GraphNodeGenerator<T>
Models & Types · AiDotNet.FederatedLearning.Graph
Generates pseudo-node features for missing cross-client neighbors using a learned model.
Inheritance: FederatedLearningComponentBase<T> → GraphNodeGenerator<T>
For Beginners
When a GNN on Client A needs features from a node on Client B (a cross-client neighbor), it can’t directly access them. This generator learns to produce realistic pseudo-node features based on the local graph structure:
How It Works
Training: The generator is trained locally on each client using known edges — it learns to predict a node’s features from its neighbors’ features. Then it’s used to predict features for missing (cross-client) neighbors.
Homophily assumption: Works best on homophilic graphs where connected nodes tend to have similar features (social networks, citation networks). Less effective on heterophilic graphs.
Constructors
| Constructor | Summary |
|---|---|
GraphNodeGenerator(int featureDim, int hiddenDim, double learningRate) |
Initializes a new instance of GraphNodeGenerator. |
Properties
| Property | Summary |
|---|---|
bool IsTrained { get; } |
Gets a value indicating whether the generator has been trained. |
Methods
| Method | Summary |
|---|---|
Tensor<T> Generate(Tensor<T> borderNodeFeatures) |
Generates a pseudo-node feature vector for a missing cross-client neighbor. |
void Train(Tensor<T> adjacency, Tensor<T> nodeFeatures, int featureDim, int epochs) |
Trains the generator on known node-neighbor pairs from the local subgraph. |