Predictive model importation

• US 10,157,343 B1
• Filed: 06/16/2015
• Issued: 12/18/2018
• Est. Priority Date: 05/09/2011
• Status: Active Grant

• Current Assignee: Google LLC (Alphabet Inc.)
• Original Assignee: Google LLC (Alphabet Inc.)
• Inventors: Lin, Wei-Hao, Green, Travis H. K., Kaplow, Robert, Fu, Gang, Mann, Gideon S.
• Primary Examiner(s): Holmes, Michael B
• Application Number:
  US14/740,942
• Time in Patent Office: 1,281 Days
• Field of Search: 706 49
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 3/08 Learning methods
• CPC Class Codes: G06N 5/04 Inference or reasoning models
• CPC Class Codes: H04L 67/10 in which an application is distributed across nodes in the network software deployment G06F8/60; multiprogramming arrangements G06F9/46

Abstract

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for obtaining a plurality of model representations of predictive models, each model representation associated with a respective user and expresses a respective predictive model, and selecting a model implementation for each of the model representations based on one or more system usage properties associated with the user associated with the corresponding model representation.

Claims

1. A computer-implemented method, the method comprising:

for each representation of a plurality of representations of predictive models, selecting a model implementation from a plurality of model implementation for the representation, wherein each representation is associated with a respective user and comprises a description of a respective predictive model, the selection based on a status of the representation'"'"'s respective user;

associating each of the model implementations with a respective node in a directed graph wherein for one or more ordered pairs of nodes in the graph the prediction output of a model implementation associated with a first node in the pair serves as input to a model implementation associated with a second node in the pair;

determining, from the directed graph, a subset of model implementations that can be executed in parallel based on edge dependencies in the directed graph; and

causing the models in the subset of model implementations to be executed in parallel.

2. The computer-implemented method of claim 1, wherein for one or more ordered pairs of nodes in the graph the prediction output of a model implementation associated with a tail node in the pair serves as input to a model implementation associated with a head node in the pair.

3. The computer-implemented method of claim 1, further comprising:

assigning a system resource quota to at least one respective user;

determining an amount of system resources that the model implementations for the respective user have consumed over a given period of time; and

selecting the model implementation based on an amount of unconsumed system resources

according to the system resource quota, wherein an amount of system resources that can be consumed by the selected model implementation does not exceed the amount of unconsumed system resources.

4. The computer-implemented method of claim 1, further comprising:

selecting a computationally-intensive model implementation for the model representation based on a determination that system resources were previously used during an off-peak time of day.

5. The computer-implemented method of claim 1, wherein:

for at least a first node in the directed graph the node is realized by special purpose circuitry, and a model implementation associated with the first node is configured to be executed by the special purpose circuitry; and

for at least a second node in the directed graph the node is realized by a general purpose microprocessor, and a model implementation associated with the second node is configured to be executed by the general purpose microprocessor.

6. The computer-implemented method of claim 5, wherein the special purpose circuitry is an ASIC.

7. A computer-implemented method, the method comprising:

for each representation of a plurality of representations of predictive models, selecting a model implementation from a plurality of model implementation for the representation, wherein each representation comprises a description of a respective predictive model;

associating each of the model implementations with a respective node in a directed graph wherein for one or more ordered pairs of nodes in the graph the prediction output of a model implementation associated with a first node in the pair serves as input to a model implementation associated with a second node in the pair, wherein;

for at least a first node in the directed graph the node is realized by special purpose circuitry, and a model implementation associated with the first node and configured to be executed by the special purpose circuitry is selected for the first node based on its configuration to be executed by the special purpose circuitry; and

for at least a second node in the directed graph the node is realized by a general purpose microprocessor, and a model implementation associated with the second node and configured to be executed by the general purpose microprocessor is selected for the second node based on its configuration to be executed by the general purpose microprocessor.

8. The computer-implemented method of claim 7, further comprising:

determining, from the directed graph, a subset of model implementations that can be executed in parallel based on edge dependencies in the directed graph; and

causing the models in the subset of model implementations to be executed in parallel.

9. The computer-implemented method of claim 8, wherein for one or more ordered pairs of nodes in the graph the prediction output of a model implementation associated with a tail node in the pair serves as input to a model implementation associated with a head node in the pair.

