💻 computer science

Trusted and Explainable Collective Learning for Conflictive Multi-View Decision-Making

This paper proposes Trusted and Explainable Collective Learning (TECL), a novel method that actively leverages conflicts in multi-view data to assess individual consistency, prioritize views, and generate reliable, interpretable collective decisions, rather than simply eliminating such conflicts.

Original authors: Nengjun Zhu, Zhiyu Zhang, Shenghui Lan, Jian Cao, Siji Zhu

Published 2026-06-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nengjun Zhu, Zhiyu Zhang, Shenghui Lan, Jian Cao, Siji Zhu

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

Imagine you are a patient with a complex health issue, and you walk into a hospital room filled with five different doctors. You have a bone problem, but one doctor is a heart specialist, another is a stomach expert, and a third is an orthopedist.

The Problem: The "Conflict" is Usually Seen as Bad
In the past, when these doctors disagreed, computer systems (and even some human teams) would try to silence the disagreement. They would say, "Okay, three doctors say 'Heart,' two say 'Bone.' Let's just ignore the bone doctors and go with the majority," or "Let's average their opinions until the disagreement disappears."

The authors of this paper argue that this is a mistake. They say: Disagreement is actually useful information. When doctors disagree, it tells us why they are thinking differently. The heart doctor is looking at your heart; the bone doctor is looking at your skeleton. If we just mash their opinions together without understanding the conflict, we lose the nuance that could save your life.

The Solution: TECL (The "Trustworthy Team Leader")
The paper introduces a new computer method called TECL (Trusted and Explainable Collective Learning). Think of TECL not as a referee who silences the team, but as a super-intelligent team leader who runs a meeting where everyone is heard, but the final decision is based on who knows the most about this specific patient.

Here is how TECL works, broken down into simple steps:

1. The "Dual-Concept" Glasses (Seeing the Details)

Imagine every doctor puts on a special pair of glasses.

Feature Glasses: These glasses highlight the specific body parts the doctor is looking at. The heart doctor's glasses zoom in on the heart; the bone doctor's glasses zoom in on the skeleton. This helps the computer understand what each doctor is focusing on.
Decision Glasses: These glasses show the doctor's final verdict and their attitude. It answers: "Do you support the group's final plan, or do you strongly oppose it?"

By using these "glasses," the system doesn't just see a "Yes/No" answer; it sees the reasoning behind the answer.

2. The "Evidence Vault" (Checking the Facts)

Instead of just trusting a doctor because they have a fancy title (like "Chief Cardiologist"), TECL asks: "What evidence do you have for this specific patient?"

It uses a mathematical tool (called "Evidential Deep Learning") to check how strong the evidence is.
If a doctor says, "This patient needs bone surgery," TECL checks the data. If the evidence is weak, the doctor's opinion is marked as "uncertain." If the evidence is strong, it's marked as "reliable."
Crucially, TECL keeps the conflict alive. It doesn't delete the bone doctor's opinion just because the heart doctor disagrees. It records that there is a disagreement and calculates how much that disagreement lowers the overall confidence in the final decision.

3. The "Dynamic Priority" System (Who Leads the Meeting?)

This is the most clever part. In a real meeting, the "Chief Doctor" usually leads. But in TECL, the leader changes depending on the situation.

Inherent Importance: This is the doctor's permanent rank (like their job title).
Dynamic Perception: This is how relevant the doctor is right now.
- Example: If the patient has a broken leg, the Orthopedist (who might be a junior doctor) suddenly becomes the "Team Leader" for this specific case, even if the Chief Cardiologist is in the room. TECL automatically boosts the Orthopedist's weight because their expertise matches the patient's specific problem.
- If the patient has a heart attack, the Chief Cardiologist gets the boost.

4. The Final Verdict (The "Reliable" Decision)

TECL combines all these opinions using a special formula.

If everyone agrees, the final decision is High Reliability.
If they disagree, the system doesn't just pick a winner. It says, "We have a conflict. The Orthopedist is right about the bone, but the Cardiologist is worried about the heart. The final decision is Bone Surgery, but the Reliability Score is Medium because of the heart concern."

