PrototypeFederatedGraphLearning<T>
Models & Types · AiDotNet.FederatedLearning.Graph
Prototype-based federated graph learning: clients share class prototypes instead of full model parameters.
Inheritance: FederatedLearningComponentBase<T> → PrototypeFederatedGraphLearning<T>
For Beginners
Sharing full GNN model parameters raises two concerns: (1) model parameters can leak information about the training graph, and (2) different clients may have very different graph structures that benefit from different model architectures. Prototype-based FGL addresses both:
How It Works
Benefits: Prototypes are much smaller than full models (K prototypes of dimension D vs. millions of parameters), more robust to topology heterogeneity, and reveal less about graph structure.
Constructors
| Constructor | Summary |
|---|---|
PrototypeFederatedGraphLearning(FederatedGraphOptions options, int prototypeDim, int numClasses) |
Initializes a new instance of PrototypeFederatedGraphLearning. |
Properties
| Property | Summary |
|---|---|
IReadOnlyDictionary<int, Tensor<T>> GlobalPrototypes { get; } |
Gets the current global prototypes. |
Methods
| Method | Summary |
|---|---|
Dictionary<int, Tensor<T>> AggregatePrototypes() |
Aggregates prototypes from all registered clients into global prototypes. |
Dictionary<int, Tensor<T>> ComputePrototypes(Tensor<T> nodeEmbeddings, Tensor<T> labels, int embeddingDim) |
Computes a client’s class prototypes from node embeddings and labels. |
Dictionary<int, Tensor<T>> ComputePrototypes(Tensor<T> nodeEmbeddings, Tensor<T> labels, int embeddingDim, Dictionary<int, int> classSampleCounts) |
Computes a client’s class prototypes and the number of nodes behind each one (the weights used by RegisterClientPrototypes). |
double ComputePrototypeLoss(Dictionary<int, Tensor<T>> clientPrototypes, double lambda) |
Computes prototype regularization loss: pull client prototypes toward global prototypes. |
void RegisterClientPrototypes(int clientId, Dictionary<int, Tensor<T>> prototypes, Dictionary<int, int> classSampleCounts) |
Registers a client’s class prototypes for the current round. |