10. The computer-implemented method of claim 7, further comprising:

assigning a system resource quota to a user associated with a representation;

determining an amount of system resources that model implementations for the respective user have consumed over a given period of time; and

selecting the model implementation based on an amount of unconsumed system resources

according to the system resource quota, wherein an amount of system resources that can be consumed by the selected model implementation does not exceed the amount of unconsumed system resources.

11. The computer-implemented method of claim 7, wherein the special purpose circuitry is an ASIC.

13. The system of claim 11, wherein for one or more ordered pairs of nodes in the graph the prediction output of a model implementation associated with a tail node in the pair serves as input to a model implementation associated with a head node in the pair.

12. A system, comprising:

a plurality of processing devices in data communication, wherein the processing devices are operable to perform operations comprising;

for each representation of a plurality of representations of predictive models, selecting a model implementation from a plurality of model implementation for the representation, wherein each representation is associated with a respective user and comprises a description of a respective predictive model, the selection based on a status of the representation'"'"'s respective user;

associating each of the model implementations with a respective node in a directed graph wherein for one or more ordered pairs of nodes in the graph the prediction output of a model implementation associated with a first node in the pair serves as input to a model implementation associated with a second node in the pair;

determining, from the directed graph, a subset of model implementations that can be executed in parallel based on edge dependencies in the directed graph; and

causing the models in the subset of model implementations to be executed in parallel.

14. The system of claim 12, the operations further comprising:

assigning a system resource quota to at least one respective user;

determining an amount of system resources that the model implementations for the respective user have consumed over a given period of time; and

selecting the model implementation based on an amount of unconsumed system resources

according to the system resource quota, wherein an amount of system resources that can be consumed by the selected model implementation does not exceed the amount of unconsumed

system resources.

15. The system of claim 12, the operations further comprising:

selecting a computationally-intensive model implementation for the model representation based on a determination that system resources were previously used during an off-peak time of day.

16. The system of claim 12, wherein:

for at least a first node in the directed graph the system realizes the node by special purpose circuitry, and a model implementation associated with the first node is configured to be executed by the special purpose circuitry; and

for at least a second node in the directed graph system realizes the node by a general purpose microprocessor, and a model implementation associated with the second node is configured to be executed by the general purpose microprocessor.

17. The computer-implemented method of claim 16, wherein the special purpose circuitry is an ASIC.

18. A system, comprising:

a plurality of processing devices in data communication, wherein the processing devices are operable to;

for each representation of a plurality of representations of predictive models, selecting a model implementation from a plurality of model implementation for the representation, wherein each representation comprises a description of a respective predictive model;

associating each of the model implementations with a respective node in a directed graph wherein for one or more ordered pairs of nodes in the graph the prediction output of a model implementation associated with a first node in the pair serves as input to a model implementation associated with a second node in the pair, wherein;

for at least a first node in the directed graph the system realizes the node by special purpose circuitry, and a model implementation associated with the first node and configured to be executed by the special purpose circuitry is selected for the first node based on its configuration to be executed by the special purpose circuitry; and

for at least a second node in the directed graph the system realizes the node by a general purpose microprocessor, and a model implementation associated with the second node and configured to be executed by the general purpose microprocessor is selected for the second node based on its configuration to be executed by the general purpose microprocessor.

19. The system of claim 18, further comprising:

determining, from the directed graph, a subset of model implementations that can be executed in parallel based on edge dependencies in the directed graph; and

causing the models in the subset of model implementations to be executed in parallel.

20. The system of claim 19, wherein for one or more ordered pairs of nodes in the graph the prediction output of a model implementation associated with a tail node in the pair serves as input to a model implementation associated with a head node in the pair.

21. The system of claim 18, further comprising:

assigning a system resource quota to a user associated with a representation;

determining an amount of system resources that model implementations for the respective user have consumed over a given period of time; and

selecting the model implementation based on an amount of unconsumed system resources

according to the system resource quota, wherein an amount of system resources that can be consumed by the selected model implementation does not exceed the amount of unconsumed system resources.

22. The system of claim 18, wherein the special purpose circuitry is an ASIC.