Graph neural network training methods and systems
View Patent ↗Methods, systems, and apparatus for training a graph neural network. An example method includes obtaining a complete graph; dividing the complete graph into a plurality of subgraphs; obtaining a training graph to participate in graph neural network training based on selecting at least one subgraph from the plurality of subgraphs; obtaining, based on the training graph, a node feature vector of each node in the training graph; obtaining a node fusion vector of each current node in the training graph; determining a loss function based on node labels and the node fusion vectors in the training graph; and iteratively training the graph neural network to update parameter values of the graph neural network based on optimizing the loss function.
1. A computer-implemented method for training a graph neural network, wherein the method comprises:
obtaining a graph that comprises a plurality of nodes and edges between the plurality of nodes that represent relationships between the plurality of nodes;
dividing the graph into a plurality of subgraphs by using a community discovery algorithm, wherein dividing the graph comprises grouping nodes that are more related to each other into a same subgraph;
obtaining at least one subgraph from the plurality of subgraphs;
obtaining, for each node in the at least one subgraph, a node feature vector;
obtaining, for each node in the at least one subgraph, a node fusion vector based on performing propagation and aggregation using the node feature vector for each node in the at least one subgraph and edges connected to each node in the at least one subgraph; and
training the graph neural network by using the at least one subgraph and based on optimizing a loss function that is computed based on the node fusion vector for each node in the at least one subgraph.
2. The computer-implemented method according to claim 1 , wherein the community discovery algorithm is one of a label propagation algorithm, a Girvan-Newman algorithm, or a hop attention & node preference algorithm.
3. The computer-implemented method according to claim 1 , wherein the plurality of nodes corresponds to respective samples and the edges between the plurality of nodes correspond to relationships between the respective samples.
4. The computer-implemented method according to claim 3 , wherein the relationships between the respective samples measure a frequency of interactions between the respective samples.
5. The computer-implemented method according to claim 1 , wherein obtaining the graph comprises:
obtaining sample data; and
generating the graph based on the sample data.
6. The computer-implemented method according to claim 1 , wherein obtaining the graph comprises:
obtaining the graph from a database, including from a database on a terminal device, or from a data source.
7. The computer-implemented method according to claim 1 , wherein obtaining the at least one subgraph from the plurality of subgraphs comprises:
obtaining two or more subgraphs from the plurality of subgraphs; and
determining a union set of the two or more subgraphs.
8. A computer-implemented system, comprising:
one or more computers; and
one or more computer memory devices interoperably coupled with the one or more computers and having tangible, non-transitory, machine-readable media storing one or more instructions that, when executed by the one or more computers, perform one or more operations for training a graph neural network, wherein the operations comprise:
obtaining a graph that comprises a plurality of nodes and edges between the plurality of nodes that represent relationships between the plurality of nodes;
dividing the graph into a plurality of subgraphs by using a community discovery algorithm, wherein dividing the graph comprises grouping nodes that are more related to each other into a same subgraph;
obtaining at least one subgraph from the plurality of subgraphs;
obtaining, for each node in the at least one subgraph, a node feature vector;
obtaining, for each node in the at least one subgraph, a node fusion vector based on performing propagation and aggregation using the node feature vector for each node in the at least one subgraph and edges connected to each node in the at least one subgraph; and
training the graph neural network by using the at least one subgraph and based on optimizing a loss function that is computed based on the node fusion vector for each node in the at least one subgraph.
9. The computer-implemented system according to claim 8 , wherein the community discovery algorithm is one of a label propagation algorithm, a Girvan-Newman algorithm, or a hop attention & node preference algorithm.
10. The computer-implemented system according to claim 8 , wherein the plurality of nodes corresponds to respective samples and the edges between the plurality of nodes correspond to relationships between the respective samples.
11. The computer-implemented system according to claim 10 , wherein the relationships between the respective samples measure a frequency of interactions between the respective samples.
12. The computer-implemented system according to claim 8 , wherein obtaining the graph comprises:
obtaining sample data; and
generating the graph based on the sample data.
13. The computer-implemented system according to claim 8 , wherein obtaining the graph comprises:
obtaining the graph from a database, including from a database on a terminal device, or from a data source.
14. The computer-implemented system according to claim 8 , wherein obtaining the at least one subgraph from the plurality of subgraphs comprises:
obtaining two or more subgraphs from the plurality of subgraphs; and
determining a union set of the two or more subgraphs.
15. A non-transitory, computer-readable medium storing one or more instructions executable by a computer system to perform operations for training a graph neural network, wherein the operations comprise:
obtaining a graph that comprises a plurality of nodes and edges between the plurality of nodes that represent relationships between the plurality of nodes;
dividing the graph into a plurality of subgraphs by using a community discovery algorithm, wherein dividing the graph comprises grouping nodes that are more related to each other into a same subgraph;
obtaining at least one subgraph from the plurality of subgraphs;
obtaining, for each node in the at least one subgraph, a node feature vector;
obtaining, for each node in the at least one subgraph, a node fusion vector based on performing propagation and aggregation using the node feature vector for each node in the at least one subgraph and edges connected to each node in the at least one subgraph; and
training the graph neural network by using the at least one subgraph and based on optimizing a loss function that is computed based on the node fusion vector for each node in the at least one subgraph.
16. The non-transitory, computer-readable medium according to claim 15 , wherein the community discovery algorithm is one of a label propagation algorithm, a Girvan-Newman algorithm, or a hop attention & node preference algorithm.
17. The non-transitory, computer-readable medium according to claim 15 , wherein the plurality of nodes corresponds to respective samples and the edges between the plurality of nodes correspond to relationships between the respective samples.
18. The non-transitory, computer-readable medium according to claim 17 , wherein the relationships between the respective samples measure a frequency of interactions between the respective samples.
19. The non-transitory, computer-readable medium according to claim 15 , wherein obtaining the graph comprises:
obtaining sample data; and
generating the graph based on the sample data.
20. The non-transitory, computer-readable medium according to claim 15 , wherein obtaining the graph comprises:
obtaining the graph from a database, including from a database on a terminal device, or from a data source.