AgnosticFairnessObjective<T>
Models & Types · AiDotNet.FederatedLearning.Fairness
Implements AFL (Agnostic Federated Learning) — minimax fairness optimization.
Inheritance: FederatedLearningComponentBase<T> → AgnosticFairnessObjective<T>
For Beginners
Standard FL minimizes the average loss across all clients. This can leave some clients with terrible performance (e.g., a hospital with rare diseases). AFL instead optimizes for the worst-performing client — it’s “agnostic” to which client distribution the model will be tested on. This ensures no client is left behind, at the cost of slightly lower average performance.
How It Works
Objective:
The lambda weights dynamically increase for clients with high loss.
Reference: Mohri, M., et al. (2019). “Agnostic Federated Learning.” ICML 2019.
Constructors
| Constructor | Summary |
|---|---|
AgnosticFairnessObjective(double lambdaLearningRate) |
Creates a new AFL objective. |
Properties
| Property | Summary |
|---|---|
double LambdaLearningRate { get; } |
Gets the lambda learning rate. |
Methods
| Method | Summary |
|---|---|
Dictionary<int, double> ComputeWeights(Dictionary<int, double> clientLosses) |
Computes AFL aggregation weights based on client losses. |