🤖 machine learning

Cooperation of Experts: Fusing Heterogeneous Information with Large Margin

The paper proposes the Cooperation of Experts (CoE) framework, which fuses heterogeneous information into unified multiplex networks using domain-specific encoders that collaborate via a novel large margin mechanism to achieve superior performance in capturing complex data structures.

Original authors: Shuo Wang, Shunyang Huang, Jinghui Yuan, Zhixiang Shen, Zhao Kang

Published 2026-06-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Shuo Wang, Shunyang Huang, Jinghui Yuan, Zhixiang Shen, Zhao Kang

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 Problem: A Messy Library

Imagine you are trying to understand a complex story, like a mystery novel. But instead of having one book, the story is scattered across five different formats: a text manuscript, a series of maps, a collection of audio recordings, a set of photos, and a list of financial receipts.

In the world of data, this is called heterogeneous information. It's messy because:

Different Languages: The text speaks one way, the maps another, and the audio another.
Different Connections: The text might link characters by "friendship," while the maps link them by "travel routes," and the receipts link them by "money spent."

Old methods tried to solve this by forcing all these different formats into a single, rigid box. They treated the text, maps, and audio as if they were all the same kind of puzzle piece. The paper argues this is a mistake. It's like trying to read a map by listening to an audio recording; you lose the unique value of each format.

The Solution: The "Cooperation of Experts" (CoE)

The authors propose a new system called CoE. Instead of one giant brain trying to understand everything at once, they build a team of specialized "experts."

Think of it like a high-end detective agency solving a complex case:

The Low-Level Experts: These are specialists who only look at one specific type of evidence.
- Expert A only looks at the text manuscripts.
- Expert B only studies the maps.
- Expert C only listens to the audio.
- Because they focus on just one thing, they become incredibly good at spotting patterns in that specific format without getting confused by the others.
The High-Level Experts: These are the "synthesizers." They don't look at the raw evidence; they look at the conclusions the Low-Level experts have drawn. They figure out how the text clues connect to the map clues to reveal the bigger picture.

The Secret Sauce: The "Large Margin" Mechanism

In many previous systems (called "Mixture of Experts"), the experts were like a committee where only the loudest voice won. If Expert A said "It's a cat" and Expert B said "It's a dog," the system might just pick the one with the highest confidence score and ignore the other. This is competition.

The CoE system changes the rules to cooperation.

Imagine a panel of judges deciding a winner. Instead of just picking the highest score, they use a special rule called the Large Margin Mechanism:

The Confidence Tensor: This is like a dynamic voting system. It asks each expert, "How sure are you about your answer?" and "How sure are you about other people's answers?"
The Gap (Margin): The system tries to make sure that the "correct" answer is not just the winner, but the clear winner. It pushes the top choice far away from the second-best choice.
- Analogy: If the correct answer is "The sky is blue," the system wants the experts to be so confident that the gap between "Blue" and "Green" is huge. If the gap is small (e.g., 51% vs 49%), the system knows the team is confused and needs to adjust.

By maximizing this "gap," the system forces the experts to agree on a strong, clear conclusion, reducing the chance of errors.

How They Built It

Cleaning the Data: Real-world data is often noisy (like a map with missing roads or a text with typos). The system first uses a "Graph Structure Learning" tool to clean up the connections, filling in missing roads and fixing typos before the experts even start looking.
The Two-Level Team: They train the Low-Level experts on the cleaned data, then train the High-Level experts to combine those insights.
The Optimization: They use a mathematical trick to ensure the "voting system" (the confidence tensor) is fair and that the experts aren't just guessing.

What They Found

The authors tested this "Detective Agency" on several real-world datasets (like academic papers, movie reviews, and social networks).

The Result: The CoE team consistently solved the puzzles better than any other method they compared it against.
The Stability: Even when they intentionally "attacked" the data (by deleting random connections or adding fake ones), the CoE system remained stable. It was like a detective who could still solve the case even if half the evidence was stolen.
The Theory: They also proved mathematically that this method is stable and won't get stuck in a "local optimum" (a dead end where it thinks it's done but isn't).

Summary

The paper introduces a new way to handle complex, mixed-up data. Instead of forcing everything into one mold, it creates a team of specialists who focus on their own strengths. Then, it uses a special "margin" rule to force them to collaborate and agree on a clear, confident answer, rather than just letting the loudest voice win. This results in a system that is more accurate, more stable, and better at handling messy real-world data.

Technical Summary: Cooperation of Experts (CoE)

Problem Statement

Fusing heterogeneous information remains a persistent challenge in modern data analysis. Real-world objects and data often exist in multiple types or diverse interaction forms (e.g., multimodal data combining images and text, or social networks with varied tie types like friendship and professional relationships). Existing approaches for modeling these intricate relationships often fail to account for the inherent heterogeneity of object patterns across different semantic spaces. Specifically, methods that train a single predictor across an entire network tend to produce suboptimal results by neglecting the intrinsic heterogeneity and varying characteristics of different link patterns. Furthermore, while Mixture of Experts (MoE) frameworks exist, they typically rely on gating mechanisms that activate only a subset of experts, fostering competition rather than collaboration, which limits the full exploitation of diverse information embedded in heterogeneous data.

