graph_neural_networks (1321B)
1 # Graph neural networks (GNN) 2 3 A kind of [distributed_algorithm] on a [graph], related to [message_passing], to compute a [function] 4 5 - initially every node of the graph is mapped to a vector of [real_numbers] 6 - at every step, you [aggregate] coordinatewise the state of the [neighbors], and then you use a [combination_function] with the previous node state and the aggregate states of the neighbors 7 - the [aggregation] function can be [sum], [mean], or [max] (coordinatewise) 8 - the [combination_function] is computed by a [feedforward_neural_network] (multilayer [perceptron]) 9 10 Extensions: 11 - [global_aggregation]: add a value which is obtained by aggregating the state of all nodes, and is passed to the [combination_function] 12 - [random_initialization]: include in the initial label of the nodes one [random] number 13 14 The network computes a value on every node. We can also use a [readout_function] to get a final value out of the entire graph: computed by a [feedforward_neural_network] from the result of applying an [aggregation_function] to the state of all nodes 15 16 Observation: 17 - On [graph_isomorphic] [graphs], the value of the function computed by a GNN is the same 18 19 - [graph_neural_network_expressiveness] 20 21 Other model: [GNN_recurrent] 22 23 Up: [graph], [neural_network] 24 25 Aliases: graph neural network, GNN, GNNs