Source-linked AI summary
Long-Term Behavioral Evaluation for Trusted Collaborator Selection via Bidirectional Mamba
Botao Zhu, Xianbin Wang
TL;DR
The paper addresses the challenge of evaluating device behavior from limited historical collaborations and one-directional temporal information. It proposes BM, which builds short-slot collaboration graphs and fuses their behavioral representations with bidirectional Mamba. BM achieves higher accuracy and stability than baseline methods, supporting collaborator selection that maximizes VoC.
Problem
Accurate collaborator selection requires long-term behavioral evaluation that captures both forward and backward temporal dependencies, beyond limited or one-directional historical assessments.
Method
BM aggregates device behavioral features from collaboration graphs constructed in short time slots, then fuses them across the full time horizon with bidirectional Mamba.
Results
BM achieves the lowest RMSE and MAE, stable long-term performance, and the highest VoC across trust-threshold levels in the reported comparisons.
Takeaways & Limitations
BM supports reliable collaborator selection by producing stable long-term behavioral evaluations that help maximize task-completion value.
Abstract
from arXiv · showhide
Effective selection of trustworthy collaborators is crucial to ensuring the successful completion of collaborative tasks, which requires accurate assessments of both long-term device behavior and short-term collaborative dynamics. Consistent device behavior patterns, which are learned from historical collaborations, can be used to predict their reliability in future collaborations. However, accurately assessing device behavior based on historical collaborations remains challenging. First, behavior assessment from limited historical collaborations captures only instantaneous past behavior, failing to represent the devices' true behavior. Second, due to the temporal dependencies of device behavior, a unidirectional evaluation that relies only on earlier collaborations loses the opportunity to learn from subsequent collaborations. Addressing these challenges requires evaluating device behavior based on long-term collaborations while considering both forward and backward temporal dependencies. To this end, this work proposes a bidirectional Mamba-enabled model (BM) for long-term behavioral evaluation. For each short time slot, a graph is constructed among devices based on historical collaborations, and device behavioral features within the slot are then aggregated accordingly. Subsequently, a bidirectional Mamba model integrates these short-term representations across all time intervals, producing a stable and reliable long-term behavior evaluation for each device. Experimental results demonstrate that BM achieves higher evaluation accuracy than baseline methods, thereby enabling the selection of collaborators that maximize the value of task completion.
I. INTRODUCTION
Reliable collaborator selection requires evaluating both long-term device behavior and short-term collaborative dynamics. BM addresses temporal and relational challenges by aggregating within-slot behavior and fusing it across time with bidirectional Mamba.
- Historical collaborations reveal behavioral patterns that help predict a device’s reliability in future collaborative tasks.
- Long-term evaluation is needed because device behavior develops across multiple short collaborations, while temporal features have both forward and backward dependencies.
- BM constructs collaboration graphs for short time slots and aggregates devices’ historical reliability through their collaboration relationships.
- The framework uses historical behavior to produce accurate assessments for selecting collaborators.
- Bidirectional Mamba fuses short-term representations across long sequences while efficiently capturing temporal dependencies.
A. Trust Model
The trust model evaluates collaborator reliability from historical behavior and available resources, while task execution models quantify transmission and computation costs.
- A collaborator’s trustworthiness depends on its past collaborative behavior and available resources.
- Historical reliability and task-specific resource trustworthiness are represented as separate components of the trust evaluation.
- The model computes transmission performance using channel bandwidth, transmission power, noise power, and channel gain between devices.
- Task execution includes transmission time and energy, followed by local computation time and energy at the collaborator.
- The computation-energy coefficient ε is set to 10^-11 according to measurements.
C. Value as a Metric
Value of task completion (VoC) measures the task owner’s satisfaction with a collaborator’s execution outcome using time and energy considerations.
- VoC quantifies the task owner’s satisfaction when collaborator kj executes task θ.
- When collaborator execution takes longer than local execution, time-based satisfaction falls below 1 and decreases as execution time increases.
- Energy-based satisfaction compares the collaborator’s energy consumption with the task owner’s local energy consumption.
D. Problem Formulation
The formulation selects a trusted collaborator that maximizes task-completion value while satisfying a minimum trust threshold and using collected historical collaboration data.
- The optimization objective is to identify a trusted collaborator that maximizes VoC for the task owner.
- The selected collaborator must satisfy the minimum trust threshold θtrust.
- Historical evaluation must rely entirely on collected historical collaboration data.
III. BIDIRECTIONAL MAMBA-ENABLED LONG TERM BEHAVIORAL EVALUATION
The BM model evaluates long-term device behavior by combining fine-grained short-term collaboration features with bidirectional Mamba processing. Its design uses state-space modeling to capture temporal dependencies efficiently across extended sequences.
- Mamba captures long-range temporal dependencies while maintaining linear computational cost.Its selective state-space mechanism adaptively retains or discards information at each time step.
- BM constructs historical collaboration graphs, fuses device reliability within each time slot, and applies Mamba for long-term evaluation.The three stages are shown as graph-sequence construction, within-slot reliability fusion, and long-term reliability evaluation.
- The continuous-time state-space model uses input sequences and learnable parameter matrices before discretization for discrete inputs.The discrete formulation is obtained with the zero-order hold method and a specified sampling timescale.