This gives the final answer two things:

The Decision: What should we do?
The Explanation: Why did we choose this? (e.g., "We chose the bone surgery because the bone expert had the strongest evidence for this specific case, even though the heart expert disagreed.")

Why is this better?

It listens to the minority: If a junior doctor has the right answer for a specific problem, TECL lets them speak up, rather than being drowned out by the "Chief."
It explains itself: It can show you which features (like a specific gene or symptom) caused the doctors to disagree.
It is honest about uncertainty: If the doctors are fighting, the system admits, "We aren't 100% sure," rather than pretending to be confident.

In Summary:
The paper proposes a smart system that treats disagreement as a valuable clue rather than a mistake. It acts like a wise meeting moderator who knows when to let the heart specialist lead and when to let the bone specialist lead, ensuring the final decision is not just accurate, but also explainable and honest about its confidence.

Note: The paper tested this method on a real dataset of breast cancer consultations involving multiple doctors and found it performed better than existing methods at making accurate decisions and explaining why.

Technical Summary: Trusted and Explainable Collective Learning for Conflictive Multi-View Decision-Making

1. Problem Definition

The paper addresses the challenge of multi-view decision-making where different data sources (views) or experts provide conflicting opinions. In real-world scenarios, such as multidisciplinary medical consultations, different views often possess unique insights that lead to disagreements. Existing methods typically treat conflict as a negative factor, aiming to eliminate or minimize it through negative operations (e.g., aligning views or discarding conflicting instances). However, this approach often overlooks the value of conflict in revealing view distinctions and assessing the reliability of decisions.

Furthermore, current multi-view learning methods suffer from two main limitations:

Lack of Reliability Modeling: They often fail to quantify the uncertainty and reliability of both individual view decisions and the final collective opinion, especially when conflicts arise.
Lack of Explainability: While some interpretable models exist (e.g., Concept Bottleneck Models), they often focus on the final decision without explaining the feature-level reasoning of individual views or the root causes of conflicts between them.

The goal is to develop a method that actively models conflicts rather than suppressing them, thereby enhancing the trustworthiness and interpretability of collective decisions while accurately predicting both individual view outputs and the final consensus.

2. Methodology: Trusted and Explainable Collective Learning (TECL)

The authors propose TECL (and its variant TrustCL), a framework designed to handle conflictive multi-view data through three core modules:

A. Dual-Concepts Learning Module

This module aims to capture concepts at both the feature level and the decision level to enhance interpretability.

Feature Concepts: An attention mechanism is applied to input features ( $x_v$ ) to generate feature concepts ( $a_v$ ), representing the degree of attention a specific view pays to sample features.
Decision Concepts: Enhanced instances ( $\hat{x}_v$ $\overset{x}{^}_{v}$ ), guided by feature concepts, are fed into an attitude capture network and an evidential Deep Neural Network (DNN). This produces:
- Individual Decisions: The predicted category for the view.
- Supporting Attitudes: A binary indicator (0 or 1) showing whether the view supports the eventual collective decision.
Consensus Judgment: If all views are consistent (no conflict), a highly reliable collective decision is generated immediately. If conflicts exist, the process moves to the next module.

B. View-Specific Evidence Learning Module

This module utilizes Evidential Deep Learning (EDL) to quantify the reliability of each view's opinion.

Evidence Acquisition: Instead of standard classification, the model outputs evidence vectors ( $e_v$ ) for each view.
Reliability Calculation: Using Subjective Logic and the Dirichlet distribution, the model converts evidence into a multinomial opinion $w_v = \{b_v, u_v\}$ , where $b_v$ represents belief masses and $u_v$ represents uncertainty.
Key Innovation: The model uses the individual view's ground truth ( $y_v$ ) as a supervised signal rather than just the collective truth. This preserves the inherent conflicts and differences between views, allowing the DNN to learn specific knowledge about each viewpoint and its associated uncertainty.

C. Reliability-Oriented Collective Learning (RoCL) Module

This module fuses conflicting opinions by calculating a dynamic priority for each view.

