LightweightShapleyEvaluator<T>
Models & Types · AiDotNet.FederatedLearning.Fairness
Implements Lightweight Shapley — O(n) Shapley value approximation using gradient similarity.
Inheritance: FederatedLearningComponentBase<T> → LightweightShapleyEvaluator<T>
Implements: IClientContributionEvaluator<T>
For Beginners
Exact Shapley values require evaluating all 2^n subsets of clients — impossibly expensive for even 20 clients. Lightweight Shapley approximates each client’s contribution by measuring how similar their gradient is to the ideal global gradient. Clients whose updates are well-aligned with the consensus get high Shapley values, while adversarial or low-quality updates get low values. This runs in O(n) time.
How It Works
Approximation:
Reference: Lightweight Shapley for Federated Contribution Evaluation (2024).
Constructors
| Constructor | Summary |
|---|---|
LightweightShapleyEvaluator(double freeRiderThreshold) |
Creates a new Lightweight Shapley evaluator. |
Properties
| Property | Summary |
|---|---|
double FreeRiderThreshold { get; } |
Gets the free-rider threshold. |
string MethodName { get; } |
Methods
| Method | Summary |
|---|---|
Dictionary<int, double> EvaluateContributions(Dictionary<int, Tensor<T>> clientModels, Tensor<T> globalModel, Dictionary<int, List<Tensor<T>>> clientHistories) |
|
HashSet<int> IdentifyFreeRiders(Dictionary<int, double> contributionScores) |