Source-linked AI summary
Multi-View Trust Evaluation for Collaborator Selection via Evidential Deep Learning
Botao Zhu, Xianbin Wang
TL;DR
The paper addresses trustworthy collaborator selection when historical trust data are heterogeneous across task-owner views and uneven in quality. It proposes MVE, combining multi-view temporal modeling, evidential uncertainty quantification, and dynamic fusion. Experiments report improved trust-evaluation accuracy and task success rate, with robustness under high-conflict conditions.
Problem
Trust evaluation is difficult because collaborator histories contain context-dependent, multi-source observations with uneven quality, while fusion must account for differing certainty across views.
Method
MVE treats each prior task owner as an observation view, uses Mamba for view-specific temporal dynamics, models uncertainty evidentially, and dynamically fuses the resulting evidence.
Results
MVE outperforms baselines in trust evaluation accuracy and task success rate, maintaining 68.6% accuracy at η = 50% after degrading by only 17.1%.
Takeaways & Limitations
MVE provides uncertainty-aware, evidence-based trust fusion that identifies and filters unreliable observations for collaborator selection.
Abstract
from arXiv · showhide
Selection of trustworthy collaborators in distributed systems is critical for efficient task completion, necessitating the inference of trustworthiness from their past collaboration experience. However, as a collaborator serves distinct devices across diverse scenarios in past collaborations, its trust-related data, observed from different device-specific views, is inherently multi-source, heterogeneous, and uneven in quality. Consequently, achieving accurate trust evaluations for collaborator selection remains a major challenge. To tackle these issues, we propose a novel multi-view evidential learning (MVE) based trust evaluation method. First, to accommodate the multi-source heterogeneity of observed trust-related data, we model each task owner who has interacted with a potential collaborator as an independent observational view, enabling the evaluation of the collaborator's view-specific trust. Second, to address the dynamic evolution of trust under changing conditions, we leverage the powerful long-sequence modeling capability of the Mamba model to capture the deep temporal patterns of a collaborator's trust state within each view. Furthermore, to quantify the certainty levels of view-specific trust assessments, we incorporate an evidential deep learning mechanism in MVE, which outputs trust evaluation results while quantifying the subjective uncertainty underlying them. Finally, we employ a dynamic evidential fusion strategy to adaptively integrate the multi-view evidence based on their respective quantified uncertainties, thereby yielding a final trust evaluation for the collaborator. Extensive experiments demonstrate that the proposed MVE method outperforms baselines in both trust evaluation accuracy and task success rate.
I. Introduction
The paper frames trustworthy collaborator selection as a trust-evaluation problem involving heterogeneous, view-dependent, and uneven-quality historical data. It proposes MVE to model view-specific trust, quantify uncertainty, and dynamically fuse assessments.
- Collaborator selection is difficult because task complexity, system dynamism, and device heterogeneity complicate reliable performance assessment.
- Historical trust data forms distinct views because different task owners observe collaborators under task-specific conditions and varying noise levels.
- Aggregating all historical records discards source structure and can obscure context-dependent differences in collaborator performance.
- Fusion is challenging because view assessments differ in certainty, while averaging or weighted fusion may implicitly treat them as equally reliable.
- MVE evaluates collaborators from multiple views, quantifies trust and uncertainty, and dynamically fuses view-specific assessments.
- The method combines a multi-view framework, Mamba-based temporal modeling, and evidential learning for diverse-source and varying-quality data.
B. Problem Formulation
The system formulates collaborator selection as a constrained optimization problem. A selected collaborator must satisfy task requirements, including a minimum trust level and sufficient collaboration history.
- The system selects the most trustworthy collaborator through a constrained optimization formulation.
- The selected collaborator must meet the minimum trust level required by the task.
- Each potential collaborator must have collaborated with at least one task owner to provide minimum data for trust evaluation.
III. MVE for Trust Evaluation
MVE constructs collaborator-specific multi-view historical datasets and processes each view as a chronological trust sequence. It then applies temporal representation learning before uncertainty-aware evidential fusion.
