Fig. 7: Model architecture.
From: Enhancing link prediction in biomedical knowledge graphs with BioPathNet
a, Overview: to predict a query relation between a source node and other nodes, the source node is initialized with the query embedding, while others are set to zero (boundary condition). Message passing is performed over t GNN layers (for example, t = 4), producing embeddings that represent paths from the source node. These embeddings are then fed into a multilayer perceptron (MLP) to predict the query relation. b, Message passing (for example, using the DistMult function (multiplication) for messages and summation for aggregation): the embedding of node v (relative to source node u) is updated on the basis of incoming messages from neighbours x1, x2 and the boundary condition. Messages are computed by multiplying neighbour embeddings (\({h}_{{x}_{1}}^{(t-1)}\), \({h}_{{x}_{2}}^{(t-1)}\)) with relation embeddings (wq(x1, r1, v), wq(x2, r2, v)) (purple), aggregated via summation (pink) and combined with \({h}_{v}^{(0)}\) through a linear layer, layer norm and activation function (blue). Relation embeddings (wq(x1, r1, v), wq(x2, r2, v)) can be learned dynamically dependently on the query q as Wrq + br.
