TiltedERMFairness<T>
Models & Types · AiDotNet.FederatedLearning.Fairness
Implements TERM (Tilted Empirical Risk Minimization) for fairness in FL.
Inheritance: FederatedLearningComponentBase<T> → TiltedERMFairness<T>
For Beginners
TERM smoothly interpolates between average and worst-case optimization using a tilt parameter t. When t=0, it’s standard average loss. When t>0, it up-weights high-loss clients (moving toward worst-case). When t<0, it focuses on easy clients (useful for outlier robustness). This gives a smooth, differentiable fairness objective that’s easier to optimize than the hard minimax of AFL.
How It Works
Objective:
Reference: Li, T., et al. (2021). “Tilted Empirical Risk Minimization.” ICLR 2021.
Constructors
| Constructor | Summary |
|---|---|
TiltedERMFairness(double tilt) |
Creates a new TERM fairness instance. |
Properties
| Property | Summary |
|---|---|
double Tilt { get; } |
Gets the tilt parameter. |
Methods
| Method | Summary |
|---|---|
Dictionary<int, double> ComputeWeights(Dictionary<int, double> clientLosses) |
Computes TERM aggregation weights based on client losses. |
double ComputeObjective(Dictionary<int, double> clientLosses) |
Computes the TERM objective value. |