problem: The old sorter looks at items one by one. It doesn't realize that if Item A looks like Item B, and Item B looks like Item C, then A and C probably belong in the same bin.
The Solution: The authors added a Map Maker. This tool draws invisible lines connecting similar data points, creating a web or a map.

Analogy: Imagine a party where you don't know anyone. A normal sorter asks, "Who are you?" and decides based on that. The Map Maker looks at the room and sees, "Oh, everyone wearing red shirts is standing in a circle talking." It preserves the geometric shape of the crowd. Even if one person in the red circle is wearing a blue hat (noise), the map knows they are still part of the red circle group. This helps the model generalize better to new data.

3. The "Team of Experts" (Multiview Learning)

The Problem: The old sorter only looks at the mail from one angle (e.g., just the color). But what if the color is misleading?
The Solution: The authors created a Team of Experts. They give the sorter two different "views" of the data at the same time.

Analogy: Imagine you are trying to identify a fruit.
- View 1 (Expert A): Looks at the fruit's color (Red?).
- View 2 (Expert B): Looks at the fruit's texture (Smooth?).
- The Team: Instead of asking just one expert, the model asks both. If Expert A says "Apple" (because it's red) and Expert B says "Apple" (because it's smooth), the model is very confident. If they disagree, the model uses its math to find the best compromise. This allows the model to use complementary information to make smarter decisions.

How They Put It All Together

The new model, IFGRVFL-MV, is like a super-team of mail sorters:

It looks at the mail from two different angles (Multiview).
It draws a map to see how the mail items relate to each other (Graph Embedding).
It uses a Fuzzy Judge to decide how much to trust each piece of mail, ignoring the ones that are too weird or damaged (Intuitionistic Fuzzy).

The Results: Did It Work?

The authors tested their new model against the old ones using standard datasets (like the UCI and KEEL repositories, which are like public libraries of test problems).

The Score: The new model achieved an average accuracy of 81.06%, beating the next best model (79.50%) and the basic models (around 70-72%).
The Proof: They didn't just say "it looks better." They ran strict statistical tests (like the Friedman and Wilcoxon tests). These are like referees checking the scorecards to ensure the win wasn't just luck. The results showed that the new model is statistically significantly better than the others.

Summary

In simple terms, the authors took a fast but slightly clumsy sorting machine and gave it:

Better eyes to see patterns (Graph Embedding).
A second opinion to check its work (Multiview Learning).
A smarter brain to handle confusion and mistakes (Intuitionistic Fuzzy Logic).

The result is a machine that sorts data more accurately, especially when the data is messy, noisy, or complex. The paper concludes that this approach is a promising step forward for handling uncertainty in machine learning.

Technical Summary: Intuitionistic Fuzzy Graph Embedded Random Vector Functional Link with Multiview Learning (IFGRVFL-MV)

Problem Statement
Random Vector Functional Link (RVFL) networks are recognized for their fast training speeds and universal approximation capabilities, achieved by randomly initializing hidden layer weights and analytically solving for output weights. However, standard RVFL models face specific limitations: they often struggle to preserve the intrinsic geometric relationships within data, are sensitive to noisy or uncertain datasets, and do not effectively leverage complementary information from multiple feature views (Multiview Learning). While existing variants have addressed these issues in isolation—such as Intuitionistic Fuzzy RVFL (IFRVFL) for uncertainty, Graph Embedded RVFL (GE-RVFL) for geometric structure, and Multiview RVFL (MvRVFL) for feature integration—no unified framework currently combines all three to simultaneously handle uncertainty, preserve topology, and exploit multiview data.

