RelationalAggregation

Enums · AiDotNet.Enums

How an R-GCN layer normalizes the messages a node receives along one relation type: the c_{i,r} of Schlichtkrull et al. 2018, Eq. 2, h_i’ = sigma(W_0 h_i + sum_r sum_{j in N_i^r} (1/c_{i,r}) W_r h_j).

For Beginners

Each node adds up the transformed states of its neighbours under every relation. Sum uses that sum as it is; Mean divides it by how many neighbours the node has under that relation, so a node with many neighbours is not drowned out by sheer count.

Fields

Field Summary
RelationalAggregation Mean c_{i,r} = the node’s weighted in-degree under relation r, which for a 0/1 adjacency is the |N_i^r| the paper uses.
RelationalAggregation Sum c_{i,r} = 1: the adjacency weights are used as given, so pass a pre-normalized adjacency for any other normalization.
int value__