B. Historical Collaboration Graph Sequence Construction
The model represents historical collaborations as a chronological sequence of weighted graphs. Edge weights encode direct reliability observations, while graph propagation aggregates neighboring devices’ recommendations into time-slot behavior features.
- The continuous history is divided into S equal-length slots, each represented by a collaboration graph based on interactions in that slot.Graphs are arranged chronologically as {G(t1), . . . , G(tS)} to form the historical collaboration sequence.
- A directed edge from ki to kj is added when kj assists ki, and its weight represents kj’s reliability as evaluated by ki.The weight is computed from direct collaboration records between the devices.
- Edge reliability combines transmission and computation success using weights α1 and α2 satisfying α1 + α2 = 1.Transmission and computation outcomes are binary, and N(ki,kj) counts tasks received by kj from ki.
- Within each time slot, a device aggregates messages from its one-hop in-degree neighbors to infer its historical reliability.Each message incorporates the recommending device’s embedding, the interaction encoding, and the time-slot encoding.
- Stacking L propagation and aggregation layers enables multi-hop information flow and produces a device embedding for each time slot.Concatenating the slot embeddings yields a representation X_kj that captures temporal dynamics across the historical period.
D. Bidirectional Mamba–Aided Long-Term Behavior Fusion
The long-term fusion module integrates each device’s sequence of short-term embeddings with forward and backward Mamba scans. Residual fusion, pooling, and an MLP then produce long-term reliability evaluations trained against historical ground truth.
- Bidirectional Mamba captures both forward and backward temporal dependencies in each device’s historical embedding sequence.Each layer contains separate forward and backward Mamba blocks.
- The Mamba outputs are fused with the original sequence through a residual connection to produce the final temporal embedding Y_kj.The fused representation remains organized across the S time slots.
- Max pooling over the temporal representation yields a single final device embedding in R^da.This embedding summarizes the most significant features across the long-term sequence.
- The long-term reliability of kj from ki’s perspective is predicted by concatenating their final embeddings and feeding them to an MLP.The resulting output vector is reduced to the maximum reliability value used for evaluation.
- BM is trained by minimizing cross-entropy between computed reliability values and ground-truth historical values.The computed and observed reliability sets are denoted T_beh and T_col, respectively.
E. Task-Specific Resource Trust Evaluation
Task-specific resource trust complements long-term behavioral evaluation by checking whether a potential collaborator is willing and has adequate communication and computation resources. Devices meeting the trust threshold are then compared by task value of completion.
- Resource trust evaluates each potential collaborator through willingness, communication resource, and computation resource components.These components reflect the collaborator’s willingness to participate and its ability to transmit and execute the task.
- Willingness trust is binary: it is one when kj is willing to collaborate and zero otherwise.
- Communication-resource trust considers channel quality, bandwidth, data rate, latency, jitter, and packet loss.Together, these factors indicate whether task transmission can be stable and timely.
- Computation-resource trust considers processing capacity, memory, storage availability, and operational stability.These factors indicate whether the collaborator can execute assigned tasks reliably and efficiently.
- A collaborator receives resource trust only when all requirements are met; eligible devices are filtered by θ_trust and the one maximizing VoC is selected.
A. Experimental Settings
The experiments evaluate BM against baseline methods using simulated wireless-system settings and compare accuracy and long-term stability. Results are reported with RMSE, MAE, and VoC-related comparisons.
- The wireless system is implemented in the NS-3 discrete-event network simulator with Python bindings and includes 500 devices.Devices use 100 mW transmission power, 80 mW reception power, and CPU frequencies randomly selected from {2, 4, 6} GHz.
- Figure 2 compares evaluation accuracy using RMSE and MAE, with BM achieving the lowest values.
- Figure 3 evaluates RMSE and VoC over a longer time horizon, showing BM’s fluctuation and VoC performance.
- The reported experimental results use a learning rate of 10^-2, an L2 regularization coefficient of 10^-5, and a dropout rate of 0.
B. Comparison of Evaluation Accuracy
BM is evaluated for historical reliability accuracy and long-term performance. It achieves the strongest reported accuracy and stability, while consistently producing the highest VoC across trust thresholds.
- B. Comparison of Evaluation Accuracy: BM achieves the lowest RMSE and MAE values in historical reliability assessment, outperforming the baseline methods.The results average performance over 10 runs.
- B. Comparison of Evaluation Accuracy: LSTM outperforms GNN and QS-Trust but remains inferior to BM in evaluation accuracy.The passage attributes GNN’s limitation to spatial-only fusion and QS-Trust’s to rule-based evaluation without temporal dynamics.
- C. Long-Term Performance Comparison: Across 500 time slots, BM shows the smallest fluctuation and consistently lower RMSE than the comparison algorithms.These results indicate stable and accurate long-term inference.
- C. Long-Term Performance Comparison: As the minimum trust threshold θtrust increases, BM consistently achieves the highest VoC across all threshold levels.
- V. CONCLUSION: BM combines short-term reliability features with bidirectional Mamba fusion to produce stable long-term evaluations supporting collaborator selection that maximizes VoC.