Distributed multi-hop neighborhood extraction for graph machine-learning via push-lazy-push-traversal
A sampling procedure is performed for paths on a multi-hop distributed graph that includes vertices partitioned on a plurality of machines; a sampled path includes first and second vertices hosted on first and second machines respectively. The sampling procedure includes communicating, by the first machine to the second machine, first path information comprising an identifier for the second vertex and an identifier for a target vertex of the sampled path; the target vertex is hosted on a target host machine. The procedure further includes communicating edge information by the first machine to the target host machine, and communicating feature information by the second machine to the target host machine. Communication of the edge information and communication of the feature information are deferred relative to communication of the first path information.
1. A computer-implemented method comprising:
performing a sampling procedure for paths on a multi-hop distributed graph, the graph including a plurality of vertices partitioned on a plurality of machines, a sampled path comprising a first vertex hosted on a first machine of the plurality of machines and a second vertex hosted on a second machine of the plurality of machines and neighboring the first vertex, the sampling procedure including:
communicating, by the first machine to the second machine, first path information comprising an identifier for the second vertex and an identifier for a target vertex of the sampled path, the target vertex hosted on a target host machine of the plurality of machines;
storing, by the first machine in a first data structure, a first indication that communication of edge information regarding the first vertex and the second vertex is required;
storing, by the second machine in a second data structure, a second indication that communication of feature information regarding the second vertex is required;
communicating, by the first machine to the target host machine, the edge information; and
communicating, by the second machine to the target host machine, the feature information;
wherein communication of the edge information and communication of the feature information are deferred to a lazy push stage that occurs after communication of the first path information.
2. The computer-implemented method according to claim 1 , wherein at least one of the communication of the edge information and the communication of the feature information is performed subsequent to communication of the first path information.
3. The computer-implemented method according to claim 1 , wherein at least one of the communication of the edge information and the communication of the feature information is performed asynchronously with respect to communication of the first path information.
4. The computer-implemented method according to claim 1 , wherein the first vertex is the target vertex of the sampled path.
5. The computer-implemented method according to claim 1 , wherein the feature information is extracted from the graph.
6. The computer-implemented method according to claim 1 , wherein the feature information is extracted from a distributed key-value store.
7. The computer-implemented method according to claim 1 , wherein the sampling procedure is performed in accordance with a sampling configuration.
8. The computer-implemented method according to claim 1 , wherein the target host machine assembles a data batch based on the sampled path, the feature information and the edge information.
9. The computer-implemented method according to claim 8 , further comprising training a graph machine learning (ML) model based on the data batch.
10. A non-transitory computer-readable medium comprising instructions executable by a processor to perform a sampling procedure for paths on a distributed graph, the graph including a plurality of vertices partitioned on a plurality of machines, a sampled path comprising a first vertex hosted on a first machine of the plurality of machines and a second vertex hosted on a second machine of the plurality of machines and neighboring the first vertex, the sampling procedure including:
communicating, by the first machine to the second machine, first path information comprising an identifier for the second vertex and an identifier for a target vertex of the sampled path, the target vertex hosted on a target host machine of the plurality of machines;
storing, by the first machine in a first data structure, a first indication that communication of edge information regarding the first vertex and the second vertex is required;
storing, by the second machine in a second data structure, a second indication that communication of feature information regarding the second vertex is required;
communicating, by the first machine to the target host machine, the edge information; and
communicating, by the second machine to the target host machine, the feature information;
wherein communication of the edge information and communication of the feature information are deferred to a lazy push stage that occurs after communication of the first path information.
11. The non-transitory computer-readable medium of claim 10 , wherein at least one of the communication of the edge information and the communication of the feature information is performed subsequent to communication of the first path information.
12. The non-transitory computer-readable medium of claim 10 , wherein at least one of the communication of the edge information and the communication of the feature information is performed asynchronously with respect to communication of the first path information.
13. The non-transitory computer-readable medium of claim 10 , wherein the first vertex is the target vertex of the sampled path.
14. The non-transitory computer-readable medium of claim 10 , wherein the sampling procedure is performed in accordance with a sampling configuration.
15. The non-transitory computer-readable medium of claim 10 , wherein the target host machine assembles a data batch based on the sampled path, the feature information and the edge information, and wherein the instructions further comprise instructions for training a graph machine learning (ML) model based on the data batch.
16. A system comprising:
a processor; and
a memory that stores executable instructions that, when executed by the processor, facilitate performance of operations, the operations comprising:
performing a sampling procedure for paths on a multi-hop distributed graph, the graph including a plurality of vertices hosted on a plurality of machines, a sampled path comprising a first vertex hosted on a first machine of the plurality of machines and a second vertex hosted on a second machine of the plurality of machines and neighboring the first vertex, the sampling procedure including:
communicating, by the first machine to the second machine, first path information comprising an identifier for the second vertex and an identifier for a target vertex of the sampled path, the target vertex hosted on a target host machine of the plurality of machines;
storing, by the first machine in a first data structure, a first indication that communication of edge information regarding the first vertex and the second vertex is required;
storing, by the second machine in a second data structure, a second indication that communication of feature information regarding the second vertex is required;
communicating, by the first machine to the target host machine, the edge information; and
communicating, by the second machine to the target host machine, the feature information;
wherein communication of the edge information and communication of the feature information are deferred to a lazy push stage that occurs after communication of the first path information.
17. The system of claim 16 , wherein at least one of the communication of the edge information and the communication of the feature information is performed subsequent to communication of the first path information.
18. The system of claim 16 , wherein at least one of the communication of the edge information and the communication of the feature information is performed asynchronously with respect to communication of the first path information.
19. The system of claim 16 , wherein the first vertex is the target vertex of the sampled path.
20. The system of claim 16 , wherein the target host machine assembles a data batch based on the sampled path, the feature information and the edge information, and wherein the operations further comprise training a graph machine learning (ML) model based on the data batch.