IGraphAggregationStrategy<T>
Interfaces · AiDotNet.FederatedLearning.Graph
Graph-aware model aggregation strategy for federated GNN training.
Known implementations: FedGnnAggregationStrategy<T>
For Beginners
Standard FedAvg treats all model parameters equally during aggregation. But GNNs have distinct parameter types — message-passing weights, attention heads, node embeddings, readout layers — that benefit from different aggregation treatment.
How It Works
Example: Message-passing layers should be weighted by subgraph size (number of edges), while readout layers should be weighted by number of labeled nodes. A graph-aware strategy handles this distinction.
Properties
| Property | Summary |
|---|---|
string StrategyName { get; } |
Gets the name of this graph aggregation strategy. |
Methods
| Method | Summary |
|---|---|
Tensor<T> Aggregate(Dictionary<int, Tensor<T>> clientModels, Dictionary<int, ClientGraphStats> clientGraphStats) |
Aggregates GNN model parameters from multiple clients with graph-aware weighting. |