VerticalPartyLabelHolder<T>

Models & Types · AiDotNet.FederatedLearning.Vertical

Represents the label-holding party in vertical federated learning.

Inheritance: FederatedLearningComponentBase<T> → VerticalPartyLabelHolder<T>

Implements: IVerticalParty<T>

For Beginners

In VFL, one special party holds the labels (prediction targets). For example, a hospital knows patient outcomes while partner banks only have financial data. The label holder plays a unique role:

How It Works

Security note: The label holder’s gradients can reveal label information. For example, large gradient magnitudes suggest the model made a large error, which hints at the true label. Always use label protection in production deployments.

Constructors

Constructor Summary
VerticalPartyLabelHolder(string partyId, Tensor<T> localData, Tensor<T> labels, IReadOnlyList<string> entityIds, int embeddingDimension, int? seed) Initializes a new instance of VerticalPartyLabelHolder.

Properties

Property Summary
bool IsLabelHolder { get; }
int EmbeddingDimension { get; }
int FeatureCount { get; }
string PartyId { get; }

Methods

Method Summary
IReadOnlyList<Tensor<T>> GetParameters()
IReadOnlyList<string> GetEntityIds()
Tensor<T> ComputeForward(IReadOnlyList<int> alignedIndices)
Tensor<T> GetLabels(IReadOnlyList<int> alignedIndices) Gets the labels for the specified aligned entity indices.
ValueTuple<double, Tensor<T>> ComputeLoss(Tensor<T> predictions, Tensor<T> batchLabels) Computes the mean squared error loss and its gradient.
void ApplyBackward(Tensor<T> gradients, double learningRate)
void SetParameters(IReadOnlyList<Tensor<T>> parameters)