Methodology: The Cooperation of Experts (CoE) Framework

The authors propose the Cooperation of Experts (CoE) framework, which encodes multi-typed information into unified heterogeneous multiplex networks. The framework is designed to transcend modality and connection differences through a two-level expert architecture and a novel large margin optimization mechanism.

1. Encoding Heterogeneous Information

The framework begins by modeling arbitrary forms of heterogeneous information as a set of $V$ networks, $G = \{G_1, ..., G_V\}$ . To address the issue of sparse and noisy real-world network structures, CoE employs a Graph Structure Learning (GSL) strategy. Using a Simple Graph Convolution (SGC) network, the method refines the network structure and reconstructs adjacency matrices via K-nearest neighbors (KNN), ensuring the final networks adhere to properties like non-negativity, symmetry, and normalization.

2. Two-Level Expert Design

The core of CoE is a hierarchical expert system designed to disentangle and integrate heterogeneous information:

Low-Level Experts: These components focus on learning patterns from individual networks (specific layers). They are trained to maximize the mutual information between the specific network representation and the node labels, $I(G'_i; Y)$ .
High-Level Experts: These operate on fused representations to capture cross-network relationships and higher-order dependencies that cannot be uncovered when networks are analyzed independently. They are trained to maximize mutual information between different network representations, $I(G'_i; G'_j)$ , and between individual networks and the total fused network.

The fusion of networks is achieved by optimizing a loss function that maximizes mutual information, effectively capturing complementary information across layers.

3. Large Margin Mechanism for Collaboration

Unlike traditional MoE approaches that select a subset of experts, CoE advocates for the collaboration of all experts in the decision-making process.

Confidence Tensor: A learnable confidence tensor $\Theta$ is introduced to assign weighted influence to each expert based on its proficiency in capturing distinct patterns. This tensor dynamically adjusts the authority of each expert.
Large Margin Optimization: To ensure effective collaboration and prevent a single expert's errors from misleading the system, the framework introduces a large margin mechanism. The optimization strategy maximizes the margin between the two most confidently predicted outcomes. This is formalized by minimizing a loss function that balances the cross-entropy loss (ensuring correctness) and a margin-based loss (ensuring consistency among experts). The margin is defined as the difference between the highest and second-highest confidence scores in the collaborative prediction.

Key Contributions

The paper outlines three primary contributions:

Novel Framework: The proposal of the CoE framework, which extracts stable and nuanced node patterns from multiplex networks by tailoring experts to capture both specialized and shared features. It effectively disentangles and integrates heterogeneous information.
Shift from Competition to Cooperation: To the best of the authors' knowledge, this is the first framework to emphasize expert cooperation rather than competition in the context of multiplex networks. It introduces a large margin optimization strategy that maximizes the effectiveness of each expert's specialized expertise, fostering a holistic understanding of network knowledge.
Theoretical and Empirical Validation: The work provides rigorous theoretical analysis guaranteeing the framework's feasibility, stability, and convergence. Extensive experiments across diverse benchmarks demonstrate superior performance compared to state-of-the-art methods.

Experimental Results

The authors evaluated CoE on five benchmark network datasets (ACM, DBLP, Yelp, MAG, Amazon) and four multimodal datasets (ESP, Flickr, IAPR, NUS).

Node Classification: CoE achieved the highest accuracy across all datasets, consistently outperforming supervised and unsupervised Graph Structure Learning (GSL) methods, multiplex learning methods, and Graph-MoE baselines (e.g., GMoE, Mowst). For instance, on the ACM dataset, CoE achieved 94.21% accuracy compared to the runner-up InfoMGF at 92.81%.
Multimodal Classification: CoE significantly outperformed baseline models in multimodal settings, demonstrating stability and adaptability.
Ablation Studies: Removing high-level experts or the GSL process resulted in degraded performance, confirming the necessity of both components. Replacing the large margin mechanism with Random Forest (RF) or Weighted Random Forest (WRF) yielded lower performance, validating the superiority of the proposed collaboration mechanism.
Robustness: Under structural attacks (random edge addition or deletion), CoE exhibited strong resilience, significantly outperforming fixed-structure models and other GSL methods, particularly under extreme perturbations.
Sensitivity: The model demonstrated low sensitivity to variations in key hyperparameters ( $\gamma$ and $\alpha$ ), indicating stability.

Significance and Claims

The paper claims that CoE represents a novel paradigm for fusing diverse information. Its significance lies in being the first study to explore expert learning specifically within multiplex networks and, more importantly, pioneering a shift from the traditional emphasis on expert competition to expert cooperation. By leveraging a large margin mechanism to balance contributions and address expertise disparities, the framework offers a more robust and comprehensive approach to modeling complex, heterogeneous data. The authors conclude that this work opens avenues for further research into applying the "Collaboration of Experts" mechanism to more complex scenarios.