Methodology
The proposed IFGRVFL-MV model integrates Intuitionistic Fuzzy Sets (IFS), Graph Embedding (GE), and Multiview Learning (MVL) into a unified RVFL framework. The methodology proceeds as follows:

Intuitionistic Fuzzy Weighting: The model assigns intuitionistic fuzzy membership ( $f$ ) and non-membership ( $f^*$ ) values to data points in each view.
- The membership value is determined by the distance of a sample from its class center relative to the class radius.
- The non-membership value is derived from the ratio of heterogeneous neighbors (samples with different class labels) within a proximity threshold.
- A score function ( $s_i$ ) is computed based on these values to quantify the reliability of each sample, effectively down-weighting outliers and noisy data during training.
Graph Embedding (GE): To preserve the topological structure of the data, the model constructs intrinsic and penalty graphs over the concatenated feature matrix of each view. This framework ensures that samples with similar labels remain close (intrinsic graph) while samples with different labels are pushed apart (penalty graph), enhancing generalization.
Multiview Learning (MVL): The model processes two distinct views of the data (View-P and View-Q). It employs an early fusion strategy where the optimization objective jointly minimizes the error for both views while enforcing consistency between them.
Optimization Formulation: The core of the model is a regularized optimization problem that minimizes a weighted sum of squared errors, output weight norms, and graph embedding penalties.
- The error terms for each view are weighted by the square roots of the intuitionistic fuzzy score matrices ( $S_1^{1/2}$ and $S_2^{1/2}$ ), reducing the influence of uncertain samples.
- A cross-view consistency term ( $\rho \xi_1^T S_1^{1/2} S_2^{1/2} \xi_2$ ) aligns the predictions of the two views.
- The solution is derived analytically by solving a system of linear equations involving the Lagrangian multipliers, resulting in a closed-form solution for the output weights ( $\beta_1, \beta_2$ ).

Key Contributions
The paper outlines four primary contributions:

Integration of IFS into RVFL: The introduction of intuitionistic fuzzy sets into the RVFL framework to handle data uncertainty and imprecision. Membership and non-membership values are assigned based on sample distance from class centers and neighbor heterogeneity.
Graph Embedding Enhancement: The incorporation of the GE framework to capture geometric relationships within the data. By defining intrinsic and penalty graphs over the feature matrix, the model preserves topological structures to improve generalization.
Multiview Learning Utilization: The application of MVL to combine information from multiple feature views, allowing the model to capture complex patterns not apparent in a single view.
Empirical Validation: Extensive experiments on benchmark datasets from the UCI and KEEL repositories, supported by statistical analyses (Friedman, Wilcoxon, and win-tie-loss tests) to validate the model's superiority.

Results
The IFGRVFL-MV model was evaluated on eight benchmark datasets (e.g., pittsburgh-bridges, breast-cancer, mammographic, Pima).

Performance: The proposed model achieved the highest average classification accuracy of 81.06%, outperforming baseline models including RVFL1 (71.99%), RVFL2 (70.38%), GE-IFWRVFL (76.00%), and GRVFL-MV (79.50%).
Statistical Significance:
- The Friedman test yielded a statistic of 23.714 with a p-value of 0.00009, rejecting the null hypothesis and confirming significant performance differences among the models. The proposed model achieved the best mean rank (1.19).
- The Nemenyi post hoc test identified significant gaps between the proposed model and others, particularly against GRVFL-MV.
- The Wilcoxon signed-rank test confirmed that IFGRVFL-MV significantly outperforms RVFL1, RVFL2, and GE-IFWRVFL (p-values < 0.05).
- Win-Tie-Loss Analysis: Across the datasets, IFGRVFL-MV secured between 5 to 7 wins against the baselines, demonstrating consistent superiority.

Significance and Claims
The paper claims that IFGRVFL-MV represents a promising advancement in the domain of uncertainty and multiview learning environments. By seamlessly combining intuitionistic fuzzy theory, graph embedding, and multiview learning, the model addresses the limitations of existing approaches that treat these techniques in isolation. The authors assert that the model effectively handles noisy datasets and leverages complementary feature information to achieve superior classification accuracy. The work establishes a new benchmark for RVFL-based models in uncertain environments, though the authors note that future work will focus on optimizing computational efficiency and exploring scalability on larger, real-world datasets.