Priority Definition: The priority $c_v$ $c_{v}$ of a view is determined by two factors:
1. Inherent Importance ( $\mu$ ): A static prior based on the view's historical status or role (e.g., a Chief Cardiologist vs. a General Practitioner).
2. Dynamic Perception ( $\sigma$ ): An instance-specific importance derived from the view's relevance to the specific data instance (e.g., an orthopedist's higher relevance for a bone injury).
Fusion Mechanism: The priorities are modeled using a Normal distribution sampled via a Dirichlet distribution. The final collective opinion is generated by fusing view-specific opinions ( $w_v$ $w_{v}$ ) using these priorities as weights.
- The fusion rule (Equation 8) mathematically combines belief and uncertainty. If views conflict, the uncertainty of the collective opinion increases, reflecting the reduced reliability of the consensus.
- If views agree, the belief is reinforced.

D. Loss Function

The training objective combines three components:

Evidence-based Cross-Entropy ( $L_{ace}$ ): Ensures correct label prediction using the Dirichlet distribution to capture second-order uncertainty.
KL Divergence ( $L_{KL}$ ): Minimizes the difference between the predicted Dirichlet distribution and a uniform distribution for incorrect labels, ensuring non-misleading evidence is suppressed.
Binary Cross-Entropy ( $L_{bce}$ ): Trains the model to predict whether a view's decision aligns with the collective opinion.
The total loss penalizes both individual view errors and the collective prediction error, treating conflict as a motive for learning rather than an error to be eliminated.

3. Key Contributions

Active Conflict Modeling: The paper proposes TrustCL and TECL, which actively model conflicts among individuals to highlight view distinctions and reveal the reliability of decisions, rather than suppressing them.
Dual-Level Interpretability: TECL models concepts at both the feature level (what features a view focuses on) and the decision level (why a view supports a specific outcome), providing explanations for both individual and collective decisions.
Dynamic Priority & Reliability Fusion: The method dynamically determines view priorities based on both inherent status and instance-specific contributions. It combines these priorities with view-specific evidence to generate a final collective opinion and its reliability score.
Empirical Validation: The approach is validated on a real-world multidisciplinary medical consultation dataset (BCDB), demonstrating superiority over baselines in accuracy, recall, and Mean Reciprocal Rank (MRR).

4. Experimental Results

The authors evaluated TECL and TrustCL on the BCDB (Breast Cancer Database) dataset, which contains 9,015 instances of multidisciplinary medical recommendations involving six doctors (views).

Performance: TECL outperformed state-of-the-art baselines (including TMC, ECML, and MLP variants) in both single-view prediction (individual doctor accuracy) and collective prediction (final consensus accuracy).
- Collective Accuracy (CA): TECL achieved 78.5%, compared to 77.9% for the next best (ECMLboth).
- Individual Accuracy (IA): TECL achieved 74.5%, outperforming ECML (73.5%) and TMC (70.6%).
Ablation Studies: Removing the dynamic perception component ( $\sigma$ ) or relying solely on inherent importance ( $\mu$ ) significantly degraded performance, confirming the necessity of dynamically adjusting view importance based on specific instances.
Interpretability Analysis: Visualizations showed that the model correctly identified when a specific doctor (e.g., an orthopedist for a bone issue) should have higher priority despite having a lower "title" than a generalist. It also successfully quantified that collective reliability decreases when views strongly disagree, providing a transparent measure of decision confidence.

5. Significance and Claims

The paper claims that TECL represents a novel branch of reliable and explainable multi-view learning. Its primary significance lies in shifting the paradigm from conflict reduction to conflict utilization. By treating conflicting opinions as valuable information for reliability modeling, the method:

Enhances the credibility of collective decisions by explicitly modeling uncertainty.
Provides explainability for why a decision was made and why conflicts occurred, which is critical for safety-critical applications like medical diagnosis.
Ensures that the "voice of the minority" (e.g., a specialist with a specific focus) is not drowned out by the majority or by static hierarchy, provided the evidence supports it.

The authors conclude that their method effectively balances the need for accurate collective decisions with the need to understand the underlying reasoning and reliability of those decisions in complex, multi-disciplinary environments. Future work is suggested to explore integrating Large Language Models (LLMs) for large-scale semantic contexts.