RepuLink is a network-based reputation system that combines traditional forward propagation with a novel backward propagation mechanism to adjust reputations based on both interactions and endorsements. The algorithm leverages endorsement penalty and reward signals to provide a more robust, calibrated, and explainable reputation evaluation in decentralized networks.
In decentralized networks, establishing trust between unknown or partially trusted nodes is a key challenge. Repulink introduces a two-layer reputation model that:
-
Integrates multiple types of feedback:
- Interaction data: Captures individual feedback (positive/negative) between nodes.
- Endorsement data: Captures social trust signals through endorsements.
-
Applies backward propagation mechanisms:
- Penalty Propagation (BEPP): Propagates negative feedback from misbehaving nodes back to endorsers.
- Reward Propagation (BERP): Rewards endorsers when the nodes they endorse perform well.
-
Calibrates final reputations:
- Combines forward reputation computed via power iteration with backward adjustments.
- Maintains normalized reputation distributions to reflect network performance.
The simulation runs via the main driver script main.py, which loads processed data, instantiates managers, performs reputation computation, and records node histories.
To run the simulation and save output to a file, you can either use command-line redirection or configure logging in the script.
Example using command-line redirection:
python3 repulink_example_full.py Check the results saved under the folder repulink_multislot_core_modules.
All configurable parameters are stored in config.py, including:
EPSILON: Small constant to prevent division by zero.ALPHA: Weight between local trust and endorsement contribution.LAMBDA: Reward sensitivity coefficient.BETA: Penalty sensitivity coefficient.GAMMA: Penalty (reward) discount factor.MAX_ITER: Maximum iterations for convergence.CONVERGENCE_THRESHOLD: Convergence threshold for the reputation vector.
Feel free to adjust these parameters as needed for experimentation and simulation tuning.