- MVE constructs a multi-view historical dataset, learns view-specific temporal representations, quantifies uncertainty, and aggregates views through evidential fusion.
- Each task owner that previously collaborated with a collaborator is treated as an independent observation view.
- Historical records within each view are organized chronologically as a time-series sequence.
- Each collaboration record represents communication and computation outcomes, including a binary success indicator.
- View-specific trust sequences are obtained by computing trust scores for the collaborations within each task-owner view.
B. Temporal Trust Representation Learning via Mamba
MVE uses Mamba to encode each view’s chronological trust sequence into a view-level embedding. The representation pipeline projects scalar trust inputs, applies causal and selective state-space processing, and retains information adaptively.
- Mamba encodes each view-level trust sequence while offering linear sequence-length complexity and input-dependent state transitions.
- Each scalar trust value is projected into a d-dimensional embedding space using shared trainable parameters.
- The embedded sequence is fed into a Mamba block after input embedding.
- Mamba first processes the input with a causal convolution and SiLU activation.
- The model generates input-dependent state-space parameters and discretized state-transition matrices for temporal processing.
- Selective scanning and gated output processing produce temporal features that are projected back to the original feature dimension.
- Adaptive step sizes let Mamba retain or discard historical information, with the final output serving as the view-level embedding.
C. View-Specific Evidential Trust Learning
MVE converts each collaborator view’s temporal representation into an uncertainty-aware evidential opinion rather than a single over-confident trust estimate. Evidence is encoded with a Dirichlet distribution and represented through belief and uncertainty masses for subsequent fusion.
- C. View-Specific Evidential Trust Learning: Evidential deep learning replaces point trust estimates with view-specific opinions that express both trust-level support and unresolved uncertainty.Direct activation-based mapping can produce over-confident outputs that fail to represent a view’s confidence.
- C. View-Specific Evidential Trust Learning: Each view embedding is passed through a linear layer and nonnegative activation to generate a K-dimensional evidence vector.Softplus ensures nonnegative evidence, and each component represents evidence for one trust level.
- C. View-Specific Evidential Trust Learning: Collected evidence parameterizes a Dirichlet distribution over the probability simplex, with a uniform base-rate prior preventing parameters from collapsing without evidence.The resulting distribution represents uncertainty over the probability vector rather than only a single trust score.
- C. View-Specific Evidential Trust Learning: The Dirichlet parameters are transformed into a subjective opinion containing a belief mass vector and an uncertainty mass.A larger uncertainty mass indicates weaker evidence and lower confidence, while a smaller mass indicates stronger evidence.
- C. View-Specific Evidential Trust Learning: For every collaborator view, belief and uncertainty satisfy a valid mass assignment whose opinion records both trust-level preference and associated uncertainty.Applying the construction across views yields an uncertainty-aware characterization of collaborator trustworthiness.
D. Evidential Multi-View Trust Fusion
MVE combines task-owner-specific opinions using a conflictive evidential aggregation rule and then converts the fused opinion into a historical trust score. The rule preserves stronger evidence while accounting for less informative views and residual uncertainty.
- D. Evidential Multi-View Trust Fusion: MVE iteratively aggregates the diverse view-specific opinions of each collaborator using a conflictive opinion aggregation rule.The resulting fused opinion contains a combined belief vector and uncertainty mass.
- D. Evidential Multi-View Trust Fusion: A confident view with belief mass (0.05, 0.15, 0.70) and uncertainty 0.10 is fused with a flatter view having belief mass (0.20, 0.25, 0.25) and uncertainty 0.30.These values illustrate fusion between a strongly informative opinion and a less decisive one.
- D. Evidential Multi-View Trust Fusion: The example fusion produces belief mass (0.088, 0.175, 0.587) and uncertainty 0.150, with the third trust level remaining dominant.The result demonstrates preservation of confident evidence while mitigating uncertainty from the less informative view.
- D. Evidential Multi-View Trust Fusion: After fusion, residual uncertainty is redistributed across trust levels using the uniform base rate before projected probabilities are computed.The final historical trust score is the expectation of ordinal trust-level scores mapped to [0, 1].
