IVerticalParty<T>

Interfaces · AiDotNet.FederatedLearning.Vertical

Represents a party in vertical federated learning that holds a subset of features.

Known implementations: VerticalPartyClient<T>, VerticalPartyLabelHolder<T>

For Beginners

In vertical FL, each party holds different features (columns) for the same entities (rows). For example, a bank has income and credit score columns, while a hospital has diagnosis and prescription columns. Each party runs a local “bottom model” on its features to produce an embedding (a compressed representation).

How It Works

The party interface abstracts away the local computation: the VFL trainer asks each party to compute forward passes on its local data and backward passes using gradients received from the coordinator.

Properties

Property Summary
bool IsLabelHolder { get; } Gets whether this party holds the labels for supervised learning.
int EmbeddingDimension { get; } Gets the output dimension of this party’s bottom model (embedding size).
int FeatureCount { get; } Gets the number of features held by this party.
string PartyId { get; } Gets the unique identifier for this party.

Methods

Method Summary
IReadOnlyList<Tensor<T>> GetParameters() Gets the current parameters of this party’s bottom model for checkpointing.
IReadOnlyList<string> GetEntityIds() Gets the entity IDs that this party has data for.
Tensor<T> ComputeForward(IReadOnlyList<int> alignedIndices) Computes the forward pass of this party’s bottom model on the given entity indices.
void ApplyBackward(Tensor<T> gradients, double learningRate) Applies the backward pass using gradients received from the coordinator.
void SetParameters(IReadOnlyList<Tensor<T>> parameters) Sets the parameters of this party’s bottom model (for loading checkpoints or unlearning).