E. Loss Function
MVE trains jointly on fused and view-specific evidential classification losses, while a consistency loss reduces conflicts among opinions. Trade-off hyperparameters balance these objectives.
- E. Loss Function: The loss combines evidential classification losses for the fused opinion and individual views with a consistency loss.The consistency term explicitly aligns trust evaluations across different opinions during training.
- E. Loss Function: Expected cross-entropy and annealed Kullback-Leibler divergence regularization reduce incorrect-label evidence and prevent premature convergence of misclassified instances to a uniform distribution.The classification losses are applied to both fused and individual-view evidential outputs.
- E. Loss Function: The hyperparameters β and γ control the trade-offs among the loss components.The complete derivation of the evidence-based loss and conflictive penalty is referred to prior work.
IV. Resource Evaluation and Collaborator Selection
For task-specific collaborator selection, the system combines historical trust with current resource trust. Resource trust evaluates whether available storage, CPU frequency, and collaborative willingness satisfy task requirements, and the highest final trust value determines selection.
- IV. Resource Evaluation and Collaborator Selection: The task owner’s resource evaluation uses each collaborator’s available storage size, CPU frequency, and collaborative willingness.These real-time resource dimensions are compared with the corresponding task requirements.
- IV. Resource Evaluation and Collaborator Selection: Resource trustworthiness is calculated with a multiplicative approach based on dimension-wise satisfaction degrees.Each satisfaction degree Φ_f belongs to {0, 1}.
- IV. Resource Evaluation and Collaborator Selection: The task-specific trust model jointly considers collaborators’ currently available resources and historical performance records.A trust server evaluates potential collaborators for the requested task and recommends an optimal collaborator.
- IV. Resource Evaluation and Collaborator Selection: The task owner selects the collaborator with the highest final trust value T_ai→a_j.The final decision uses the combined historical and resource trust values.
A. Experimental Setup
Experiments use two computationally different tasks, heterogeneous devices, controlled resource conditions, and five representative operational scenarios.
- Face recognition and virus scanning are evaluated with a default task size of 500 MB.
- DELL 5200, DELL 5820, and DELL 7060 devices are tested under controlled computational and network resources.HeavyLoad controls computational resources, while Clumsy controls network resources.
- Five operational scenarios—elite, stable, strategic, selfish, and failed—are constructed from device behaviour features.Each device executes 200 task instances under each scenario.
B. Macro-F1 and MAE Comparisons
MVE achieves the strongest historical trust evaluation performance and remains robust when uncertainty filtering or inter-view conflict challenges the assessments.
- Macro-F1 and MAE Comparisons: MVE outperforms DEF by 4.86% in Macro-F1 and 14.6% in MAE, achieving the best performance on both metrics.QS-Trust has the lowest performance, while TMC and DEF improve through evidential multi-view fusion mechanisms.
- Macro-F1 and MAE Comparisons: The evaluation compares MVE and baselines on historical trust using Macro-F1 and mean absolute error across 5 runs.
- Uncertainty Analysis: MVE provides the most discriminative uncertainty estimates for achieving higher accuracy.The uncertainty analysis evaluates accuracy after retaining samples with progressively lower uncertainty.
- Robustness to Conflict: Under inter-view conflict ratios varying from 0% to 50%, MVE degrades by only 17.1% and retains 68.6% accuracy at η = 50%.Conflict is introduced at inference time by bit-wise flipping selected views, while training uses clean data.
D. Task Success Rate Comparison Under Varying Requirements
MVE achieves higher task success rates than QS-Trust as task requirements increase, supported by task-specific resource trust evaluation and per-dimension capability checks.
- MVE yields a significantly higher task success rate than QS-Trust, with the performance gap widening as task requirements increase.The comparison varies the task requirements τ_thr and τ_CPU.
- Task-specific resource trust evaluation performs binary checks against each task demand rather than using a coarse-grained generic resource score.This filters candidates whose CPU frequency, bandwidth, or other critical capabilities fall short.
- MVE’s evaluations improve both trust evaluation accuracy and task success rate in extensive experiments.